added repo files

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 Stability AI
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

LICENSES/LICENSE_ADP.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 archinet.ai
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

LICENSES/LICENSE_DESCRIPT.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023-present, Descript
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

LICENSES/LICENSE_META.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) Meta Platforms, Inc. and affiliates.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

LICENSES/LICENSE_NVIDIA.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 NVIDIA CORPORATION.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software. 
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

LICENSES/LICENSE_XTRANSFORMERS.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020 Phil Wang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

app.py

@@ -1,7 +1,20 @@
 import gradio as gr
+from stable_audio_tools import get_pretrained_model
+from stable_audio_tools.interface.gradio import create_ui
+import json 
+
+import torch
 
 def greet(name):
-    return "Hello " + name + "!!"
+    torch.manual_seed(42)
+
+    interface = create_ui(
+        model_config_path = "model_config_float_conditioning_dit_all.json", 
+        ckpt_path="epoch=684-step=319200.ckpt", 
+        pretrained_name="", 
+        pretransform_ckpt_path="",
+        model_half=False
+    )
+    interface.queue()
+    interface.launch(share=True)
 
-demo = gr.Interface(fn=greet, inputs="text", outputs="text")
-demo.launch()

defaults.ini

@@ -0,0 +1,56 @@
+
+[DEFAULTS]
+
+#name of the run
+name = stable_audio_tools
+
+# the batch size
+batch_size = 8 
+
+# number of GPUs to use for training
+num_gpus = 1 
+
+# number of nodes to use for training
+num_nodes = 1 
+
+# Multi-GPU strategy for PyTorch Lightning
+strategy = ""
+
+# Precision to use for training
+precision = "16-mixed"
+
+# number of CPU workers for the DataLoader
+num_workers = 8
+
+# the random seed
+seed = 42
+
+# Batches for gradient accumulation
+accum_batches = 1
+
+# Number of steps between checkpoints
+checkpoint_every = 10000                              
+                     
+# trainer checkpoint file to restart training from
+ckpt_path = ''
+
+# model checkpoint file to start a new training run from
+pretrained_ckpt_path = ''
+
+# Checkpoint path for the pretransform model if needed
+pretransform_ckpt_path = ''
+
+# configuration model specifying model hyperparameters
+model_config = ''
+
+# configuration for datasets
+dataset_config = ''
+
+# directory to save the checkpoints in
+save_dir = ''
+
+# gradient_clip_val passed into PyTorch Lightning Trainer
+gradient_clip_val = 0.0
+
+# remove the weight norm from the pretransform model
+remove_pretransform_weight_norm = ''

docs/autoencoders.md

@@ -0,0 +1,354 @@
+# Autoencoders
+At a high level, autoencoders are models constructed of two parts: an *encoder*, and a *decoder*. 
+
+The *encoder* takes in an sequence (such as mono or stereo audio) and outputs a compressed representation of that sequence as a d-channel "latent sequence", usually heavily downsampled by a constant factor.
+
+The *decoder* takes in a d-channel latent sequence and upsamples it back to the original input sequence length, reversing the compression of the encoder.
+
+Autoencoders are trained with a combination of reconstruction and adversarial losses in order to create a compact and invertible representation of raw audio data that allows downstream models to work in a data-compressed "latent space", with various desirable and controllable properties such as reduced sequence length, noise resistance, and discretization.
+
+The autoencoder architectures defined in `stable-audio-tools` are largely fully-convolutional, which allows autoencoders trained on small lengths to be applied to arbitrary-length sequences. For example, an autoencoder trained on 1-second samples could be used to encode 45-second inputs to a latent diffusion model.
+
+# Model configs
+The model config file for an autoencoder should set the `model_type` to `autoencoder`, and the `model` object should have the following properties:
+
+- `encoder`
+    - Configuration for the autoencoder's encoder half
+- `decoder`
+    - Configuration for the autoencoder's decoder half
+- `latent_dim`
+    - Latent dimension of the autoencoder, used by inference scripts and downstream models
+- `downsampling_ratio`
+    - Downsampling ratio between the input sequence and the latent sequence, used by inference scripts and downstream models
+- `io_channels`
+    - Number of input and output channels for the autoencoder when they're the same, used by inference scripts and downstream models
+- `bottleneck`
+    - Configuration for the autoencoder's bottleneck
+    - Optional
+- `pretransform`
+    - A pretransform definition for the autoencoder, such as wavelet decomposition or another autoencoder
+    - See [pretransforms.md](pretransforms.md) for more information
+    - Optional
+- `in_channels`
+    - Specifies the number of input channels for the autoencoder, when it's different from `io_channels`, such as in a mono-to-stereo model
+    - Optional
+- `out_channels`
+    - Specifies the number of output channels for the autoencoder, when it's different from `io_channels`
+    - Optional
+
+# Training configs
+The `training` config in the autoencoder model config file should have the following properties:
+- `learning_rate`
+    - The learning rate to use during training, fixed learning rate is currently the only option
+- `use_ema`
+    - If true, a copy of the model weights is maintained during training and updated as an exponential moving average of the trained model's weights. 
+    - Optional. Default: `false`
+- `warmup_steps`
+    - The number of training steps before turning on adversarial losses
+    - Optional. Default: `0`
+- `encoder_freeze_on_warmup`
+    - If true, freezes the encoder after the warmup steps have completed, so adversarial training only affects the decoder.
+    - Optional. Default: `false`
+- `loss_configs`
+    - Configurations for the loss function calculation
+    - Optional
+
+## Loss configs
+There are few different types of losses that are used for autoencoder training, including spectral losses, time-domain losses, adversarial losses, and bottleneck-specific losses.
+
+Hyperparameters fo these losses as well as loss weighting factors can be configured in the `loss_configs` property in the `training` config.
+
+### Spectral losses
+Multi-resolution STFT losses are the main reconstruction loss used for our audio autoencoders. We use the [auraloss](https://github.com/csteinmetz1/auraloss/tree/main/auraloss) library for our spectral loss functions. 
+
+For mono autoencoders (`io_channels` == 1), we use the [MultiResolutionSTFTLoss](https://github.com/csteinmetz1/auraloss/blob/1576b0cd6e927abc002b23cf3bfc455b660f663c/auraloss/freq.py#L329) module. 
+
+For stereo autoencoders (`io_channels` == 2), we use the [SumAndDifferenceSTFTLoss](https://github.com/csteinmetz1/auraloss/blob/1576b0cd6e927abc002b23cf3bfc455b660f663c/auraloss/freq.py#L533) module. 
+
+#### Example config
+```json
+"spectral": {
+    "type": "mrstft",
+    "config": {
+        "fft_sizes": [2048, 1024, 512, 256, 128, 64, 32],
+        "hop_sizes": [512, 256, 128, 64, 32, 16, 8],
+        "win_lengths": [2048, 1024, 512, 256, 128, 64, 32],
+        "perceptual_weighting": true
+    },
+    "weights": {
+        "mrstft": 1.0
+    }
+}
+```
+
+### Time-domain loss
+We compute the L1 distance between the original audio and the decoded audio to provide a time-domain loss.
+
+#### Example config
+```json
+"time": {
+    "type": "l1",
+    "weights": {
+        "l1": 0.1
+    }
+}
+```
+
+### Adversarial losses
+Adversarial losses bring in an ensemble of discriminator models to discriminate between real and fake audio, providing a signal to the autoencoder on perceptual discrepancies to fix.
+
+We largely rely on the [multi-scale STFT discriminator](https://github.com/facebookresearch/encodec/blob/0e2d0aed29362c8e8f52494baf3e6f99056b214f/encodec/msstftd.py#L99) from the EnCodec repo
+
+#### Example config
+```json
+"discriminator": {
+    "type": "encodec",
+    "config": {
+        "filters": 32,
+        "n_ffts": [2048, 1024, 512, 256, 128],
+        "hop_lengths": [512, 256, 128, 64, 32],
+        "win_lengths": [2048, 1024, 512, 256, 128]
+    },
+    "weights": {
+        "adversarial": 0.1,
+        "feature_matching": 5.0
+    }
+}
+```
+
+## Demo config
+The only property to set for autoencoder training demos is the `demo_every` property, determining the number of steps between demos.
+
+### Example config
+```json
+"demo": {
+    "demo_every": 2000
+}
+```
+
+# Encoder and decoder types
+Encoders and decoders are defined separately in the model configuration, so encoders and decoders from different model architectures and libraries can be used interchangeably. 
+
+## Oobleck
+Oobleck is Harmonai's in-house autoencoder architecture, implementing features from a variety of other autoencoder architectures.
+
+### Example config
+```json
+"encoder": {
+    "type": "oobleck",
+    "config": {
+        "in_channels": 2,
+        "channels": 128,
+        "c_mults": [1, 2, 4, 8],
+        "strides": [2, 4, 8, 8],
+        "latent_dim": 128,
+        "use_snake": true
+    }
+},
+"decoder": {
+    "type": "oobleck",
+    "config": {
+        "out_channels": 2,
+        "channels": 128,
+        "c_mults": [1, 2, 4, 8],
+        "strides": [2, 4, 8, 8],
+        "latent_dim": 64,
+        "use_snake": true,
+        "use_nearest_upsample": false
+    }
+}
+```
+
+## DAC
+This is the Encoder and Decoder definitions from the `descript-audio-codec` repo. It's a simple fully-convolutional autoencoder with channels doubling every level. The encoder and decoder configs are passed directly into the constructors for the DAC [Encoder](https://github.com/descriptinc/descript-audio-codec/blob/c7cfc5d2647e26471dc394f95846a0830e7bec34/dac/model/dac.py#L64) and [Decoder](https://github.com/descriptinc/descript-audio-codec/blob/c7cfc5d2647e26471dc394f95846a0830e7bec34/dac/model/dac.py#L115).
+
+**Note: This does not include the DAC quantizer, and does not load pre-trained DAC models, this is just the encoder and decoder definitions.**
+
+### Example config
+```json
+"encoder": {
+    "type": "dac",
+    "config": {
+        "in_channels": 2,
+        "latent_dim": 32,
+        "d_model": 128,
+        "strides": [2, 4, 4, 4]
+    }
+},
+"decoder": {
+    "type": "dac",
+    "config": {
+        "out_channels": 2,
+        "latent_dim": 32,
+        "channels": 1536,
+        "rates": [4, 4, 4, 2]
+    }
+}
+```
+
+## SEANet
+This is the SEANetEncoder and SEANetDecoder definitions from Meta's EnCodec repo. This is the same encoder and decoder architecture used in the EnCodec models used in MusicGen, without the quantizer.
+
+The encoder and decoder configs are passed directly into the [SEANetEncoder](https://github.com/facebookresearch/encodec/blob/0e2d0aed29362c8e8f52494baf3e6f99056b214f/encodec/modules/seanet.py#L66C12-L66C12) and [SEANetDecoder](https://github.com/facebookresearch/encodec/blob/0e2d0aed29362c8e8f52494baf3e6f99056b214f/encodec/modules/seanet.py#L147) classes directly, though we reverse the input order of the strides (ratios) in the encoder to make it consistent with the order in the decoder.
+
+### Example config
+```json
+"encoder": {
+    "type": "seanet",
+    "config": {
+        "channels": 2,
+        "dimension": 128,
+        "n_filters": 64,
+        "ratios": [4, 4, 8, 8],
+        "n_residual_layers": 1,
+        "dilation_base": 2,
+        "lstm": 2,
+        "norm": "weight_norm"
+    }
+},
+"decoder": {
+    "type": "seanet",
+    "config": {
+        "channels": 2,
+        "dimension": 64,
+        "n_filters": 64,
+        "ratios": [4, 4, 8, 8],
+        "n_residual_layers": 1,
+        "dilation_base": 2,
+        "lstm": 2,
+        "norm": "weight_norm"
+    }
+},
+```
+
+# Bottlenecks
+In our terminology, the "bottleneck" of an autoencoder is a module placed between the encoder and decoder to enforce particular constraints on the latent space the encoder creates.
+
+Bottlenecks have a similar interface to the autoencoder with `encode()` and `decode()` functions defined. Some bottlenecks return extra information in addition to the output latent series, such as quantized token indices, or additional losses to be considered during training.
+
+To define a bottleneck for the autoencoder, you can provide the `bottleneck` object in the autoencoder's model configuration, with the following 
+
+## VAE
+
+The Variational Autoencoder (VAE) bottleneck splits the encoder's output in half along the channel dimension, treats the two halves as the "mean" and "scale" parameters for VAE sampling, and performs the latent sampling. At a basic level, the "scale" values determine the amount of noise to add to the "mean" latents, which creates a noise-resistant latent space where more of the latent space decodes to perceptually "valid" audio. This is particularly helpful for diffusion models where the outpus of the diffusion sampling process leave a bit of Gaussian error noise.
+
+**Note: For the VAE bottleneck to work, the output dimension of the encoder must be twice the size of the input dimension for the decoder.**
+
+### Example config
+```json
+"bottleneck": {
+    "type": "vae"
+}
+```
+
+### Extra info
+The VAE bottleneck also returns a `kl` value in the encoder info. This is the [KL divergence](https://en.wikipedia.org/wiki/Kullback%E2%80%93Leibler_divergence) between encoded/sampled latent space and a Gaussian distribution. By including this value as a loss value to optimize, we push our latent distribution closer to a normal distribution, potentially trading off reconstruction quality.
+
+### Example loss config
+```json
+"bottleneck": {
+    "type": "kl",
+    "weights": {
+        "kl": 1e-4
+    }
+}
+```
+
+## Tanh
+This bottleneck applies the tanh function to the latent series, "soft-clipping" the latent values to be between -1 and 1. This is a quick and dirty way to enforce a limit on the variance of the latent space, but training these models can be unstable as it's seemingly easy for the latent space to saturate the values to -1 or 1 and never recover.
+
+### Example config
+```json
+"bottleneck": {
+    "type": "tanh"
+}
+```
+
+## Wasserstein
+The Wasserstein bottleneck implements the WAE-MMD regularization method from the [Wasserstein Auto-Encoders](https://arxiv.org/abs/1711.01558) paper, calculating the Maximum Mean Discrepancy (MMD) between the latent space and a Gaussian distribution. Including this value as a loss value to optimize leads to a more Gaussian latent space, but does not require stochastic sampling as with a VAE, so the encoder is deterministic.
+
+The Wasserstein bottleneck also exposes the `noise_augment_dim` property, which concatenates `noise_augment_dim` channels of Gaussian noise to the latent series before passing into the decoder. This adds some stochasticity to the latents which can be helpful for adversarial training, while keeping the encoder outputs deterministic.
+
+**Note: The MMD calculation is very VRAM-intensive for longer sequence lengths, so training a Wasserstein autoencoder is best done on autoencoders with a decent downsampling factor, or on short sequence lengths. For inference, the MMD calculation is disabled.**
+
+### Example config
+```json
+"bottleneck": {
+    "type": "wasserstein"
+}
+```
+
+### Extra info
+This bottleneck adds the `mmd` value to the encoder info, representing the Maximum Mean Discrepancy.
+
+### Example loss config
+```json
+"bottleneck": {
+    "type": "mmd",
+    "weights": {
+        "mmd": 100
+    }
+}
+```
+
+## L2 normalization (Spherical autoencoder)
+The L2 normalization bottleneck normalizes the latents across the channel-dimension, projecting the latents to a d-dimensional hypersphere. This acts as a form of latent space normalization.
+
+
+### Example config
+```json
+"bottleneck": {
+    "type": "l2_norm"
+}
+```
+
+
+## RVQ
+Residual vector quantization (RVQ) is currently the leading method for learning discrete neural audio codecs (tokenizers for audio). In vector quantization, each item in the latent sequence is individually "snapped" to the nearest vector in a discrete "codebook" of learned vectors. The index of the vector in the codebook can then be used as a token index for things like autoregressive transformers. Residual vector quantization improves the precision of normal vector quantization by adding additional codebooks. For a deeper dive into RVQ, check out [this blog post by Dr. Scott Hawley](https://drscotthawley.github.io/blog/posts/2023-06-12-RVQ.html).
+
+This RVQ bottleneck uses [lucidrains' implementation](https://github.com/lucidrains/vector-quantize-pytorch/tree/master) from the `vector-quantize-pytorch` repo, which provides a lot of different quantizer options. The bottleneck config is passed through to the `ResidualVQ`  [constructor](https://github.com/lucidrains/vector-quantize-pytorch/blob/0c6cea24ce68510b607f2c9997e766d9d55c085b/vector_quantize_pytorch/residual_vq.py#L26).
+
+**Note: This RVQ implementation uses manual replacement of codebook vectors to reduce codebook collapse. This does not work with multi-GPU training as the random replacement is not synchronized across devices.** 
+
+### Example config
+```json
+"bottleneck": {
+    "type": "rvq",
+    "config": {
+        "num_quantizers": 4,
+        "codebook_size": 2048,
+        "dim": 1024,
+        "decay": 0.99,
+    }
+}
+```
+
+## DAC RVQ
+This is the residual vector quantization implementation from the `descript-audio-codec` repo. It differs from the above implementation in that it does not use manual replacements to improve codebook usage, but instead uses learnable linear layers to project the latents down to a lower-dimensional space before performing the individual quantization operations. This means it's compatible with distributed training. 
+
+The bottleneck config is passed directly into the `ResidualVectorQuantize` [constructor](https://github.com/descriptinc/descript-audio-codec/blob/c7cfc5d2647e26471dc394f95846a0830e7bec34/dac/nn/quantize.py#L97).
+
+The `quantize_on_decode` property is also exposed, which moves the quantization process to the decoder. This should not be used during training, but is helpful when training latent diffusion models that use the quantization process as a way to remove error after the diffusion sampling process.
+
+### Example config
+```json
+"bottleneck": {
+    "type": "dac_rvq",
+    "config": {
+        "input_dim": 64,
+        "n_codebooks": 9,
+        "codebook_dim": 32,
+        "codebook_size": 1024,
+        "quantizer_dropout": 0.5
+    }
+}
+```
+
+### Extra info
+The DAC RVQ bottleneck also adds the following properties to the `info` object:
+- `pre_quantizer`
+    - The pre-quantization latent series, useful in combination with `quantize_on_decode` for training latent diffusion models.
+- `vq/commitment_loss`
+    - Commitment loss for the quantizer
+- `vq/codebook_loss`
+    - Codebook loss for the quantizer

docs/conditioning.md

@@ -0,0 +1,151 @@
+# Conditioning
+Conditioning, in the context of `stable-audio-tools` is the use of additional signals in a model that are used to add an additional level of control over the model's behavior. For example, we can condition the outputs of a diffusion model on a text prompt, creating a text-to-audio model.
+
+# Conditioning types
+There are a few different kinds of conditioning depending on the conditioning signal being used.
+
+## Cross attention
+Cross attention is a type of conditioning that allows us to find correlations between two sequences of potentially different lengths. For example, cross attention allows us to find correlations between a sequence of features from a text encoder and a sequence of high-level audio features.
+
+Signals used for cross-attention conditioning should be of the shape `[batch, sequence, channels]`.
+
+## Global conditioning
+Global conditioning is the use of a single n-dimensional tensor to provide conditioning information that pertains to the whole sequence being conditioned. For example, this could be the single embedding output of a CLAP model, or a learned class embedding.
+
+Signals used for global conditioning should be of the shape `[batch, channels]`.
+
+## Input concatenation
+Input concatenation applies a spatial conditioning signal to the model that correlates in the sequence dimension with the model's input, and is of the same length. The conditioning signal will be concatenated with the model's input data along the channel dimension. This can be used for things like inpainting information, melody conditioning, or for creating a diffusion autoencoder.
+
+Signals used for input concatenation conditioning should be of the shape `[batch, channels, sequence]` and must be the same length as the model's input.
+
+# Conditioners and conditioning configs
+`stable-audio-tools` uses Conditioner modules to translate human-readable metadata such as text prompts or a number of seconds into tensors that the model can take as input. 
+
+Each conditioner has a corresponding `id` that it expects to find in the conditioning dictionary provided during training or inference. Each conditioner takes in the relevant conditioning data and returns a tuple containing the corresponding tensor and a mask.
+
+The ConditionedDiffusionModelWrapper manages the translation between the user-provided metadata dictionary (e.g. `{"prompt": "a beautiful song", "seconds_start": 22, "seconds_total": 193}`) and the dictionary of different conditioning types that the model uses (e.g. `{"cross_attn_cond": ...}`).
+
+To apply conditioning to a model, you must provide a `conditioning` configuration in the model's config. At the moment, we only support conditioning diffusion models though the `diffusion_cond` model type.
+
+The `conditioning` configuration should contain a `configs` array, which allows you to define multiple conditioning signals. 
+
+Each item in `configs` array should define the `id` for the corresponding metadata, the type of conditioner to be used, and the config for that conditioner.
+
+The `cond_dim` property is used to enforce the same dimension on all conditioning inputs, however that can be overridden with an explicit `output_dim` property on any of the individual configs.
+
+## Example config
+```json
+"conditioning": {
+    "configs": [
+        {
+            "id": "prompt",
+            "type": "t5",
+            "config": {
+                "t5_model_name": "t5-base",
+                "max_length": 77,
+                "project_out": true
+            }
+        }
+    ],
+    "cond_dim": 768
+}
+```
+
+# Conditioners
+
+## Text encoders
+
+### `t5`
+This uses a frozen [T5](https://huggingface.co/docs/transformers/model_doc/t5) text encoder from the `transformers` library to encode text prompts into a sequence of text features.
+
+The `t5_model_name` property determines which T5 model is loaded from the `transformers` library.
+
+The `max_length` property determines the maximum number of tokens that the text encoder will take in, as well as the sequence length of the output text features.
+
+If you set `enable_grad` to `true`, the T5 model will be un-frozen and saved with the model checkpoint, allowing you to fine-tune the T5 model.
+
+T5 encodings are only compatible with cross attention conditioning.
+
+#### Example config 
+```json
+{
+    "id": "prompt",
+    "type": "t5",
+    "config": {
+        "t5_model_name": "t5-base",
+        "max_length": 77,
+        "project_out": true
+    }
+}
+```
+
+### `clap_text`
+This loads the text encoder from a [CLAP](https://github.com/LAION-AI/CLAP) model, which can provide either a sequence of text features, or a single multimodal text/audio embedding.
+
+The CLAP model must be provided with a local file path, set in the `clap_ckpt_path` property,along with the correct `audio_model_type` and `enable_fusion` properties for the provided model.
+
+If the `use_text_features` property is set to `true`, the conditioner output will be a sequence of text features, instead of a single multimodal embedding. This allows for more fine-grained text information to be used by the model, at the cost of losing the ability to prompt with CLAP audio embeddings.
+
+By default, if `use_text_features` is true, the last layer of the CLAP text encoder's features are returned. You can return the text features of earlier layers by specifying the index of the layer to return in the `feature_layer_ix` property. For example, you can return the text features of the next-to-last layer of the CLAP model by setting `feature_layer_ix` to `-2`.
+
+If you set `enable_grad` to `true`, the CLAP model will be un-frozen and saved with the model checkpoint, allowing you to fine-tune the CLAP model.
+
+CLAP text embeddings are compatible with global conditioning and cross attention conditioning. If `use_text_features` is set to `true`, the features are not compatible with global conditioning.
+
+#### Example config
+```json
+{
+    "id": "prompt",
+    "type": "clap_text",
+    "config": {
+        "clap_ckpt_path": "/path/to/clap/model.ckpt",
+        "audio_model_type": "HTSAT-base",
+        "enable_fusion": true,
+        "use_text_features": true,
+        "feature_layer_ix": -2
+    }
+}
+```
+
+## Number encoders
+
+### `int`
+The IntConditioner takes in a list of integers in a given range, and returns a discrete learned embedding for each of those integers.
+
+The `min_val` and `max_val` properties set the range of the embedding values. Input integers are clamped to this range.
+
+This can be used for things like discrete timing embeddings, or learned class embeddings.
+
+Int embeddings are compatible with global conditioning and cross attention conditioning.
+
+#### Example config
+```json
+{
+    "id": "seconds_start",
+    "type": "int",
+    "config": {
+        "min_val": 0,
+        "max_val": 512
+    }
+}
+```
+
+### `number`
+The NumberConditioner takes in a a list of floats in a given range, and returns a continuous Fourier embedding of the provided floats.
+
+The `min_val` and `max_val` properties set the range of the float values. This is the range used to normalize the input float values.
+
+Number embeddings are compatible with global conditioning and cross attention conditioning.
+
+#### Example config
+```json
+{
+    "id": "seconds_total",
+    "type": "number",
+    "config": {
+        "min_val": 0,
+        "max_val": 512
+    }
+}
+```

docs/datasets.md

@@ -0,0 +1,75 @@
+# Datasets
+`stable-audio-tools` supports loading data from local file storage, as well as loading audio files and JSON files in the [WebDataset](https://github.com/webdataset/webdataset/tree/main/webdataset) format from Amazon S3 buckets.
+
+# Dataset configs
+To specify the dataset used for training, you must provide a dataset config JSON file to `train.py`.
+
+The dataset config consists of a `dataset_type` property specifying the type of data loader to use, a `datasets` array to provide multiple data sources, and a `random_crop` property, which decides if the cropped audio from the training samples is from a random place in the audio file, or always from the beginning.
+
+## Local audio files
+To use a local directory of audio samples, set the `dataset_type` property in your dataset config to `"audio_dir"`, and provide a list of objects to the `datasets` property including the `path` property, which should be the path to your directory of audio samples.
+
+This will load all of the compatible audio files from the provided directory and all subdirectories.
+
+### Example config 
+```json
+{
+    "dataset_type": "audio_dir",
+    "datasets": [
+        {
+            "id": "my_audio",
+            "path": "/path/to/audio/dataset/"
+        }
+    ],
+    "random_crop": true
+}
+```
+
+## S3 WebDataset
+To load audio files and related metadata from .tar files in the WebDataset format hosted in Amazon S3 buckets, you can set the `dataset_type` property to `s3`, and provide the `datasets` parameter with a list of objects containing the AWS S3 path to the shared S3 bucket prefix of the WebDataset .tar files. The S3 bucket will be searched recursively given the path, and assumes any .tar files found contain audio files and corresponding JSON files where the related files differ only in file extension (e.g. "000001.flac", "000001.json", "00002.flac", "00002.json", etc.)
+
+### Example config
+```json
+{
+    "dataset_type": "s3",
+    "datasets": [
+        {
+            "id": "s3-test",
+            "s3_path": "s3://my-bucket/datasets/webdataset/audio/"
+        }
+    ],
+    "random_crop": true
+}
+```
+
+# Custom metadata
+To customize the metadata provided to the conditioners during model training, you can provide a separate custom metadata module to the dataset config. This metadata module should be a Python file that must contain a function called `get_custom_metadata` that takes in two parameters, `info`, and `audio`, and returns a dictionary. 
+
+For local training, the `info` parameter will contain a few pieces of information about the loaded audio file, such as the path, and information about how the audio was cropped from the original training sample. For S3 WebDataset datasets, it will also contain the metadata from the related JSON files. 
+
+The `audio` parameter contains the audio sample that will be passed to the model at training time. This lets you analyze the audio for extra properties that you can then pass in as extra conditioning signals.
+
+The dictionary returned from the `get_custom_metadata` function will have its properties added to the `metadata` object used at training time. For more information on how conditioning works, please see the [Conditioning documentation](./conditioning.md)
+
+## Example config and custom metadata module
+```json
+{
+    "dataset_type": "audio_dir",
+    "datasets": [
+        {
+            "id": "my_audio",
+            "path": "/path/to/audio/dataset/"
+        }
+    ],
+    "custom_metadata_module": "/path/to/custom_metadata.py",
+    "random_crop": true
+}
+```
+
+`custom_metadata.py`:
+```py
+def get_custom_metadata(info, audio):
+
+    # Pass in the relative path of the audio file as the prompt
+    return {"prompt": info["relpath"]}
+```

docs/pretransforms.md

@@ -0,0 +1,43 @@
+# Pretransforms
+Many models require some fixed transform to be applied to the input audio before the audio is passed in to the trainable layers of the model, as well as a corresponding inverse transform to be applied to the outputs of the model. We refer to these as "pretransforms".
+
+At the moment, `stable-audio-tools` supports two pretransforms, frozen autoencoders for latent diffusion models and wavelet decompositions.
+
+Pretransforms have a similar interface to autoencoders with "encode" and "decode" functions defined for each pretransform.
+
+## Autoencoder pretransform
+To define a model with an autoencoder pretransform, you can define the "pretransform" property in the model config, with the `type` property set to `autoencoder`. The `config` property should be an autoencoder model definition.
+
+Example:
+```json
+"pretransform": {
+    "type": "autoencoder",
+    "config": {
+        "encoder": {
+            ...
+        },
+        "decoder": {
+            ...
+        }
+        ...normal autoencoder configuration
+    }
+}
+```
+
+### Latent rescaling
+The original [Latent Diffusion paper](https://arxiv.org/abs/2112.10752) found that rescaling the latent series to unit variance before performing diffusion improved quality. To this end, we expose a `scale` property on autoencoder pretransforms that will take care of this rescaling. The scale should be set to the original standard deviation of the latents, which can be determined experimentally, or by looking at the `latent_std` value during training. The pretransform code will divide by this scale factor in the `encode` function and multiply by this scale in the `decode` function.
+
+## Wavelet pretransform
+`stable-audio-tools` also exposes wavelet decomposition as a pretransform. Wavelet decomposition is a quick way to trade off sequence length for channels in autoencoders, while maintaining a multi-band implicit bias.
+
+Wavelet pretransforms take the following properties:
+
+- `channels`
+    - The number of input and output audio channels for the wavelet transform
+- `levels`
+    - The number of successive wavelet decompositions to perform. Each level doubles the channel count and halves the sequence length
+- `wavelet`
+    - The specific wavelet from [PyWavelets](https://pywavelets.readthedocs.io/en/latest/ref/wavelets.html) to use, currently limited to `"bior2.2", "bior2.4", "bior2.6", "bior2.8", "bior4.4", "bior6.8"`
+
+## Future work
+We hope to add more filters and transforms to this list, including PQMF and STFT transforms.

model_config_float_conditioning_dit_all.json

@@ -0,0 +1,208 @@
+{
+  "model_type": "diffusion_cond",
+  "sample_size": 1048576,
+  "sample_rate": 44100,
+  "audio_channels": 1,
+  "model": {
+    "pretransform": {
+      "type": "autoencoder",
+      "iterate_batch": true,
+      "config": {
+          "encoder": {
+              "type": "dac",
+              "config": {
+                  "in_channels": 1,
+                  "latent_dim": 128,
+                  "d_model": 128,
+                  "strides": [4, 4, 8, 8]
+              }
+          },
+          "decoder": {
+              "type": "dac",
+              "config": {
+                  "out_channels": 1,
+                  "latent_dim": 64,
+                  "channels": 1536,
+                  "rates": [8, 8, 4, 4]
+              }
+          },
+          "bottleneck": {
+              "type": "vae"
+          },
+          "latent_dim": 64,
+          "downsampling_ratio": 1024,
+          "io_channels": 1
+      }
+  },
+    "conditioning": {
+      "configs": [
+        {
+          "id": "latitude",
+          "type": "number",
+          "config": {
+            "min_val": -54.617412,
+            "max_val": -10.13994
+          }
+        },
+        {
+          "id": "longitude",
+          "type": "number",
+          "config": {
+            "min_val": 96.8233,
+            "max_val": 167.9619
+          }
+        },
+        {
+          "id": "temperature",
+          "type": "number",
+          "config": {
+            "min_val": -10.0,
+            "max_val": 55.0
+          }
+        },
+        {
+          "id": "humidity",
+          "type": "number",
+          "config": {
+            "min_val": 1,
+            "max_val": 100.0
+          }
+        },
+        {
+          "id": "wind_speed",
+          "type": "number",
+          "config": {
+            "min_val": 0,
+            "max_val": 50.0
+          }
+        },
+        {
+          "id": "pressure",
+          "type": "number",
+          "config": {
+            "min_val": 800.0,
+            "max_val": 1200.0
+          }
+        },
+        {
+          "id": "minutes_of_day",
+          "type": "number",
+          "config": {
+            "min_val": 0,
+            "max_val": 1439
+          }
+        },
+        {
+          "id": "day_of_year",
+          "type": "number",
+          "config": {
+            "min_val": 1,
+            "max_val": 366
+          }
+        },
+        {
+          "id": "seconds_start",
+          "type": "number",
+          "config": {
+            "min_val": 0,
+            "max_val": 512
+          }
+        },
+        {
+          "id": "seconds_total",
+          "type": "number",
+          "config": {
+            "min_val": 0,
+            "max_val": 512
+          }
+        }
+      ],
+      "cond_dim": 768
+    },
+    "diffusion": {
+      "cross_attention_cond_ids": ["latitude", "longitude", "temperature", "humidity", "wind_speed", "pressure", "minutes_of_day", "day_of_year","seconds_start", "seconds_total"],
+      "global_cond_ids": ["seconds_start", "seconds_total"],
+      "type": "dit",
+      "config": {
+        "io_channels": 64,
+        "embed_dim": 768,
+        "depth": 24,
+        "num_heads": 24,
+        "cond_token_dim": 768,
+        "global_cond_dim": 1536,
+        "project_cond_tokens": false,
+        "transformer_type": "continuous_transformer"
+      }
+    },
+    "io_channels": 64
+  },
+
+  "training": {
+    "use_ema": true,
+    "log_loss_info": false,
+    "optimizer_configs": {
+      "diffusion": {
+        "optimizer": {
+          "type": "AdamW",
+          "config": {
+            "lr": 5e-5,
+            "betas": [0.9, 0.999],
+            "weight_decay": 1e-3
+          }
+        },
+        "scheduler": {
+          "type": "InverseLR",
+          "config": {
+            "inv_gamma": 1000000,
+            "power": 0.5,
+            "warmup": 0.99
+          }
+        }
+      }
+    },
+    "demo": {
+      "demo_every": 2500,
+      "demo_steps": 100,
+      "num_demos": 3,
+      "demo_cfg_scales": [3, 5, 7],
+      "demo_cond": [
+        {
+          "latitude": -24.005512,
+          "longitude": 133.368348,
+          "temperature": 25.5,
+          "humidity": 60,
+          "wind_speed": 8,
+          "pressure": 1000,
+          "minutes_of_day": 400,
+          "day_of_year": 110,
+          "seconds_start": 0,
+          "seconds_total": 22
+        },
+        {
+          "latitude": -26.987815,
+          "longitude": 153.129068,
+          "temperature": 31.5,
+          "humidity": 70,
+          "wind_speed": 12,
+          "pressure": 1010,
+          "minutes_of_day": 600,
+          "day_of_year": 57,
+          "seconds_start": 0,
+          "seconds_total": 22
+        },
+        {
+          "latitude": -12.546364,
+          "longitude": 130.919605,
+          "temperature": 28.5,
+          "humidity": 60,
+          "wind_speed": 18,
+          "pressure": 1015,
+          "minutes_of_day": 1140,
+          "day_of_year": 280,
+          "seconds_start": 0,
+          "seconds_total": 22
+        }
+      ]
+    }
+  }
+}

pyproject.toml

@@ -0,0 +1,3 @@
+[build-system]
+requires = ["setuptools"]
+build-backend = "setuptools.build_meta"

run_gradio.py

@@ -0,0 +1,31 @@
+from stable_audio_tools import get_pretrained_model
+from stable_audio_tools.interface.gradio import create_ui
+import json 
+
+import torch
+
+def main(args):
+    torch.manual_seed(42)
+
+    interface = create_ui(
+        model_config_path = args.model_config, 
+        ckpt_path=args.ckpt_path, 
+        pretrained_name=args.pretrained_name, 
+        pretransform_ckpt_path=args.pretransform_ckpt_path,
+        model_half=args.model_half
+    )
+    interface.queue()
+    interface.launch(share=True, auth=(args.username, args.password) if args.username is not None else None)
+
+if __name__ == "__main__":
+    import argparse
+    parser = argparse.ArgumentParser(description='Run gradio interface')
+    parser.add_argument('--pretrained-name', type=str, help='Name of pretrained model', required=False)
+    parser.add_argument('--model-config', type=str, help='Path to model config', required=False)
+    parser.add_argument('--ckpt-path', type=str, help='Path to model checkpoint', required=False)
+    parser.add_argument('--pretransform-ckpt-path', type=str, help='Optional to model pretransform checkpoint', required=False)
+    parser.add_argument('--username', type=str, help='Gradio username', required=False)
+    parser.add_argument('--password', type=str, help='Gradio password', required=False)
+    parser.add_argument('--model-half', action='store_true', help='Whether to use half precision', required=False)
+    args = parser.parse_args()
+    main(args)

run_tests.py

@@ -0,0 +1,44 @@
+from stable_audio_tools import get_pretrained_model
+from stable_audio_tools.interface.testing import runTests
+print(runTests)  # Check if it prints a function reference
+
+
+import torch
+
+def main(args):
+    torch.manual_seed(42)
+    runTests(model_config_path = args.model_config, 
+        ckpt_path=args.ckpt_path, 
+        pretrained_name=args.pretrained_name, 
+        pretransform_ckpt_path=args.pretransform_ckpt_path,
+        model_half=args.model_half,
+        output_dir=args.output_dir,
+        json_dir=args.json_dir
+    )
+    
+
+
+
+
+if __name__ == "__main__":
+    import argparse
+    import sys
+    parser = argparse.ArgumentParser(description='Run generation tests')
+    parser.add_argument('--pretrained-name', type=str, help='Name of pretrained model', required=False)
+    parser.add_argument('--model-config', type=str, help='Path to model config', required=False)
+    parser.add_argument('--ckpt-path', type=str, help='Path to model checkpoint', required=False)
+    parser.add_argument('--pretransform-ckpt-path', type=str, help='Optional to model pretransform checkpoint', required=False)
+    parser.add_argument('--model-half', action='store_true', help='Whether to use half precision', required=False)
+    parser.add_argument('--output-dir', type=str, help='Path to output directory', required=True)
+    parser.add_argument('--json-dir', type=str, help='Path to directory containing JSON files', required=True)
+    print("Running tests")
+
+    print("Arguments provided:", sys.argv[1:])
+    
+    args = parser.parse_args()
+    print("Parsed arguments:", args)
+    main(args)
+
+    
+
+

scripts/ds_zero_to_pl_ckpt.py

@@ -0,0 +1,14 @@
+import argparse
+from lightning.pytorch.utilities.deepspeed import convert_zero_checkpoint_to_fp32_state_dict
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--save_path", type=str, help="Path to the zero checkpoint")
+    parser.add_argument("--output_path", type=str, help="Path to the output checkpoint", default="lightning_model.pt")
+    args = parser.parse_args()
+
+    # lightning deepspeed has saved a directory instead of a file
+    save_path = args.save_path
+    output_path = args.output_path
+    convert_zero_checkpoint_to_fp32_state_dict(save_path, output_path)

setup.py

@@ -0,0 +1,46 @@
+from setuptools import setup, find_packages
+
+setup(
+    name='stable-audio-tools',
+    version='0.0.12',
+    url='https://github.com/Stability-AI/stable-audio-tools.git',
+    author='Stability AI',
+    description='Training and inference tools for generative audio models from Stability AI',
+    packages=find_packages(),  
+    install_requires=[
+        'audiocraft==1.0.0',
+        'aeiou==0.0.20',
+        'alias-free-torch==0.0.6',
+        'auraloss==0.4.0',
+        'descript-audio-codec==1.0.0',
+        'einops==0.7.0',
+        'einops-exts==0.0.4',
+        'ema-pytorch==0.2.3',
+        'encodec==0.1.1',
+        'flash-attn>=2.5.0',
+        'gradio>=3.42.0',
+        'huggingface_hub',
+        'importlib-resources==5.12.0',
+        'k-diffusion==0.1.1',
+        'laion-clap==1.1.4',
+        'local-attention==1.8.6',
+        'pandas==2.0.2',
+        'pedalboard==0.7.4',
+        'prefigure==0.0.9',
+        'pytorch_lightning==2.1.0', 
+        'PyWavelets==1.4.1',
+        'safetensors',
+        'sentencepiece==0.1.99',
+        's3fs',
+        'torch>=2.0.1',
+        'torchaudio>=2.0.2',
+        'torchmetrics==0.11.4',
+        'tqdm',
+        'transformers==4.33.3',
+        'v-diffusion-pytorch==0.0.2',
+        'vector-quantize-pytorch==1.9.14',
+        'wandb==0.15.4',
+        'webdataset==0.2.48',
+        'x-transformers<1.27.0'
+    ],
+)

stable_audio_tools/__init__.py

@@ -0,0 +1,2 @@
+from .models.factory import create_model_from_config, create_model_from_config_path
+from .models.pretrained import get_pretrained_model

stable_audio_tools/configs/dataset_configs/custom_metadata/custom_md_example.py

@@ -0,0 +1,4 @@
+def get_custom_metadata(info, audio):
+
+    # Use relative path as the prompt
+    return {"prompt": info["relpath"]}

stable_audio_tools/configs/dataset_configs/local_training_example.json

@@ -0,0 +1,11 @@
+{
+    "dataset_type": "audio_dir",
+    "datasets": [
+        {
+            "id": "my_audio",
+            "path": "/path/to/audio/dataset/"
+        }
+    ],
+    "custom_metadata_module": "/path/to/custom_metadata/custom_md_example.py",
+    "random_crop": true
+}

stable_audio_tools/configs/dataset_configs/s3_wds_example.json

@@ -0,0 +1,10 @@
+{
+    "dataset_type": "s3",
+    "datasets": [
+        {
+            "id": "s3-test",
+            "s3_path": "s3://my-bucket/datasets/webdataset/audio/"
+        }
+    ],
+    "random_crop": true
+}

stable_audio_tools/configs/model_configs/autoencoders/dac_2048_32_vae.json

@@ -0,0 +1,71 @@
+{
+    "model_type": "autoencoder",
+    "sample_size": 65536,
+    "sample_rate": 44100,
+    "audio_channels": 1,
+    "model": {
+        "encoder": {
+            "type": "dac",
+            "config": {
+                "latent_dim": 64,
+                "d_model": 128,
+                "strides": [4, 8, 8, 8]
+            }
+        },
+        "decoder": {
+            "type": "dac",
+            "config": {
+                "latent_dim": 32,
+                "channels": 1536,
+                "rates": [8, 8, 8, 4]
+            }
+        },
+        "bottleneck": {
+            "type": "vae"
+        },
+        "latent_dim": 32,
+        "downsampling_ratio": 2048,
+        "io_channels": 1
+    },
+    "training": {
+        "learning_rate": 1e-4,
+        "warmup_steps": 0,
+        "use_ema": false,
+        "loss_configs": {
+            "discriminator": {
+                "type": "encodec",
+                "config": {
+                    "filters": 32,
+                    "n_ffts": [2048, 1024, 512, 256, 128, 64, 32],
+                    "hop_lengths": [512, 256, 128, 64, 32, 16, 8],
+                    "win_lengths": [2048, 1024, 512, 256, 128, 64, 32]
+                },
+                "weights": {
+                    "adversarial": 0.1,
+                    "feature_matching": 5.0
+                }
+            },
+            "spectral": {
+                "type": "mrstft",
+                "config": {
+                    "fft_sizes": [2048, 1024, 512, 256, 128, 64, 32],
+                    "hop_sizes": [512, 256, 128, 64, 32, 16, 8],
+                    "win_lengths": [2048, 1024, 512, 256, 128, 64, 32],
+                    "perceptual_weighting": true
+                },
+                "weights": {
+                    "mrstft": 1.0
+                }
+            },
+            "time": {
+                "type": "l1",
+                "weights": {
+                    "l1": 0.0
+                }
+            }
+        },
+        "demo": {
+            "demo_every": 2000
+        }
+    }
+}

stable_audio_tools/configs/model_configs/autoencoders/encodec_musicgen_rvq.json

@@ -0,0 +1,88 @@
+{
+    "model_type": "autoencoder",
+    "sample_size": 32000,
+    "sample_rate": 32000,
+    "audio_channels": 1,
+    "model": {
+        "encoder": {
+            "type": "seanet",
+            "config": {
+                "channels": 1,
+                "dimension": 128,
+                "n_filters": 64,
+                "ratios": [4, 4, 5, 8],
+                "n_residual_layers": 1,
+                "dilation_base": 2,
+                "lstm": 2,
+                "norm": "weight_norm"
+            }
+        },
+        "decoder": {
+            "type": "seanet",
+            "config": {
+                "channels": 1,
+                "dimension": 128,
+                "n_filters": 64,
+                "ratios": [4, 4, 5, 8],
+                "n_residual_layers": 1,
+                "dilation_base": 2,
+                "lstm": 2,
+                "norm": "weight_norm"
+            }
+        },
+        "bottleneck": {
+            "type": "rvq",
+            "config": {
+                "num_quantizers": 4,
+                "codebook_size": 2048,
+                "dim": 128,
+                "decay": 0.99,
+                "threshold_ema_dead_code": 2
+            }
+        },
+        "latent_dim": 128,
+        "downsampling_ratio": 640,
+        "io_channels": 1
+    },
+    "training": {
+        "learning_rate": 1e-4,
+        "warmup_steps": 0,
+        "use_ema": true,
+        "loss_configs": {
+            "discriminator": {
+                "type": "encodec",
+                "config": {
+                    "filters": 32,
+                    "n_ffts": [2048, 1024, 512, 256, 128],
+                    "hop_lengths": [512, 256, 128, 64, 32],
+                    "win_lengths": [2048, 1024, 512, 256, 128]
+                },
+                "weights": {
+                    "adversarial": 0.1,
+                    "feature_matching": 5.0
+                }
+            },
+            "spectral": {
+                "type": "mrstft",
+                "config": {
+                    "fft_sizes": [2048, 1024, 512, 256, 128, 64, 32],
+                    "hop_sizes": [512, 256, 128, 64, 32, 16, 8],
+                    "win_lengths": [2048, 1024, 512, 256, 128, 64, 32],
+                    "perceptual_weighting": true
+                },
+                "weights": {
+                    "mrstft": 1.0
+                }
+            },
+            "time": {
+                "type": "l1",
+                "weights": {
+                    "l1": 0.0
+                }
+            }
+        },
+        "demo": {
+            "demo_every": 2000
+        }
+    }
+}

stable_audio_tools/configs/model_configs/autoencoders/stable_audio_1_0_vae.json

@@ -0,0 +1,111 @@
+{
+    "model_type": "autoencoder",
+    "sample_size": 65536,
+    "sample_rate": 44100,
+    "audio_channels": 2,
+    "model": {
+        "encoder": {
+            "type": "dac",
+            "config": {
+                "in_channels": 2,
+                "latent_dim": 128,
+                "d_model": 128,
+                "strides": [4, 4, 8, 8]
+            }
+        },
+        "decoder": {
+            "type": "dac",
+            "config": {
+                "out_channels": 2,
+                "latent_dim": 64,
+                "channels": 1536,
+                "rates": [8, 8, 4, 4]
+            }
+        },
+        "bottleneck": {
+            "type": "vae"
+        },
+        "latent_dim": 64,
+        "downsampling_ratio": 1024,
+        "io_channels": 2
+    },
+    "training": {
+        "learning_rate": 1e-4,
+        "warmup_steps": 0,
+        "use_ema": true,
+        "optimizer_configs": {
+            "autoencoder": {
+                "optimizer": {
+                    "type": "AdamW",
+                    "config": {
+                        "betas": [0.8, 0.99],
+                        "lr": 1e-4
+                    }
+                },
+                "scheduler": {
+                    "type": "ExponentialLR",
+                    "config": {
+                        "gamma": 0.999996
+                    }
+                }
+            },
+            "discriminator": {
+                "optimizer": {
+                    "type": "AdamW",
+                    "config": {
+                        "betas": [0.8, 0.99],
+                        "lr": 1e-4
+                    }
+                },
+                "scheduler": {
+                    "type": "ExponentialLR",
+                    "config": {
+                        "gamma": 0.999996
+                    }
+                }
+            }
+        },
+        "loss_configs": {
+            "discriminator": {
+                "type": "encodec",
+                "config": {
+                    "filters": 32,
+                    "n_ffts": [2048, 1024, 512, 256, 128],
+                    "hop_lengths": [512, 256, 128, 64, 32],
+                    "win_lengths": [2048, 1024, 512, 256, 128]
+                },
+                "weights": {
+                    "adversarial": 0.1,
+                    "feature_matching": 5.0
+                }
+            },
+            "spectral": {
+                "type": "mrstft",
+                "config": {
+                    "fft_sizes": [2048, 1024, 512, 256, 128, 64, 32],
+                    "hop_sizes": [512, 256, 128, 64, 32, 16, 8],
+                    "win_lengths": [2048, 1024, 512, 256, 128, 64, 32],
+                    "perceptual_weighting": true
+                },
+                "weights": {
+                    "mrstft": 1.0
+                }
+            },
+            "time": {
+                "type": "l1",
+                "weights": {
+                    "l1": 0.0
+                }
+            },
+            "bottleneck": {
+                "type": "kl",
+                "weights": {
+                    "kl": 1e-6
+                }
+            }
+        },
+        "demo": {
+            "demo_every": 2000
+        }
+    }
+}

stable_audio_tools/configs/model_configs/autoencoders/stable_audio_2_0_vae.json

@@ -0,0 +1,122 @@
+{
+    "model_type": "autoencoder",
+    "sample_size": 65536,
+    "sample_rate": 44100,
+    "audio_channels": 2,
+    "model": {
+        "encoder": {
+            "type": "oobleck",
+            "config": {
+                "in_channels": 2,
+                "channels": 128,
+                "c_mults": [1, 2, 4, 8, 16],
+                "strides": [2, 4, 4, 8, 8],
+                "latent_dim": 128,
+                "use_snake": true
+            }
+        },
+        "decoder": {
+            "type": "oobleck",
+            "config": {
+                "out_channels": 2,
+                "channels": 128,
+                "c_mults": [1, 2, 4, 8, 16],
+                "strides": [2, 4, 4, 8, 8],
+                "latent_dim": 64,
+                "use_snake": true,
+                "final_tanh": false
+            }
+        },
+        "bottleneck": {
+            "type": "vae"
+        },
+        "latent_dim": 64,
+        "downsampling_ratio": 2048,
+        "io_channels": 2
+    },
+    "training": {
+        "learning_rate": 1.5e-4,
+        "warmup_steps": 0,
+        "use_ema": true,
+        "optimizer_configs": {
+            "autoencoder": {
+                "optimizer": {
+                    "type": "AdamW",
+                    "config": {
+                        "betas": [0.8, 0.99],
+                        "lr": 1.5e-4,
+                        "weight_decay": 1e-3
+                    }
+                },
+                "scheduler": {
+                    "type": "InverseLR",
+                    "config": {
+                        "inv_gamma": 200000,
+                        "power": 0.5,
+                        "warmup": 0.999
+                    }
+                }
+            },
+            "discriminator": {
+                "optimizer": {
+                    "type": "AdamW",
+                    "config": {
+                        "betas": [0.8, 0.99],
+                        "lr": 3e-4,
+                        "weight_decay": 1e-3
+                    }
+                },
+                "scheduler": {
+                    "type": "InverseLR",
+                    "config": {
+                        "inv_gamma": 200000,
+                        "power": 0.5,
+                        "warmup": 0.999
+                    }
+                }
+            }
+        },
+        "loss_configs": {
+            "discriminator": {
+                "type": "encodec",
+                "config": {
+                    "filters": 64,
+                    "n_ffts": [2048, 1024, 512, 256, 128],
+                    "hop_lengths": [512, 256, 128, 64, 32],
+                    "win_lengths": [2048, 1024, 512, 256, 128]
+                },
+                "weights": {
+                    "adversarial": 0.1,
+                    "feature_matching": 5.0
+                }
+            },
+            "spectral": {
+                "type": "mrstft",
+                "config": {
+                    "fft_sizes": [2048, 1024, 512, 256, 128, 64, 32],
+                    "hop_sizes": [512, 256, 128, 64, 32, 16, 8],
+                    "win_lengths": [2048, 1024, 512, 256, 128, 64, 32],
+                    "perceptual_weighting": true
+                },
+                "weights": {
+                    "mrstft": 1.0
+                }
+            },
+            "time": {
+                "type": "l1",
+                "weights": {
+                    "l1": 0.0
+                }
+            },
+            "bottleneck": {
+                "type": "kl",
+                "weights": {
+                    "kl": 1e-4
+                }
+            }
+        },
+        "demo": {
+            "demo_every": 2000
+        }
+    }
+}

stable_audio_tools/configs/model_configs/dance_diffusion/dance_diffusion_base.json

@@ -0,0 +1,18 @@
+{
+    "model_type": "diffusion_uncond",
+    "sample_size": 65536,
+    "sample_rate": 48000,
+    "model": {
+        "type": "DAU1d",
+        "config": {
+            "n_attn_layers": 5
+        }
+    },
+    "training": {
+        "learning_rate": 1e-4,
+        "demo": {
+            "demo_every": 2000,
+            "demo_steps": 250
+        }
+    }
+}

stable_audio_tools/configs/model_configs/dance_diffusion/dance_diffusion_base_16k.json

@@ -0,0 +1,18 @@
+{
+    "model_type": "diffusion_uncond",
+    "sample_size": 65536,
+    "sample_rate": 16000,
+    "model": {
+        "type": "DAU1d",
+        "config": {
+            "n_attn_layers": 5
+        }
+    },
+    "training": {
+        "learning_rate": 1e-4,
+        "demo": {
+            "demo_every": 2000,
+            "demo_steps": 250
+        }
+    }
+}

stable_audio_tools/configs/model_configs/dance_diffusion/dance_diffusion_base_44k.json

@@ -0,0 +1,18 @@
+{
+    "model_type": "diffusion_uncond",
+    "sample_size": 65536,
+    "sample_rate": 44100,
+    "model": {
+        "type": "DAU1d",
+        "config": {
+            "n_attn_layers": 5
+        }
+    },
+    "training": {
+        "learning_rate": 4e-5,
+        "demo": {
+            "demo_every": 2000,
+            "demo_steps": 250
+        }
+    }
+}

stable_audio_tools/configs/model_configs/dance_diffusion/dance_diffusion_large.json

@@ -0,0 +1,18 @@
+{
+    "model_type": "diffusion_uncond",
+    "sample_size": 131072,
+    "sample_rate": 48000,
+    "model": {
+        "type": "DAU1d",
+        "config": {
+            "n_attn_layers": 5
+        }
+    },
+    "training": {
+        "learning_rate": 1e-4,
+        "demo": {
+            "demo_every": 2000,
+            "demo_steps": 250
+        }
+    }
+}

stable_audio_tools/configs/model_configs/txt2audio/musicgen_small_finetune.json

@@ -0,0 +1,22 @@
+{
+    "model_type": "musicgen",
+    "sample_size": 320000,
+    "sample_rate": 32000,
+    "audio_channels": 1,
+    "model": {
+        "pretrained": "small"
+    },
+    "training": {
+        "learning_rate": 1e-4,
+        "demo": {
+            "demo_every": 2000,
+            "demo_cond": [
+                {"prompt": "Keywords: Atmospheres, Orchestral Drone, Bass, Sci-Fi Ambient Soundscape, Synthesiser, Middle Eastern Vocal, dramatic piano"},
+                {"prompt": "Genre: Corporate|Instruments: Ukulele, Drums, Clapping, Glockenspiel"},
+                {"prompt": "Description: 116 BPM rock drums, drum track for a rock song"},
+                {"prompt": "A grand orchestral arrangement with thunderous percussion, epic brass fanfares, and soaring strings, creating a cinematic atmosphere fit for a heroic battle."}
+            ],
+            "demo_cfg_scales": [3, 6, 9]
+        }
+    }
+}

stable_audio_tools/configs/model_configs/txt2audio/stable_audio_1_0.json

@@ -0,0 +1,107 @@
+{
+    "model_type": "diffusion_cond",
+    "sample_size": 4194304,
+    "sample_rate": 44100,
+    "audio_channels": 2,
+    "model": {
+        "pretransform": {
+            "type": "autoencoder",
+            "iterate_batch": true,
+            "config": {
+                "encoder": {
+                    "type": "dac",
+                    "config": {
+                        "in_channels": 2,
+                        "latent_dim": 128,
+                        "d_model": 128,
+                        "strides": [4, 4, 8, 8]
+                    }
+                },
+                "decoder": {
+                    "type": "dac",
+                    "config": {
+                        "out_channels": 2,
+                        "latent_dim": 64,
+                        "channels": 1536,
+                        "rates": [8, 8, 4, 4]
+                    }
+                },
+                "bottleneck": {
+                    "type": "vae"
+                },
+                "latent_dim": 64,
+                "downsampling_ratio": 1024,
+                "io_channels": 2
+            }
+        },
+        "conditioning": {
+            "configs": [
+                {
+                    "id": "prompt",
+                    "type": "clap_text",
+                    "config": {
+                        "audio_model_type": "HTSAT-base",
+                        "enable_fusion": true,
+                        "clap_ckpt_path": "/path/to/clap.ckpt",
+                        "use_text_features": true,
+                        "feature_layer_ix": -2
+                    }
+                },
+                {
+                    "id": "seconds_start",
+                    "type": "int",
+                    "config": {
+                        "min_val": 0,
+                        "max_val": 512
+                    }
+                },
+                {
+                    "id": "seconds_total",
+                    "type": "int",
+                    "config": {
+                        "min_val": 0,
+                        "max_val": 512
+                    }
+                }
+            ],
+            "cond_dim": 768
+        },
+        "diffusion": {
+            "type": "adp_cfg_1d",
+            "cross_attention_cond_ids": ["prompt", "seconds_start", "seconds_total"],
+            "config": {
+                "in_channels": 64,
+                "context_embedding_features": 768,
+                "context_embedding_max_length": 79,
+                "channels": 256,
+                "resnet_groups": 16,
+                "kernel_multiplier_downsample": 2,
+                "multipliers": [4, 4, 4, 5, 5],
+                "factors": [1, 2, 2, 4],
+                "num_blocks": [2, 2, 2, 2],
+                "attentions": [1, 3, 3, 3, 3],
+                "attention_heads": 16,
+                "attention_multiplier": 4,
+                "use_nearest_upsample": false,
+                "use_skip_scale": true,
+                "use_context_time": true
+            }
+        },
+        "io_channels": 64
+    },
+    "training": {
+        "learning_rate": 4e-5,
+        "demo": {
+            "demo_every": 2000,
+            "demo_steps": 250,
+            "num_demos": 4,
+            "demo_cond": [
+                {"prompt": "A beautiful piano arpeggio", "seconds_start": 0, "seconds_total": 95},
+                {"prompt": "A tropical house track with upbeat melodies, a driving bassline, and cheery vibes", "seconds_start": 0, "seconds_total": 90},
+                {"prompt": "A cool 80s glam rock song with driving drums and distorted guitars", "seconds_start": 0, "seconds_total": 180},
+                {"prompt": "A grand orchestral arrangement with thunderous percussion, epic brass fanfares, and soaring strings, creating a cinematic atmosphere fit for a heroic battle.", "seconds_start": 0, "seconds_total": 60}
+            ],
+            "demo_cfg_scales": [3, 6, 9]
+        }
+    }
+}

stable_audio_tools/configs/model_configs/txt2audio/stable_audio_2_0.json

@@ -0,0 +1,127 @@
+{
+    "model_type": "diffusion_cond",
+    "sample_size": 12582912,
+    "sample_rate": 44100,
+    "audio_channels": 2,
+    "model": {
+        "pretransform": {
+            "type": "autoencoder",
+            "iterate_batch": true,
+            "config": {
+                "encoder": {
+                    "type": "oobleck",
+                    "config": {
+                        "in_channels": 2,
+                        "channels": 128,
+                        "c_mults": [1, 2, 4, 8, 16],
+                        "strides": [2, 4, 4, 8, 8],
+                        "latent_dim": 128,
+                        "use_snake": true
+                    }
+                },
+                "decoder": {
+                    "type": "oobleck",
+                    "config": {
+                        "out_channels": 2,
+                        "channels": 128,
+                        "c_mults": [1, 2, 4, 8, 16],
+                        "strides": [2, 4, 4, 8, 8],
+                        "latent_dim": 64,
+                        "use_snake": true,
+                        "final_tanh": false
+                    }
+                },
+                "bottleneck": {
+                    "type": "vae"
+                },
+                "latent_dim": 64,
+                "downsampling_ratio": 2048,
+                "io_channels": 2
+            }
+        },
+        "conditioning": {
+            "configs": [
+                {
+                    "id": "prompt",
+                    "type": "clap_text",
+                    "config": {
+                        "audio_model_type": "HTSAT-base",
+                        "enable_fusion": true,
+                        "clap_ckpt_path": "/path/to/clap.ckpt",
+                        "use_text_features": true,
+                        "feature_layer_ix": -2
+                    }
+                },
+                {
+                    "id": "seconds_start",
+                    "type": "number",
+                    "config": {
+                        "min_val": 0,
+                        "max_val": 512
+                    }
+                },
+                {
+                    "id": "seconds_total",
+                    "type": "number",
+                    "config": {
+                        "min_val": 0,
+                        "max_val": 512
+                    }
+                }
+            ],
+            "cond_dim": 768
+        },
+        "diffusion": {
+            "cross_attention_cond_ids": ["prompt", "seconds_start", "seconds_total"],
+            "global_cond_ids": ["seconds_start", "seconds_total"],
+            "type": "dit",
+            "config": {
+                "io_channels": 64,
+                "embed_dim": 1536,
+                "depth": 24,
+                "num_heads": 24,
+                "cond_token_dim": 768,
+                "global_cond_dim": 1536,
+                "project_cond_tokens": false,
+                "transformer_type": "continuous_transformer"
+            }
+        },
+        "io_channels": 64
+    },
+    "training": {
+        "use_ema": true,
+        "log_loss_info": false,
+        "optimizer_configs": {
+            "diffusion": {
+                "optimizer": {
+                    "type": "AdamW",
+                    "config": {
+                        "lr": 5e-5,
+                        "betas": [0.9, 0.999],
+                        "weight_decay": 1e-3
+                    }
+                },
+                "scheduler": {
+                    "type": "InverseLR",
+                    "config": {
+                        "inv_gamma": 1000000,
+                        "power": 0.5,
+                        "warmup": 0.99
+                    }
+                }
+            }
+        },
+        "demo": {
+            "demo_every": 2000,
+            "demo_steps": 250,
+            "num_demos": 4,
+            "demo_cond": [
+                {"prompt": "A beautiful piano arpeggio", "seconds_start": 0, "seconds_total": 80},
+                {"prompt": "A tropical house track with upbeat melodies, a driving bassline, and cheery vibes", "seconds_start": 0, "seconds_total": 250},
+                {"prompt": "A cool 80s glam rock song with driving drums and distorted guitars", "seconds_start": 0, "seconds_total": 180},
+                {"prompt": "A grand orchestral arrangement with thunderous percussion, epic brass fanfares, and soaring strings, creating a cinematic atmosphere fit for a heroic battle.", "seconds_start": 0, "seconds_total": 190}
+            ],
+            "demo_cfg_scales": [3, 6, 9]
+        }
+    }
+}

stable_audio_tools/data/dataset.py

@@ -0,0 +1,597 @@
+import importlib
+import numpy as np
+import io
+import os
+import posixpath
+import random
+import re
+import subprocess
+import time
+import torch
+import torchaudio
+import webdataset as wds
+
+from aeiou.core import is_silence
+from os import path
+from pedalboard.io import AudioFile
+from torchaudio import transforms as T
+from typing import Optional, Callable, List
+
+from .utils import Stereo, Mono, PhaseFlipper, PadCrop_Normalized_T
+
+AUDIO_KEYS = ("flac", "wav", "mp3", "m4a", "ogg", "opus")
+
+# fast_scandir implementation by Scott Hawley originally in https://github.com/zqevans/audio-diffusion/blob/main/dataset/dataset.py
+
+def fast_scandir(
+    dir:str,  # top-level directory at which to begin scanning
+    ext:list,  # list of allowed file extensions,
+    #max_size = 1 * 1000 * 1000 * 1000 # Only files < 1 GB
+    ):
+    "very fast `glob` alternative. from https://stackoverflow.com/a/59803793/4259243"
+    subfolders, files = [], []
+    ext = ['.'+x if x[0]!='.' else x for x in ext]  # add starting period to extensions if needed
+    try: # hope to avoid 'permission denied' by this try
+        for f in os.scandir(dir):
+            try: # 'hope to avoid too many levels of symbolic links' error
+                if f.is_dir():
+                    subfolders.append(f.path)
+                elif f.is_file():
+                    file_ext = os.path.splitext(f.name)[1].lower()
+                    is_hidden = os.path.basename(f.path).startswith(".")
+
+                    if file_ext in ext and not is_hidden:
+                        files.append(f.path)
+            except:
+                pass 
+    except:
+        pass
+
+    for dir in list(subfolders):
+        sf, f = fast_scandir(dir, ext)
+        subfolders.extend(sf)
+        files.extend(f)
+    return subfolders, files
+
+def keyword_scandir(
+    dir: str,  # top-level directory at which to begin scanning
+    ext: list,  # list of allowed file extensions
+    keywords: list,  # list of keywords to search for in the file name
+):
+    "very fast `glob` alternative. from https://stackoverflow.com/a/59803793/4259243"
+    subfolders, files = [], []
+    # make keywords case insensitive
+    keywords = [keyword.lower() for keyword in keywords]
+    # add starting period to extensions if needed
+    ext = ['.'+x if x[0] != '.' else x for x in ext]
+    banned_words = ["paxheader", "__macosx"]
+    try:  # hope to avoid 'permission denied' by this try
+        for f in os.scandir(dir):
+            try:  # 'hope to avoid too many levels of symbolic links' error
+                if f.is_dir():
+                    subfolders.append(f.path)
+                elif f.is_file():
+                    is_hidden = f.name.split("/")[-1][0] == '.'
+                    has_ext = os.path.splitext(f.name)[1].lower() in ext
+                    name_lower = f.name.lower()
+                    has_keyword = any(
+                        [keyword in name_lower for keyword in keywords])
+                    has_banned = any(
+                        [banned_word in name_lower for banned_word in banned_words])
+                    if has_ext and has_keyword and not has_banned and not is_hidden and not os.path.basename(f.path).startswith("._"):
+                        files.append(f.path)
+            except:
+                pass
+    except:
+        pass
+
+    for dir in list(subfolders):
+        sf, f = keyword_scandir(dir, ext, keywords)
+        subfolders.extend(sf)
+        files.extend(f)
+    return subfolders, files
+
+def get_audio_filenames(
+    paths: list,  # directories in which to search
+    keywords=None,
+    exts=['.wav', '.mp3', '.flac', '.ogg', '.aif', '.opus']
+):
+    "recursively get a list of audio filenames"
+    filenames = []
+    if type(paths) is str:
+        paths = [paths]
+    for path in paths:               # get a list of relevant filenames
+        if keywords is not None:
+            subfolders, files = keyword_scandir(path, exts, keywords)
+        else:
+            subfolders, files = fast_scandir(path, exts)
+        filenames.extend(files)
+    return filenames
+
+class SampleDataset(torch.utils.data.Dataset):
+    def __init__(
+        self, 
+        paths, 
+        sample_size=65536, 
+        sample_rate=48000, 
+        keywords=None, 
+        relpath=None, 
+        random_crop=True,
+        force_channels="stereo",
+        custom_metadata_fn: Optional[Callable[[str], str]] = None
+    ):
+        super().__init__()
+        self.filenames = []
+        self.relpath = relpath
+
+        self.augs = torch.nn.Sequential(
+            PhaseFlipper(),
+        )
+
+        self.pad_crop = PadCrop_Normalized_T(sample_size, sample_rate, randomize=random_crop)
+
+        self.force_channels = force_channels
+
+        self.encoding = torch.nn.Sequential(
+            Stereo() if self.force_channels == "stereo" else torch.nn.Identity(),
+            Mono() if self.force_channels == "mono" else torch.nn.Identity(),
+        )
+
+        self.filenames = get_audio_filenames(paths, keywords)
+
+        print(f'Found {len(self.filenames)} files')
+
+        self.sr = sample_rate
+
+        self.custom_metadata_fn = custom_metadata_fn
+
+    def load_file(self, filename):
+        ext = filename.split(".")[-1]
+
+        if ext == "mp3":
+            with AudioFile(filename) as f:
+                audio = f.read(f.frames)
+                audio = torch.from_numpy(audio)
+                in_sr = f.samplerate
+        else:
+            audio, in_sr = torchaudio.load(filename, format=ext)
+
+        if in_sr != self.sr:
+            resample_tf = T.Resample(in_sr, self.sr)
+            audio = resample_tf(audio)
+
+        return audio
+
+    def __len__(self):
+        return len(self.filenames)
+
+    def __getitem__(self, idx):
+        audio_filename = self.filenames[idx]
+        try:
+            start_time = time.time()
+            audio = self.load_file(audio_filename)
+
+            audio, t_start, t_end, seconds_start, seconds_total, padding_mask = self.pad_crop(audio)
+
+            # Run augmentations on this sample (including random crop)
+            if self.augs is not None:
+                audio = self.augs(audio)
+
+            audio = audio.clamp(-1, 1)
+
+            # Encode the file to assist in prediction
+            if self.encoding is not None:
+                audio = self.encoding(audio)
+
+            info = {}
+
+            info["path"] = audio_filename
+
+            if self.relpath is not None:
+                info["relpath"] = path.relpath(audio_filename, self.relpath)
+
+            info["timestamps"] = (t_start, t_end)
+            info["seconds_start"] = seconds_start
+            info["seconds_total"] = seconds_total
+            info["padding_mask"] = padding_mask
+
+            end_time = time.time()
+
+            info["load_time"] = end_time - start_time
+
+            if self.custom_metadata_fn is not None:
+                custom_metadata = self.custom_metadata_fn(info, audio)
+                info.update(custom_metadata)
+
+                if "__reject__" in info and info["__reject__"]:
+                    return self[random.randrange(len(self))]
+
+            return (audio, info)
+        except Exception as e:
+            print(f'Couldn\'t load file {audio_filename}: {e}')
+            return self[random.randrange(len(self))]
+
+def group_by_keys(data, keys=wds.tariterators.base_plus_ext, lcase=True, suffixes=None, handler=None):
+    """Return function over iterator that groups key, value pairs into samples.
+    :param keys: function that splits the key into key and extension (base_plus_ext)
+    :param lcase: convert suffixes to lower case (Default value = True)
+    """
+    current_sample = None
+    for filesample in data:
+        assert isinstance(filesample, dict)
+        fname, value = filesample["fname"], filesample["data"]
+        prefix, suffix = keys(fname)
+        if wds.tariterators.trace:
+            print(
+                prefix,
+                suffix,
+                current_sample.keys() if isinstance(current_sample, dict) else None,
+            )
+        if prefix is None:
+            continue
+        if lcase:
+            suffix = suffix.lower()
+        if current_sample is None or prefix != current_sample["__key__"]:
+            if wds.tariterators.valid_sample(current_sample):
+                yield current_sample
+            current_sample = dict(__key__=prefix, __url__=filesample["__url__"])
+        if suffix in current_sample:
+            print(f"{fname}: duplicate file name in tar file {suffix} {current_sample.keys()}")
+        if suffixes is None or suffix in suffixes:
+            current_sample[suffix] = value
+    if wds.tariterators.valid_sample(current_sample):
+        yield current_sample
+
+wds.tariterators.group_by_keys = group_by_keys
+
+# S3 code and WDS preprocessing code based on implementation by Scott Hawley originally in https://github.com/zqevans/audio-diffusion/blob/main/dataset/dataset.py
+
+def get_s3_contents(dataset_path, s3_url_prefix=None, filter='', recursive=True, debug=False, profile=None):
+    """
+    Returns a list of full S3 paths to files in a given S3 bucket and directory path.
+    """
+    # Ensure dataset_path ends with a trailing slash
+    if dataset_path != '' and not dataset_path.endswith('/'):
+        dataset_path += '/'
+    # Use posixpath to construct the S3 URL path
+    bucket_path = posixpath.join(s3_url_prefix or '', dataset_path)
+    # Construct the `aws s3 ls` command
+    cmd = ['aws', 's3', 'ls', bucket_path]
+
+    if profile is not None:
+        cmd.extend(['--profile', profile])
+
+    if recursive:
+        # Add the --recursive flag if requested
+        cmd.append('--recursive')
+    
+    # Run the `aws s3 ls` command and capture the output
+    run_ls = subprocess.run(cmd, capture_output=True, check=True)
+    # Split the output into lines and strip whitespace from each line
+    contents = run_ls.stdout.decode('utf-8').split('\n')
+    contents = [x.strip() for x in contents if x]
+    # Remove the timestamp from lines that begin with a timestamp
+    contents = [re.sub(r'^\S+\s+\S+\s+\d+\s+', '', x)
+                if re.match(r'^\S+\s+\S+\s+\d+\s+', x) else x for x in contents]
+    # Construct a full S3 path for each file in the contents list
+    contents = [posixpath.join(s3_url_prefix or '', x)
+                for x in contents if not x.endswith('/')]
+    # Apply the filter, if specified
+    if filter:
+        contents = [x for x in contents if filter in x]
+    # Remove redundant directory names in the S3 URL
+    if recursive:
+        # Get the main directory name from the S3 URL
+        main_dir = "/".join(bucket_path.split('/')[3:])
+        # Remove the redundant directory names from each file path
+        contents = [x.replace(f'{main_dir}', '').replace(
+            '//', '/') for x in contents]
+    # Print debugging information, if requested
+    if debug:
+        print("contents = \n", contents)
+    # Return the list of S3 paths to files
+    return contents
+
+
+def get_all_s3_urls(
+    names=[],           # list of all valid [LAION AudioDataset] dataset names
+    # list of subsets you want from those datasets, e.g. ['train','valid']
+    subsets=[''],
+    s3_url_prefix=None,  # prefix for those dataset names
+    recursive=True,     # recursively list all tar files in all subdirs
+    filter_str='tar',   # only grab files with this substring
+    # print debugging info -- note: info displayed likely to change at dev's whims
+    debug=False,
+    profiles={},        # dictionary of profiles for each item in names, e.g. {'dataset1': 'profile1', 'dataset2': 'profile2'}
+):
+    "get urls of shards (tar files) for multiple datasets in one s3 bucket"
+    urls = []
+    for name in names:
+        # If s3_url_prefix is not specified, assume the full S3 path is included in each element of the names list
+        if s3_url_prefix is None:
+            contents_str = name
+        else:
+            # Construct the S3 path using the s3_url_prefix and the current name value
+            contents_str = posixpath.join(s3_url_prefix, name)
+        if debug:
+            print(f"get_all_s3_urls: {contents_str}:")
+        for subset in subsets:
+            subset_str = posixpath.join(contents_str, subset)
+            if debug:
+                print(f"subset_str = {subset_str}")
+            # Get the list of tar files in the current subset directory
+            profile = profiles.get(name, None)
+            tar_list = get_s3_contents(
+                subset_str, s3_url_prefix=None, recursive=recursive, filter=filter_str, debug=debug, profile=profile)
+            for tar in tar_list:
+                # Escape spaces and parentheses in the tar filename for use in the shell command
+                tar = tar.replace(" ", "\ ").replace(
+                    "(", "\(").replace(")", "\)")
+                # Construct the S3 path to the current tar file
+                s3_path = posixpath.join(name, subset, tar) + " -"
+                # Construct the AWS CLI command to download the current tar file
+                if s3_url_prefix is None:
+                    request_str = f"pipe:aws s3 --cli-connect-timeout 0 cp {s3_path}"
+                else:
+                    request_str = f"pipe:aws s3 --cli-connect-timeout 0 cp {posixpath.join(s3_url_prefix, s3_path)}"
+                if profiles.get(name):
+                    request_str += f" --profile {profiles.get(name)}"
+                if debug:
+                    print("request_str = ", request_str)
+                # Add the constructed URL to the list of URLs
+                urls.append(request_str)
+    return urls
+
+
+def log_and_continue(exn):
+    """Call in an exception handler to ignore any exception, isssue a warning, and continue."""
+    print(f"Handling webdataset error ({repr(exn)}). Ignoring.")
+    return True
+
+
+def is_valid_sample(sample):
+    has_json = "json" in sample
+    has_audio = "audio" in sample
+    is_silent = is_silence(sample["audio"])
+    is_rejected = "__reject__" in sample["json"] and sample["json"]["__reject__"]
+
+    return has_json and has_audio and not is_silent and not is_rejected
+
+class S3DatasetConfig:
+    def __init__(
+        self,
+        id: str,
+        s3_path: str,
+        custom_metadata_fn: Optional[Callable[[str], str]] = None,
+        profile: Optional[str] = None,
+    ):
+        self.id = id
+        self.s3_path = s3_path
+        self.custom_metadata_fn = custom_metadata_fn
+        self.profile = profile
+        self.urls = []
+
+    def load_data_urls(self):
+        self.urls = get_all_s3_urls(
+            names=[self.s3_path],
+            s3_url_prefix=None,
+            recursive=True,
+            profiles={self.s3_path: self.profile} if self.profile else {},
+        )
+
+        return self.urls
+
+def audio_decoder(key, value):
+    # Get file extension from key
+    ext = key.split(".")[-1]
+
+    if ext in AUDIO_KEYS:
+        return torchaudio.load(io.BytesIO(value))
+    else:
+        return None
+
+def collation_fn(samples):
+        batched = list(zip(*samples))
+        result = []
+        for b in batched:
+            if isinstance(b[0], (int, float)):
+                b = np.array(b)
+            elif isinstance(b[0], torch.Tensor):
+                b = torch.stack(b)
+            elif isinstance(b[0], np.ndarray):
+                b = np.array(b)
+            else:
+                b = b
+            result.append(b)
+        return result
+
+class S3WebDataLoader():
+    def __init__(
+        self,
+        datasets: List[S3DatasetConfig],
+        batch_size,
+        sample_size,
+        sample_rate=48000,
+        num_workers=8,
+        epoch_steps=1000,
+        random_crop=True,
+        force_channels="stereo",
+        augment_phase=True,
+        **data_loader_kwargs
+    ):
+
+        self.datasets = datasets
+
+        self.sample_size = sample_size
+        self.sample_rate = sample_rate
+        self.random_crop = random_crop
+        self.force_channels = force_channels
+        self.augment_phase = augment_phase
+
+        urls = [dataset.load_data_urls() for dataset in datasets]
+
+        # Flatten the list of lists of URLs
+        urls = [url for dataset_urls in urls for url in dataset_urls]
+
+        self.dataset = wds.DataPipeline(
+            wds.ResampledShards(urls),
+            wds.tarfile_to_samples(handler=log_and_continue),
+            wds.decode(audio_decoder, handler=log_and_continue),
+            wds.map(self.wds_preprocess, handler=log_and_continue),
+            wds.select(is_valid_sample),
+            wds.to_tuple("audio", "json", handler=log_and_continue),
+            wds.batched(batch_size, partial=False, collation_fn=collation_fn),
+        ).with_epoch(epoch_steps//num_workers if num_workers > 0 else epoch_steps)
+
+        self.data_loader = wds.WebLoader(self.dataset, num_workers=num_workers, **data_loader_kwargs)
+
+    def wds_preprocess(self, sample):
+
+        found_key, rewrite_key = '', ''
+        for k, v in sample.items():  # print the all entries in dict
+            for akey in AUDIO_KEYS:
+                if k.endswith(akey):
+                    # to rename long/weird key with its simpler counterpart
+                    found_key, rewrite_key = k, akey
+                    break
+            if '' != found_key:
+                break
+        if '' == found_key:  # got no audio!
+            return None  # try returning None to tell WebDataset to skip this one
+
+        audio, in_sr = sample[found_key]
+        if in_sr != self.sample_rate:
+            resample_tf = T.Resample(in_sr, self.sample_rate)
+            audio = resample_tf(audio)
+
+        if self.sample_size is not None:
+            # Pad/crop and get the relative timestamp
+            pad_crop = PadCrop_Normalized_T(
+                self.sample_size, randomize=self.random_crop, sample_rate=self.sample_rate)
+            audio, t_start, t_end, seconds_start, seconds_total, padding_mask = pad_crop(
+                audio)
+            sample["json"]["seconds_start"] = seconds_start
+            sample["json"]["seconds_total"] = seconds_total
+            sample["json"]["padding_mask"] = padding_mask
+        else:
+            t_start, t_end = 0, 1
+
+        # Check if audio is length zero, initialize to a single zero if so
+        if audio.shape[-1] == 0:
+            audio = torch.zeros(1, 1)
+
+        # Make the audio stereo and augment by randomly inverting phase
+        augs = torch.nn.Sequential(
+            Stereo() if self.force_channels == "stereo" else torch.nn.Identity(),
+            Mono() if self.force_channels == "mono" else torch.nn.Identity(),
+            PhaseFlipper() if self.augment_phase else torch.nn.Identity()
+        )
+
+        audio = augs(audio)
+
+        sample["json"]["timestamps"] = (t_start, t_end)
+
+        if "text" in sample["json"]:
+            sample["json"]["prompt"] = sample["json"]["text"]
+
+        # Check for custom metadata functions
+        for dataset in self.datasets:
+            if dataset.custom_metadata_fn is None:
+                continue
+        
+            if dataset.s3_path in sample["__url__"]:
+                custom_metadata = dataset.custom_metadata_fn(sample["json"], audio)
+                sample["json"].update(custom_metadata)
+
+        if found_key != rewrite_key:   # rename long/weird key with its simpler counterpart
+            del sample[found_key]
+
+        sample["audio"] = audio
+
+        # Add audio to the metadata as well for conditioning
+        sample["json"]["audio"] = audio
+        
+        return sample
+
+def create_dataloader_from_config(dataset_config, batch_size, sample_size, sample_rate, audio_channels=2, num_workers=4):
+
+    dataset_type = dataset_config.get("dataset_type", None)
+
+    assert dataset_type is not None, "Dataset type must be specified in dataset config"
+
+    if audio_channels == 1:
+        force_channels = "mono"
+    else:
+        force_channels = "stereo"
+
+    if dataset_type == "audio_dir":
+
+        audio_dir_configs = dataset_config.get("datasets", None)
+
+        assert audio_dir_configs is not None, "Directory configuration must be specified in datasets[\"dataset\"]"
+
+        training_dirs = []
+
+        custom_metadata_fn = None
+        custom_metadata_module_path = dataset_config.get("custom_metadata_module", None)
+
+        if custom_metadata_module_path is not None:
+            spec = importlib.util.spec_from_file_location("metadata_module", custom_metadata_module_path)
+            metadata_module = importlib.util.module_from_spec(spec)
+            spec.loader.exec_module(metadata_module)                
+
+            custom_metadata_fn = metadata_module.get_custom_metadata
+
+        for audio_dir_config in audio_dir_configs:
+            audio_dir_path = audio_dir_config.get("path", None)
+            assert audio_dir_path is not None, "Path must be set for local audio directory configuration"
+            training_dirs.append(audio_dir_path)
+
+        train_set = SampleDataset(
+            training_dirs,
+            sample_rate=sample_rate,
+            sample_size=sample_size,
+            random_crop=dataset_config.get("random_crop", True),
+            force_channels=force_channels,
+            custom_metadata_fn=custom_metadata_fn,
+            relpath=training_dirs[0] #TODO: Make relpath relative to each training dir
+        )
+
+        return torch.utils.data.DataLoader(train_set, batch_size, shuffle=True,
+                                num_workers=num_workers, persistent_workers=True, pin_memory=True, drop_last=True, collate_fn=collation_fn)
+
+    elif dataset_type == "s3":
+        dataset_configs = []
+
+        for s3_config in dataset_config["datasets"]:
+
+            custom_metadata_fn = None
+            custom_metadata_module_path = s3_config.get("custom_metadata_module", None)
+
+            if custom_metadata_module_path is not None:
+                spec = importlib.util.spec_from_file_location("metadata_module", custom_metadata_module_path)
+                metadata_module = importlib.util.module_from_spec(spec)
+                spec.loader.exec_module(metadata_module)                
+
+                custom_metadata_fn = metadata_module.get_custom_metadata
+
+            dataset_configs.append(
+                S3DatasetConfig(
+                    id=s3_config["id"],
+                    s3_path=s3_config["s3_path"],
+                    custom_metadata_fn=custom_metadata_fn,
+                    profile=s3_config.get("profile", None),
+                )
+            )
+
+        return S3WebDataLoader(
+            dataset_configs,
+            sample_rate=sample_rate,
+            sample_size=sample_size,
+            batch_size=batch_size,
+            random_crop=dataset_config.get("random_crop", True),
+            num_workers=num_workers,
+            persistent_workers=True,
+            force_channels=force_channels,
+            epoch_steps=dataset_config.get("epoch_steps", 2000),
+        ).data_loader

stable_audio_tools/data/utils.py

@@ -0,0 +1,96 @@
+import math
+import random
+import torch
+
+from torch import nn
+from typing import Tuple
+
+class PadCrop(nn.Module):
+    def __init__(self, n_samples, randomize=True):
+        super().__init__()
+        self.n_samples = n_samples
+        self.randomize = randomize
+
+    def __call__(self, signal):
+        n, s = signal.shape
+        start = 0 if (not self.randomize) else torch.randint(0, max(0, s - self.n_samples) + 1, []).item()
+        end = start + self.n_samples
+        output = signal.new_zeros([n, self.n_samples])
+        output[:, :min(s, self.n_samples)] = signal[:, start:end]
+        return output
+
+class PadCrop_Normalized_T(nn.Module):
+    
+    def __init__(self, n_samples: int, sample_rate: int, randomize: bool = True):
+        
+        super().__init__()
+        
+        self.n_samples = n_samples
+        self.sample_rate = sample_rate
+        self.randomize = randomize
+
+    def __call__(self, source: torch.Tensor) -> Tuple[torch.Tensor, float, float, int, int]:
+        
+        n_channels, n_samples = source.shape
+        
+        # If the audio is shorter than the desired length, pad it
+        upper_bound = max(0, n_samples - self.n_samples)
+        
+        # If randomize is False, always start at the beginning of the audio
+        offset = 0
+        if(self.randomize and n_samples > self.n_samples):
+            offset = random.randint(0, upper_bound)
+
+        # Calculate the start and end times of the chunk
+        t_start = offset / (upper_bound + self.n_samples)
+        t_end = (offset + self.n_samples) / (upper_bound + self.n_samples)
+
+        # Create the chunk
+        chunk = source.new_zeros([n_channels, self.n_samples])
+
+        # Copy the audio into the chunk
+        chunk[:, :min(n_samples, self.n_samples)] = source[:, offset:offset + self.n_samples]
+        
+        # Calculate the start and end times of the chunk in seconds
+        seconds_start = math.floor(offset / self.sample_rate)
+        seconds_total = math.ceil(n_samples / self.sample_rate)
+
+        # Create a mask the same length as the chunk with 1s where the audio is and 0s where it isn't
+        padding_mask = torch.zeros([self.n_samples])
+        padding_mask[:min(n_samples, self.n_samples)] = 1
+        
+        
+        return (
+            chunk,
+            t_start,
+            t_end,
+            seconds_start,
+            seconds_total,
+            padding_mask
+        )
+
+class PhaseFlipper(nn.Module):
+    "Randomly invert the phase of a signal"
+    def __init__(self, p=0.5):
+        super().__init__()
+        self.p = p
+    def __call__(self, signal):
+        return -signal if (random.random() < self.p) else signal
+        
+class Mono(nn.Module):
+  def __call__(self, signal):
+    return torch.mean(signal, dim=0, keepdims=True) if len(signal.shape) > 1 else signal
+
+class Stereo(nn.Module):
+  def __call__(self, signal):
+    signal_shape = signal.shape
+    # Check if it's mono
+    if len(signal_shape) == 1: # s -> 2, s
+        signal = signal.unsqueeze(0).repeat(2, 1)
+    elif len(signal_shape) == 2:
+        if signal_shape[0] == 1: #1, s -> 2, s
+            signal = signal.repeat(2, 1)
+        elif signal_shape[0] > 2: #?, s -> 2,s
+            signal = signal[:2, :]    
+
+    return signal

stable_audio_tools/inference/generation.py

@@ -0,0 +1,243 @@
+import numpy as np
+import torch 
+import typing as tp
+import math 
+from torchaudio import transforms as T
+
+from .utils import prepare_audio
+from .sampling import sample, sample_k
+from ..data.utils import PadCrop
+
+def generate_diffusion_uncond(
+        model,
+        steps: int = 250,
+        batch_size: int = 1,
+        sample_size: int = 2097152,
+        seed: int = -1,
+        device: str = "cuda",
+        init_audio: tp.Optional[tp.Tuple[int, torch.Tensor]] = None,
+        init_noise_level: float = 1.0,
+        return_latents = False,
+        **sampler_kwargs
+        ) -> torch.Tensor:
+    
+    # The length of the output in audio samples 
+    audio_sample_size = sample_size
+
+    # If this is latent diffusion, change sample_size instead to the downsampled latent size
+    if model.pretransform is not None:
+        sample_size = sample_size // model.pretransform.downsampling_ratio
+        
+    # Seed
+    # The user can explicitly set the seed to deterministically generate the same output. Otherwise, use a random seed.
+    seed = seed if seed != -1 else np.random.randint(0, 2**32 - 1, dtype=np.uint32)
+    print(seed)
+    torch.manual_seed(seed)
+    # Define the initial noise immediately after setting the seed
+    noise = torch.randn([batch_size, model.io_channels, sample_size], device=device)
+
+    if init_audio is not None:
+        # The user supplied some initial audio (for inpainting or variation). Let us prepare the input audio.
+        in_sr, init_audio = init_audio
+
+        io_channels = model.io_channels
+
+        # For latent models, set the io_channels to the autoencoder's io_channels
+        if model.pretransform is not None:
+            io_channels = model.pretransform.io_channels
+
+        # Prepare the initial audio for use by the model
+        init_audio = prepare_audio(init_audio, in_sr=in_sr, target_sr=model.sample_rate, target_length=audio_sample_size, target_channels=io_channels, device=device)
+
+        # For latent models, encode the initial audio into latents
+        if model.pretransform is not None:
+            init_audio = model.pretransform.encode(init_audio)
+
+        init_audio = init_audio.repeat(batch_size, 1, 1)
+    else:
+        # The user did not supply any initial audio for inpainting or variation. Generate new output from scratch. 
+        init_audio = None
+        init_noise_level = None
+
+    # Inpainting mask
+    
+    if init_audio is not None:
+        # variations
+        sampler_kwargs["sigma_max"] = init_noise_level
+        mask = None 
+    else:
+        mask = None
+
+    # Now the generative AI part:
+    # k-diffusion denoising process go!
+    sampled = sample_k(model.model, noise, init_audio, mask, steps, **sampler_kwargs, device=device)
+
+    # Denoising process done. 
+    # If this is latent diffusion, decode latents back into audio
+    if model.pretransform is not None and not return_latents:
+        sampled = model.pretransform.decode(sampled)
+
+    # Return audio
+    return sampled
+
+
+def generate_diffusion_cond(
+        model,
+        steps: int = 250,
+        cfg_scale=6,
+        conditioning: dict = None,
+        conditioning_tensors: tp.Optional[dict] = None,
+        negative_conditioning: dict = None,
+        negative_conditioning_tensors: tp.Optional[dict] = None,
+        batch_size: int = 1,
+        sample_size: int = 2097152,
+        sample_rate: int = 48000,
+        seed: int = -1,
+        device: str = "cuda",
+        init_audio: tp.Optional[tp.Tuple[int, torch.Tensor]] = None,
+        init_noise_level: float = 1.0,
+        mask_args: dict = None,
+        return_latents = False,
+        **sampler_kwargs
+        ) -> torch.Tensor: 
+    """
+    Generate audio from a prompt using a diffusion model.
+    
+    Args:
+        model: The diffusion model to use for generation.
+        steps: The number of diffusion steps to use.
+        cfg_scale: Classifier-free guidance scale 
+        conditioning: A dictionary of conditioning parameters to use for generation.
+        conditioning_tensors: A dictionary of precomputed conditioning tensors to use for generation.
+        batch_size: The batch size to use for generation.
+        sample_size: The length of the audio to generate, in samples.
+        sample_rate: The sample rate of the audio to generate (Deprecated, now pulled from the model directly)
+        seed: The random seed to use for generation, or -1 to use a random seed.
+        device: The device to use for generation.
+        init_audio: A tuple of (sample_rate, audio) to use as the initial audio for generation.
+        init_noise_level: The noise level to use when generating from an initial audio sample.
+        return_latents: Whether to return the latents used for generation instead of the decoded audio.
+        **sampler_kwargs: Additional keyword arguments to pass to the sampler.    
+    """
+
+    # The length of the output in audio samples 
+    audio_sample_size = sample_size
+
+    # If this is latent diffusion, change sample_size instead to the downsampled latent size
+    if model.pretransform is not None:
+        sample_size = sample_size // model.pretransform.downsampling_ratio
+        
+    # Seed
+    # The user can explicitly set the seed to deterministically generate the same output. Otherwise, use a random seed.
+    seed = seed if seed != -1 else np.random.randint(0, 2**32 - 1)
+    print(seed)
+    torch.manual_seed(seed)
+    # Define the initial noise immediately after setting the seed
+    noise = torch.randn([batch_size, model.io_channels, sample_size], device=device)
+
+    # Conditioning
+    assert conditioning is not None or conditioning_tensors is not None, "Must provide either conditioning or conditioning_tensors"
+    if conditioning_tensors is None:
+        conditioning_tensors = model.conditioner(conditioning, device)
+    conditioning_tensors = model.get_conditioning_inputs(conditioning_tensors)
+
+    if negative_conditioning is not None or negative_conditioning_tensors is not None:
+        
+        if negative_conditioning_tensors is None:
+            negative_conditioning_tensors = model.conditioner(negative_conditioning, device)
+            
+        negative_conditioning_tensors = model.get_conditioning_inputs(negative_conditioning_tensors, negative=True)
+    else:
+        negative_conditioning_tensors = {}
+
+    if init_audio is not None:
+        # The user supplied some initial audio (for inpainting or variation). Let us prepare the input audio.
+        in_sr, init_audio = init_audio
+
+        io_channels = model.io_channels
+
+        # For latent models, set the io_channels to the autoencoder's io_channels
+        if model.pretransform is not None:
+            io_channels = model.pretransform.io_channels
+
+        # Prepare the initial audio for use by the model
+        init_audio = prepare_audio(init_audio, in_sr=in_sr, target_sr=model.sample_rate, target_length=audio_sample_size, target_channels=io_channels, device=device)
+
+        # For latent models, encode the initial audio into latents
+        if model.pretransform is not None:
+            init_audio = model.pretransform.encode(init_audio)
+
+        init_audio = init_audio.repeat(batch_size, 1, 1)
+    else:
+        # The user did not supply any initial audio for inpainting or variation. Generate new output from scratch. 
+        init_audio = None
+        init_noise_level = None
+        mask_args = None
+
+    # Inpainting mask
+    if init_audio is not None and mask_args is not None:
+        # Cut and paste init_audio according to cropfrom, pastefrom, pasteto
+        # This is helpful for forward and reverse outpainting
+        cropfrom = math.floor(mask_args["cropfrom"]/100.0 * sample_size)
+        pastefrom = math.floor(mask_args["pastefrom"]/100.0 * sample_size)
+        pasteto = math.ceil(mask_args["pasteto"]/100.0 * sample_size)
+        assert pastefrom < pasteto, "Paste From should be less than Paste To"
+        croplen = pasteto - pastefrom
+        if cropfrom + croplen > sample_size:
+            croplen = sample_size - cropfrom 
+        cropto = cropfrom + croplen
+        pasteto = pastefrom + croplen
+        cutpaste = init_audio.new_zeros(init_audio.shape)
+        cutpaste[:, :, pastefrom:pasteto] = init_audio[:,:,cropfrom:cropto]
+        #print(cropfrom, cropto, pastefrom, pasteto)
+        init_audio = cutpaste
+        # Build a soft mask (list of floats 0 to 1, the size of the latent) from the given args
+        mask = build_mask(sample_size, mask_args)
+        mask = mask.to(device)
+    elif init_audio is not None and mask_args is None:
+        # variations
+        sampler_kwargs["sigma_max"] = init_noise_level
+        mask = None 
+    else:
+        mask = None
+
+    # Now the generative AI part:
+    # k-diffusion denoising process go!
+    sampled = sample_k(model.model, noise, init_audio, mask, steps, **sampler_kwargs, **conditioning_tensors, **negative_conditioning_tensors, cfg_scale=cfg_scale, batch_cfg=True, rescale_cfg=True, device=device)
+
+    # v-diffusion: 
+    #sampled = sample(model.model, noise, steps, 0, **conditioning_tensors, embedding_scale=cfg_scale)
+
+    # Denoising process done. 
+    # If this is latent diffusion, decode latents back into audio
+    if model.pretransform is not None and not return_latents:
+        #cast sampled latents to pretransform dtype
+        sampled = sampled.to(next(model.pretransform.parameters()).dtype)
+        sampled = model.pretransform.decode(sampled)
+
+    # Return audio
+    return sampled
+
+# builds a softmask given the parameters
+# returns array of values 0 to 1, size sample_size, where 0 means noise / fresh generation, 1 means keep the input audio, 
+# and anything between is a mixture of old/new
+# ideally 0.5 is half/half mixture but i haven't figured this out yet
+def build_mask(sample_size, mask_args):
+    maskstart = math.floor(mask_args["maskstart"]/100.0 * sample_size)
+    maskend = math.ceil(mask_args["maskend"]/100.0 * sample_size)
+    softnessL = round(mask_args["softnessL"]/100.0 * sample_size)
+    softnessR = round(mask_args["softnessR"]/100.0 * sample_size)
+    marination = mask_args["marination"]
+    # use hann windows for softening the transition (i don't know if this is correct)
+    hannL = torch.hann_window(softnessL*2, periodic=False)[:softnessL]
+    hannR = torch.hann_window(softnessR*2, periodic=False)[softnessR:]
+    # build the mask. 
+    mask = torch.zeros((sample_size))
+    mask[maskstart:maskend] = 1
+    mask[maskstart:maskstart+softnessL] = hannL
+    mask[maskend-softnessR:maskend] = hannR
+    # marination finishes the inpainting early in the denoising schedule, and lets audio get changed in the final rounds
+    if marination > 0:        
+        mask = mask * (1-marination) 
+    #print(mask)
+    return mask

stable_audio_tools/inference/sampling.py

@@ -0,0 +1,170 @@
+import torch
+import math
+from tqdm import trange
+
+import k_diffusion as K
+
+# Define the noise schedule and sampling loop
+def get_alphas_sigmas(t):
+    """Returns the scaling factors for the clean image (alpha) and for the
+    noise (sigma), given a timestep."""
+    return torch.cos(t * math.pi / 2), torch.sin(t * math.pi / 2)
+
+def alpha_sigma_to_t(alpha, sigma):
+    """Returns a timestep, given the scaling factors for the clean image and for
+    the noise."""
+    return torch.atan2(sigma, alpha) / math.pi * 2
+
+def t_to_alpha_sigma(t):
+    """Returns the scaling factors for the clean image and for the noise, given
+    a timestep."""
+    return torch.cos(t * math.pi / 2), torch.sin(t * math.pi / 2)
+
+@torch.no_grad()
+def sample(model, x, steps, eta, **extra_args):
+    """Draws samples from a model given starting noise. v-diffusion"""
+    ts = x.new_ones([x.shape[0]])
+
+    # Create the noise schedule
+    t = torch.linspace(1, 0, steps + 1)[:-1]
+
+    alphas, sigmas = get_alphas_sigmas(t)
+
+    # The sampling loop
+    for i in trange(steps):
+
+        # Get the model output (v, the predicted velocity)
+        with torch.cuda.amp.autocast():
+            v = model(x, ts * t[i], **extra_args).float()
+
+        # Predict the noise and the denoised image
+        pred = x * alphas[i] - v * sigmas[i]
+        eps = x * sigmas[i] + v * alphas[i]
+
+        # If we are not on the last timestep, compute the noisy image for the
+        # next timestep.
+        if i < steps - 1:
+            # If eta > 0, adjust the scaling factor for the predicted noise
+            # downward according to the amount of additional noise to add
+            ddim_sigma = eta * (sigmas[i + 1]**2 / sigmas[i]**2).sqrt() * \
+                (1 - alphas[i]**2 / alphas[i + 1]**2).sqrt()
+            adjusted_sigma = (sigmas[i + 1]**2 - ddim_sigma**2).sqrt()
+
+            # Recombine the predicted noise and predicted denoised image in the
+            # correct proportions for the next step
+            x = pred * alphas[i + 1] + eps * adjusted_sigma
+
+            # Add the correct amount of fresh noise
+            if eta:
+                x += torch.randn_like(x) * ddim_sigma
+
+    # If we are on the last timestep, output the denoised image
+    return pred
+
+# Soft mask inpainting is just shrinking hard (binary) mask inpainting
+# Given a float-valued soft mask (values between 0 and 1), get the binary mask for this particular step
+def get_bmask(i, steps, mask):
+    strength = (i+1)/(steps)
+    # convert to binary mask
+    bmask = torch.where(mask<=strength,1,0)
+    return bmask
+
+def make_cond_model_fn(model, cond_fn):
+    def cond_model_fn(x, sigma, **kwargs):
+        with torch.enable_grad():
+            x = x.detach().requires_grad_()
+            denoised = model(x, sigma, **kwargs)
+            cond_grad = cond_fn(x, sigma, denoised=denoised, **kwargs).detach()
+            cond_denoised = denoised.detach() + cond_grad * K.utils.append_dims(sigma**2, x.ndim)
+        return cond_denoised
+    return cond_model_fn
+
+# Uses k-diffusion from https://github.com/crowsonkb/k-diffusion
+# init_data is init_audio as latents (if this is latent diffusion)
+# For sampling, set both init_data and mask to None
+# For variations, set init_data 
+# For inpainting, set both init_data & mask 
+def sample_k(
+        model_fn, 
+        noise, 
+        init_data=None,
+        mask=None,
+        steps=100, 
+        sampler_type="dpmpp-2m-sde", 
+        sigma_min=0.5, 
+        sigma_max=50, 
+        rho=1.0, device="cuda", 
+        callback=None, 
+        cond_fn=None,
+        **extra_args
+    ):
+
+    denoiser = K.external.VDenoiser(model_fn)
+
+    if cond_fn is not None:
+        denoiser = make_cond_model_fn(denoiser, cond_fn)
+
+    # Make the list of sigmas. Sigma values are scalars related to the amount of noise each denoising step has
+    sigmas = K.sampling.get_sigmas_polyexponential(steps, sigma_min, sigma_max, rho, device=device)
+    # Scale the initial noise by sigma 
+    noise = noise * sigmas[0]
+
+    wrapped_callback = callback
+
+    if mask is None and init_data is not None:
+        # VARIATION (no inpainting)
+        # set the initial latent to the init_data, and noise it with initial sigma
+        x = init_data + noise 
+    elif mask is not None and init_data is not None:
+        # INPAINTING
+        bmask = get_bmask(0, steps, mask)
+        # initial noising
+        input_noised = init_data + noise
+        # set the initial latent to a mix of init_data and noise, based on step 0's binary mask
+        x = input_noised * bmask + noise * (1-bmask)
+        # define the inpainting callback function (Note: side effects, it mutates x)
+        # See https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py#L596C13-L596C105
+        # callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        # This is called immediately after `denoised = model(x, sigmas[i] * s_in, **extra_args)`
+        def inpainting_callback(args):
+            i = args["i"]
+            x = args["x"]
+            sigma = args["sigma"]
+            #denoised = args["denoised"]
+            # noise the init_data input with this step's appropriate amount of noise
+            input_noised = init_data + torch.randn_like(init_data) * sigma
+            # shrinking hard mask
+            bmask = get_bmask(i, steps, mask)
+            # mix input_noise with x, using binary mask
+            new_x = input_noised * bmask + x * (1-bmask)
+            # mutate x
+            x[:,:,:] = new_x[:,:,:]
+        # wrap together the inpainting callback and the user-submitted callback. 
+        if callback is None: 
+            wrapped_callback = inpainting_callback
+        else:
+            wrapped_callback = lambda args: (inpainting_callback(args), callback(args))
+    else:
+        # SAMPLING
+        # set the initial latent to noise
+        x = noise
+
+
+    with torch.cuda.amp.autocast():
+        if sampler_type == "k-heun":
+            return K.sampling.sample_heun(denoiser, x, sigmas, disable=False, callback=wrapped_callback, extra_args=extra_args)
+        elif sampler_type == "k-lms":
+            return K.sampling.sample_lms(denoiser, x, sigmas, disable=False, callback=wrapped_callback, extra_args=extra_args)
+        elif sampler_type == "k-dpmpp-2s-ancestral":
+            return K.sampling.sample_dpmpp_2s_ancestral(denoiser, x, sigmas, disable=False, callback=wrapped_callback, extra_args=extra_args)
+        elif sampler_type == "k-dpm-2":
+            return K.sampling.sample_dpm_2(denoiser, x, sigmas, disable=False, callback=wrapped_callback, extra_args=extra_args)
+        elif sampler_type == "k-dpm-fast":
+            return K.sampling.sample_dpm_fast(denoiser, x, sigma_min, sigma_max, steps, disable=False, callback=wrapped_callback, extra_args=extra_args)
+        elif sampler_type == "k-dpm-adaptive":
+            return K.sampling.sample_dpm_adaptive(denoiser, x, sigma_min, sigma_max, rtol=0.01, atol=0.01, disable=False, callback=wrapped_callback, extra_args=extra_args)
+        elif sampler_type == "dpmpp-2m-sde":
+            return K.sampling.sample_dpmpp_2m_sde(denoiser, x, sigmas, disable=False, callback=wrapped_callback, extra_args=extra_args)
+        elif sampler_type == "dpmpp-3m-sde":
+            return K.sampling.sample_dpmpp_3m_sde(denoiser, x, sigmas, disable=False, callback=wrapped_callback, extra_args=extra_args)
+

stable_audio_tools/inference/utils.py

@@ -0,0 +1,35 @@
+from ..data.utils import PadCrop
+
+from torchaudio import transforms as T
+
+def set_audio_channels(audio, target_channels):
+    if target_channels == 1:
+        # Convert to mono
+        audio = audio.mean(1, keepdim=True)
+    elif target_channels == 2:
+        # Convert to stereo
+        if audio.shape[1] == 1:
+            audio = audio.repeat(1, 2, 1)
+        elif audio.shape[1] > 2:
+            audio = audio[:, :2, :]
+    return audio
+
+def prepare_audio(audio, in_sr, target_sr, target_length, target_channels, device):
+    
+    audio = audio.to(device)
+
+    if in_sr != target_sr:
+        resample_tf = T.Resample(in_sr, target_sr).to(device)
+        audio = resample_tf(audio)
+
+    audio = PadCrop(target_length, randomize=False)(audio)
+
+    # Add batch dimension
+    if audio.dim() == 1:
+        audio = audio.unsqueeze(0).unsqueeze(0)
+    elif audio.dim() == 2:
+        audio = audio.unsqueeze(0)
+
+    audio = set_audio_channels(audio, target_channels)
+
+    return audio

stable_audio_tools/interface/gradio.py

@@ -0,0 +1,782 @@
+import gc
+import numpy as np
+import gradio as gr
+import json 
+import torch
+import torchaudio
+
+from aeiou.viz import audio_spectrogram_image
+from einops import rearrange
+from safetensors.torch import load_file
+from torch.nn import functional as F
+from torchaudio import transforms as T
+
+from ..inference.generation import generate_diffusion_cond, generate_diffusion_uncond
+from ..models.factory import create_model_from_config
+from ..models.pretrained import get_pretrained_model
+from ..models.utils import load_ckpt_state_dict
+from ..inference.utils import prepare_audio
+from ..training.utils import copy_state_dict
+
+model = None
+sample_rate = 44100
+sample_size = 524288
+
+def load_model(model_config=None, model_ckpt_path=None, pretrained_name=None, pretransform_ckpt_path=None, device="cuda", model_half=False):
+    global model, sample_rate, sample_size
+    
+    if pretrained_name is not None:
+        print(f"Loading pretrained model {pretrained_name}")
+        model, model_config = get_pretrained_model(pretrained_name)
+
+    elif model_config is not None and model_ckpt_path is not None:
+        print(f"Creating model from config")
+        model = create_model_from_config(model_config)
+
+        print(f"Loading model checkpoint from {model_ckpt_path}")
+        # Load checkpoint
+        copy_state_dict(model, load_ckpt_state_dict(model_ckpt_path))
+        #model.load_state_dict(load_ckpt_state_dict(model_ckpt_path))
+
+    sample_rate = model_config["sample_rate"]
+    sample_size = model_config["sample_size"]
+
+    if pretransform_ckpt_path is not None:
+        print(f"Loading pretransform checkpoint from {pretransform_ckpt_path}")
+        model.pretransform.load_state_dict(load_ckpt_state_dict(pretransform_ckpt_path), strict=False)
+        print(f"Done loading pretransform")
+
+    model.to(device).eval().requires_grad_(False)
+
+    if model_half:
+        model.to(torch.float16)
+        
+    print(f"Done loading model")
+
+    return model, model_config
+
+def generate_cond(
+        prompt,
+        negative_prompt=None,
+        seconds_start=0,
+        seconds_total=30,
+        latitude = 0.0,
+        longitude = 0.0,
+        temperature = 0.0,
+        humidity = 0.0,
+        wind_speed = 0.0,
+        pressure = 0.0,
+        minutes_of_day = 0.0,
+        day_of_year = 0.0,
+        cfg_scale=6.0,
+        steps=250,
+        preview_every=None,
+        seed=-1,
+        sampler_type="dpmpp-2m-sde",
+        sigma_min=0.03,
+        sigma_max=50,
+        cfg_rescale=0.4,
+        use_init=False,
+        init_audio=None,
+        init_noise_level=1.0,
+        mask_cropfrom=None,
+        mask_pastefrom=None,
+        mask_pasteto=None,
+        mask_maskstart=None,
+        mask_maskend=None,
+        mask_softnessL=None,
+        mask_softnessR=None,
+        mask_marination=None,
+        batch_size=1    
+    ):
+
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+    gc.collect()
+
+    print(f"Prompt: {prompt}")
+
+    global preview_images
+    preview_images = []
+    if preview_every == 0:
+        preview_every = None
+
+    # Return fake stereo audio
+    conditioning = [{"prompt": prompt, "latitude": -latitude, "longitude": longitude, "temperature": temperature, "humidity": humidity, "wind_speed": wind_speed, "pressure": pressure, "minutes_of_day": minutes_of_day,"day_of_year": day_of_year, "seconds_start":seconds_start, "seconds_total": seconds_total }] * batch_size
+
+    if negative_prompt:
+        negative_conditioning = [{"prompt": negative_prompt, "latitude": -latitude, "longitude": longitude, "temperature": temperature, "humidity": humidity, "wind_speed": wind_speed, "pressure": pressure, "minutes_of_day": minutes_of_day,"day_of_year": day_of_year, "seconds_start":seconds_start, "seconds_total": seconds_total}] * batch_size
+    else:
+        negative_conditioning = None
+        
+    #Get the device from the model
+    device = next(model.parameters()).device
+
+    seed = int(seed)
+
+    if not use_init:
+        init_audio = None
+    
+    input_sample_size = sample_size
+
+    if init_audio is not None:
+        in_sr, init_audio = init_audio
+        # Turn into torch tensor, converting from int16 to float32
+        init_audio = torch.from_numpy(init_audio).float().div(32767)
+        
+        if init_audio.dim() == 1:
+            init_audio = init_audio.unsqueeze(0) # [1, n]
+        elif init_audio.dim() == 2:
+            init_audio = init_audio.transpose(0, 1) # [n, 2] -> [2, n]
+
+        if in_sr != sample_rate:
+            resample_tf = T.Resample(in_sr, sample_rate).to(init_audio.device)
+            init_audio = resample_tf(init_audio)
+
+        audio_length = init_audio.shape[-1]
+
+        if audio_length > sample_size:
+
+            input_sample_size = audio_length + (model.min_input_length - (audio_length % model.min_input_length)) % model.min_input_length
+
+        init_audio = (sample_rate, init_audio)
+
+    def progress_callback(callback_info):
+        global preview_images
+        denoised = callback_info["denoised"]
+        current_step = callback_info["i"]
+        sigma = callback_info["sigma"]
+
+        if (current_step - 1) % preview_every == 0:
+            if model.pretransform is not None:
+                denoised = model.pretransform.decode(denoised)
+            denoised = rearrange(denoised, "b d n -> d (b n)")
+            denoised = denoised.clamp(-1, 1).mul(32767).to(torch.int16).cpu()
+            audio_spectrogram = audio_spectrogram_image(denoised, sample_rate=sample_rate)
+            preview_images.append((audio_spectrogram, f"Step {current_step} sigma={sigma:.3f})"))
+
+    # If inpainting, send mask args
+    # This will definitely change in the future
+    if mask_cropfrom is not None: 
+        mask_args = {
+            "cropfrom": mask_cropfrom,
+            "pastefrom": mask_pastefrom,
+            "pasteto": mask_pasteto,
+            "maskstart": mask_maskstart,
+            "maskend": mask_maskend,
+            "softnessL": mask_softnessL,
+            "softnessR": mask_softnessR,
+            "marination": mask_marination,
+        }
+    else:
+        mask_args = None 
+
+    # Do the audio generation
+    audio = generate_diffusion_cond(
+        model, 
+        conditioning=conditioning,
+        negative_conditioning=negative_conditioning,
+        steps=steps,
+        cfg_scale=cfg_scale,
+        batch_size=batch_size,
+        sample_size=input_sample_size,
+        sample_rate=sample_rate,
+        seed=seed,
+        device=device,
+        sampler_type=sampler_type,
+        sigma_min=sigma_min,
+        sigma_max=sigma_max,
+        init_audio=init_audio,
+        init_noise_level=init_noise_level,
+        mask_args = mask_args,
+        callback = progress_callback if preview_every is not None else None,
+        scale_phi = cfg_rescale
+    )
+
+    # Convert to WAV file
+    audio = rearrange(audio, "b d n -> d (b n)")
+    audio = audio.to(torch.float32).div(torch.max(torch.abs(audio))).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
+    torchaudio.save("output.wav", audio, sample_rate)
+
+    # Let's look at a nice spectrogram too
+    audio_spectrogram = audio_spectrogram_image(audio, sample_rate=sample_rate)
+
+    return ("output.wav", [audio_spectrogram, *preview_images])
+
+def generate_uncond(
+        steps=250,
+        seed=-1,
+        sampler_type="dpmpp-2m-sde",
+        sigma_min=0.03,
+        sigma_max=50,
+        use_init=False,
+        init_audio=None,
+        init_noise_level=1.0,
+        batch_size=1,
+        preview_every=None
+        ):
+
+    global preview_images
+
+    preview_images = []
+
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+    gc.collect()
+
+    #Get the device from the model
+    device = next(model.parameters()).device
+
+    seed = int(seed)
+
+    if not use_init:
+        init_audio = None
+    
+    input_sample_size = sample_size
+
+    if init_audio is not None:
+        in_sr, init_audio = init_audio
+        # Turn into torch tensor, converting from int16 to float32
+        init_audio = torch.from_numpy(init_audio).float().div(32767)
+        
+        if init_audio.dim() == 1:
+            init_audio = init_audio.unsqueeze(0) # [1, n]
+        elif init_audio.dim() == 2:
+            init_audio = init_audio.transpose(0, 1) # [n, 2] -> [2, n]
+
+        if in_sr != sample_rate:
+            resample_tf = T.Resample(in_sr, sample_rate).to(init_audio.device)
+            init_audio = resample_tf(init_audio)
+
+        audio_length = init_audio.shape[-1]
+
+        if audio_length > sample_size:
+
+            input_sample_size = audio_length + (model.min_input_length - (audio_length % model.min_input_length)) % model.min_input_length
+
+        init_audio = (sample_rate, init_audio)
+
+    def progress_callback(callback_info):
+        global preview_images
+        denoised = callback_info["denoised"]
+        current_step = callback_info["i"]
+        sigma = callback_info["sigma"]
+
+        if (current_step - 1) % preview_every == 0:
+
+            if model.pretransform is not None:
+                denoised = model.pretransform.decode(denoised)
+
+            denoised = rearrange(denoised, "b d n -> d (b n)")
+
+            denoised = denoised.clamp(-1, 1).mul(32767).to(torch.int16).cpu()
+
+            audio_spectrogram = audio_spectrogram_image(denoised, sample_rate=sample_rate)
+
+            preview_images.append((audio_spectrogram, f"Step {current_step} sigma={sigma:.3f})"))
+
+    audio = generate_diffusion_uncond(
+        model, 
+        steps=steps,
+        batch_size=batch_size,
+        sample_size=input_sample_size,
+        seed=seed,
+        device=device,
+        sampler_type=sampler_type,
+        sigma_min=sigma_min,
+        sigma_max=sigma_max,
+        init_audio=init_audio,
+        init_noise_level=init_noise_level,
+        callback = progress_callback if preview_every is not None else None
+    )
+
+    audio = rearrange(audio, "b d n -> d (b n)")
+
+    audio = audio.to(torch.float32).div(torch.max(torch.abs(audio))).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
+
+    torchaudio.save("output.wav", audio, sample_rate)
+
+    audio_spectrogram = audio_spectrogram_image(audio, sample_rate=sample_rate)
+
+    return ("output.wav", [audio_spectrogram, *preview_images])
+
+def generate_lm(
+        temperature=1.0,
+        top_p=0.95,
+        top_k=0,    
+        batch_size=1,
+        ):
+    
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+    gc.collect()
+
+    #Get the device from the model
+    device = next(model.parameters()).device
+
+    audio = model.generate_audio(
+        batch_size=batch_size,
+        max_gen_len = sample_size//model.pretransform.downsampling_ratio,
+        conditioning=None,
+        temp=temperature,
+        top_p=top_p,
+        top_k=top_k,
+        use_cache=True
+    )
+
+    audio = rearrange(audio, "b d n -> d (b n)")
+
+    audio = audio.to(torch.float32).div(torch.max(torch.abs(audio))).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
+
+    torchaudio.save("output.wav", audio, sample_rate)
+
+    audio_spectrogram = audio_spectrogram_image(audio, sample_rate=sample_rate)
+
+    return ("output.wav", [audio_spectrogram])
+
+
+def create_uncond_sampling_ui(model_config):   
+    generate_button = gr.Button("Generate", variant='primary', scale=1)
+    
+    with gr.Row(equal_height=False):
+        with gr.Column():            
+            with gr.Row():
+                # Steps slider
+                steps_slider = gr.Slider(minimum=1, maximum=500, step=1, value=100, label="Steps")
+
+            with gr.Accordion("Sampler params", open=False):
+            
+                # Seed
+                seed_textbox = gr.Textbox(label="Seed (set to -1 for random seed)", value="-1")
+
+            # Sampler params
+                with gr.Row():
+                    sampler_type_dropdown = gr.Dropdown(["dpmpp-2m-sde", "dpmpp-3m-sde", "k-heun", "k-lms", "k-dpmpp-2s-ancestral", "k-dpm-2", "k-dpm-fast"], label="Sampler type", value="dpmpp-2m-sde")
+                    sigma_min_slider = gr.Slider(minimum=0.0, maximum=2.0, step=0.01, value=0.03, label="Sigma min")
+                    sigma_max_slider = gr.Slider(minimum=0.0, maximum=200.0, step=0.1, value=80, label="Sigma max")
+
+            with gr.Accordion("Init audio", open=False):
+                init_audio_checkbox = gr.Checkbox(label="Use init audio")
+                init_audio_input = gr.Audio(label="Init audio")
+                init_noise_level_slider = gr.Slider(minimum=0.0, maximum=100.0, step=0.01, value=0.1, label="Init noise level")
+
+        with gr.Column():
+            audio_output = gr.Audio(label="Output audio", interactive=False)
+            audio_spectrogram_output = gr.Gallery(label="Output spectrogram", show_label=False)
+            send_to_init_button = gr.Button("Send to init audio", scale=1)
+            send_to_init_button.click(fn=lambda audio: audio, inputs=[audio_output], outputs=[init_audio_input])
+    
+    generate_button.click(fn=generate_uncond, 
+        inputs=[
+            steps_slider, 
+            seed_textbox, 
+            sampler_type_dropdown, 
+            sigma_min_slider, 
+            sigma_max_slider,
+            init_audio_checkbox,
+            init_audio_input,
+            init_noise_level_slider,
+        ], 
+        outputs=[
+            audio_output, 
+            audio_spectrogram_output
+        ], 
+        api_name="generate")
+def create_conditioning_slider(min_val, max_val, label):
+    """
+    Create a Gradio slider for a given conditioning parameter.
+
+    Args:
+    - min_val: The minimum value for the slider.
+    - max_val: The maximum value for the slider.
+    - label: The label for the slider, which is displayed in the UI.
+
+    Returns:
+    - A gr.Slider object configured according to the provided parameters.
+    """
+    step = (max_val - min_val) / 1000
+    default_val = (max_val + min_val) / 2
+    print(f"Creating slider for {label} with min_val={min_val}, max_val={max_val}, step={step}, default_val={default_val}")
+    return gr.Slider(minimum=min_val, maximum=max_val, step=step, value=default_val, label=label)
+
+def create_sampling_ui(model_config, inpainting=False):
+    with gr.Row():
+        with gr.Column(scale=6):
+            prompt = gr.Textbox(show_label=False, placeholder="Prompt")
+            negative_prompt = gr.Textbox(show_label=False, placeholder="Negative prompt")
+        generate_button = gr.Button("Generate", variant='primary', scale=1)
+    
+    model_conditioning_config = model_config["model"].get("conditioning", None)
+
+    has_seconds_start = False
+    has_seconds_total = False
+
+    if model_conditioning_config is not None:
+        for conditioning_config in model_conditioning_config["configs"]:
+            if conditioning_config["id"] == "seconds_start":
+                has_seconds_start = True
+            if conditioning_config["id"] == "seconds_total":
+                has_seconds_total = True
+
+    with gr.Row(equal_height=False):
+        with gr.Column():
+            with gr.Row():
+                
+                seconds_start_slider = gr.Slider(minimum=0, maximum=512, step=1, value=0, label="Seconds start", visible=has_seconds_start)
+           
+                seconds_total_slider = gr.Slider(minimum=0, maximum=512, step=1, value=sample_size//sample_rate, label="Seconds total", visible=has_seconds_total)
+            
+            with gr.Row():
+                # Steps slider
+                steps_slider = gr.Slider(minimum=1, maximum=500, step=1, value=100, label="Steps")
+
+                # Preview Every slider
+                preview_every_slider = gr.Slider(minimum=0, maximum=100, step=1, value=0, label="Preview Every")
+
+                # CFG scale 
+                cfg_scale_slider = gr.Slider(minimum=0.0, maximum=25.0, step=0.1, value=0.2, label="CFG scale")
+                
+            with gr.Accordion("Climate and location", open=True):
+                latitude_config = next((item for item in model_conditioning_config["configs"] if item["id"] == "latitude"), None)
+                if latitude_config:
+                    latitude_slider = create_conditioning_slider(
+                        min_val=latitude_config["config"]["min_val"],
+                        max_val=latitude_config["config"]["max_val"],
+                        label="latitude")
+                    
+                longitude_config = next((item for item in model_conditioning_config["configs"] if item["id"] == "longitude"), None)
+                if longitude_config:
+                    longitude_slider = create_conditioning_slider(
+                        min_val=longitude_config["config"]["min_val"],
+                        max_val=longitude_config["config"]["max_val"],
+                        label="longitude")
+                    
+                temperature_config = next((item for item in model_conditioning_config["configs"] if item["id"] == "temperature"), None)
+                if temperature_config:
+                    temperature_slider = create_conditioning_slider(
+                        min_val=temperature_config["config"]["min_val"],
+                        max_val=temperature_config["config"]["max_val"],
+                        label="temperature")
+                    
+                humidity_config = next((item for item in model_conditioning_config["configs"] if item["id"] == "humidity"), None)
+                if humidity_config:
+                    humidity_slider = create_conditioning_slider(
+                        min_val=humidity_config["config"]["min_val"],
+                        max_val=humidity_config["config"]["max_val"],
+                        label="humidity")
+                    
+                wind_speed_config = next((item for item in model_conditioning_config["configs"] if item["id"] == "wind_speed"), None)
+                if wind_speed_config:
+                    wind_speed_slider = create_conditioning_slider(
+                        min_val=wind_speed_config["config"]["min_val"],
+                        max_val=wind_speed_config["config"]["max_val"],
+                        label="wind_speed")
+                    
+                pressure_config = next((item for item in model_conditioning_config["configs"] if item["id"] == "pressure"), None)
+                if pressure_config:
+                    pressure_slider = create_conditioning_slider(
+                        min_val=pressure_config["config"]["min_val"],
+                        max_val=pressure_config["config"]["max_val"],
+                        label="pressure")
+                    
+                minutes_of_day_config = next((item for item in model_conditioning_config["configs"] if item["id"] == "minutes_of_day"), None)
+                if minutes_of_day_config:
+                    minutes_of_day_slider = create_conditioning_slider(
+                        min_val=minutes_of_day_config["config"]["min_val"],
+                        max_val=minutes_of_day_config["config"]["max_val"],
+                        label="minutes_of_day")
+                    
+                day_of_year_config = next((item for item in model_conditioning_config["configs"] if item["id"] == "day_of_year"), None)
+                if day_of_year_config:
+                    day_of_year_slider = create_conditioning_slider(
+                        min_val=day_of_year_config["config"]["min_val"],
+                        max_val=day_of_year_config["config"]["max_val"],
+                        label="Day of year")
+
+            with gr.Accordion("Sampler params", open=False):
+            
+                # Seed
+                seed_textbox = gr.Textbox(label="Seed (set to -1 for random seed)", value="-1")
+
+                # Sampler params
+                with gr.Row():
+                    sampler_type_dropdown = gr.Dropdown(["dpmpp-2m-sde", "dpmpp-3m-sde", "k-heun", "k-lms", "k-dpmpp-2s-ancestral", "k-dpm-2", "k-dpm-fast"], label="Sampler type", value="dpmpp-2m-sde")
+                    sigma_min_slider = gr.Slider(minimum=0.0, maximum=2.0, step=0.01, value=0.03, label="Sigma min")
+                    sigma_max_slider = gr.Slider(minimum=0.0, maximum=200.0, step=0.1, value=80, label="Sigma max")
+                    cfg_rescale_slider = gr.Slider(minimum=0.0, maximum=1, step=0.01, value=0.2, label="CFG rescale amount")
+
+            if inpainting: 
+                # Inpainting Tab
+                with gr.Accordion("Inpainting", open=False):
+                    sigma_max_slider.maximum=1000
+                    
+                    init_audio_checkbox = gr.Checkbox(label="Do inpainting")
+                    init_audio_input = gr.Audio(label="Init audio")
+                    init_noise_level_slider = gr.Slider(minimum=0.1, maximum=100.0, step=0.1, value=80, label="Init audio noise level", visible=False) # hide this
+
+                    mask_cropfrom_slider = gr.Slider(minimum=0.0, maximum=100.0, step=0.1, value=0, label="Crop From %")
+                    mask_pastefrom_slider = gr.Slider(minimum=0.0, maximum=100.0, step=0.1, value=0, label="Paste From %")
+                    mask_pasteto_slider = gr.Slider(minimum=0.0, maximum=100.0, step=0.1, value=100, label="Paste To %")
+
+                    mask_maskstart_slider = gr.Slider(minimum=0.0, maximum=100.0, step=0.1, value=50, label="Mask Start %")
+                    mask_maskend_slider = gr.Slider(minimum=0.0, maximum=100.0, step=0.1, value=100, label="Mask End %")
+                    mask_softnessL_slider = gr.Slider(minimum=0.0, maximum=100.0, step=0.1, value=0, label="Softmask Left Crossfade Length %")
+                    mask_softnessR_slider = gr.Slider(minimum=0.0, maximum=100.0, step=0.1, value=0, label="Softmask Right Crossfade Length %")
+                    mask_marination_slider = gr.Slider(minimum=0.0, maximum=1, step=0.0001, value=0, label="Marination level", visible=False) # still working on the usefulness of this 
+
+                    inputs = [prompt, 
+                        negative_prompt,
+                        seconds_start_slider, 
+                        seconds_total_slider,
+                        latitude_slider,
+                        longitude_slider,
+                        temperature_slider,
+                        humidity_slider,
+                        wind_speed_slider,
+                        pressure_slider,
+                        minutes_of_day_slider,
+                        day_of_year_slider, 
+                        cfg_scale_slider, 
+                        steps_slider, 
+                        preview_every_slider, 
+                        seed_textbox, 
+                        sampler_type_dropdown, 
+                        sigma_min_slider, 
+                        sigma_max_slider,
+                        cfg_rescale_slider,
+                        init_audio_checkbox,
+                        init_audio_input,
+                        init_noise_level_slider,
+                        mask_cropfrom_slider,
+                        mask_pastefrom_slider,
+                        mask_pasteto_slider,
+                        mask_maskstart_slider,
+                        mask_maskend_slider,
+                        mask_softnessL_slider,
+                        mask_softnessR_slider,
+                        mask_marination_slider
+                    ]
+            else:
+                # Default generation tab
+                with gr.Accordion("Init audio", open=False):
+                    init_audio_checkbox = gr.Checkbox(label="Use init audio")
+                    init_audio_input = gr.Audio(label="Init audio")
+                    init_noise_level_slider = gr.Slider(minimum=0.1, maximum=100.0, step=0.01, value=0.1, label="Init noise level")
+
+                    inputs = [prompt, 
+                        negative_prompt,
+                        seconds_start_slider, 
+                        seconds_total_slider, 
+                        latitude_slider,
+                        longitude_slider,
+                        temperature_slider,
+                        humidity_slider,
+                        wind_speed_slider,
+                        pressure_slider,
+                        minutes_of_day_slider,
+                        day_of_year_slider,
+                        cfg_scale_slider, 
+                        steps_slider, 
+                        preview_every_slider, 
+                        seed_textbox, 
+                        sampler_type_dropdown, 
+                        sigma_min_slider, 
+                        sigma_max_slider,
+                        cfg_rescale_slider,
+                        init_audio_checkbox,
+                        init_audio_input,
+                        init_noise_level_slider
+                    ]
+
+        with gr.Column():
+            audio_output = gr.Audio(label="Output audio", interactive=False)
+            audio_spectrogram_output = gr.Gallery(label="Output spectrogram", show_label=False)
+            send_to_init_button = gr.Button("Send to init audio", scale=1)
+            send_to_init_button.click(fn=lambda audio: audio, inputs=[audio_output], outputs=[init_audio_input])
+    
+    generate_button.click(fn=generate_cond, 
+        inputs=inputs,
+        outputs=[
+            audio_output, 
+            audio_spectrogram_output
+        ], 
+        api_name="generate")
+
+
+def create_txt2audio_ui(model_config):
+    with gr.Blocks() as ui:
+        with gr.Tab("Generation"):
+            create_sampling_ui(model_config) 
+        with gr.Tab("Inpainting"):
+            create_sampling_ui(model_config, inpainting=True)    
+    return ui
+
+def create_diffusion_uncond_ui(model_config):
+    with gr.Blocks() as ui:
+        create_uncond_sampling_ui(model_config)
+    
+    return ui
+
+def autoencoder_process(audio, latent_noise, n_quantizers):
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+    gc.collect()
+
+    #Get the device from the model
+    device = next(model.parameters()).device
+
+    in_sr, audio = audio
+
+    audio = torch.from_numpy(audio).float().div(32767).to(device)
+
+    if audio.dim() == 1:
+        audio = audio.unsqueeze(0)
+    else:
+        audio = audio.transpose(0, 1)
+
+    audio = model.preprocess_audio_for_encoder(audio, in_sr)
+    # Note: If you need to do chunked encoding, to reduce VRAM, 
+    # then add these arguments to encode_audio and decode_audio: chunked=True, overlap=32, chunk_size=128
+    # To turn it off, do chunked=False
+    # Optimal overlap and chunk_size values will depend on the model. 
+    # See encode_audio & decode_audio in autoencoders.py for more info
+    # Get dtype of model
+    dtype = next(model.parameters()).dtype
+
+    audio = audio.to(dtype)
+
+    if n_quantizers > 0:
+        latents = model.encode_audio(audio, chunked=False, n_quantizers=n_quantizers)
+    else:
+        latents = model.encode_audio(audio, chunked=False)
+
+    if latent_noise > 0:
+        latents = latents + torch.randn_like(latents) * latent_noise
+
+    audio = model.decode_audio(latents, chunked=False)
+
+    audio = rearrange(audio, "b d n -> d (b n)")
+
+    audio = audio.to(torch.float32).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
+
+    torchaudio.save("output.wav", audio, sample_rate)
+
+    return "output.wav"
+
+def create_autoencoder_ui(model_config):
+
+    is_dac_rvq = "model" in model_config and "bottleneck" in model_config["model"] and model_config["model"]["bottleneck"]["type"] in ["dac_rvq","dac_rvq_vae"]
+
+    if is_dac_rvq:
+        n_quantizers = model_config["model"]["bottleneck"]["config"]["n_codebooks"]
+    else:
+        n_quantizers = 0
+
+    with gr.Blocks() as ui:
+        input_audio = gr.Audio(label="Input audio")
+        output_audio = gr.Audio(label="Output audio", interactive=False)
+        n_quantizers_slider = gr.Slider(minimum=1, maximum=n_quantizers, step=1, value=n_quantizers, label="# quantizers", visible=is_dac_rvq)
+        latent_noise_slider = gr.Slider(minimum=0.0, maximum=10.0, step=0.001, value=0.0, label="Add latent noise")
+        process_button = gr.Button("Process", variant='primary', scale=1)
+        process_button.click(fn=autoencoder_process, inputs=[input_audio, latent_noise_slider, n_quantizers_slider], outputs=output_audio, api_name="process")
+
+    return ui
+
+def diffusion_prior_process(audio, steps, sampler_type, sigma_min, sigma_max):
+
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+    gc.collect()
+
+    #Get the device from the model
+    device = next(model.parameters()).device
+
+    in_sr, audio = audio
+
+    audio = torch.from_numpy(audio).float().div(32767).to(device)
+    
+    if audio.dim() == 1:
+        audio = audio.unsqueeze(0) # [1, n]
+    elif audio.dim() == 2:
+        audio = audio.transpose(0, 1) # [n, 2] -> [2, n]
+
+    audio = audio.unsqueeze(0)
+
+    audio = model.stereoize(audio, in_sr, steps, sampler_kwargs={"sampler_type": sampler_type, "sigma_min": sigma_min, "sigma_max": sigma_max})
+
+    audio = rearrange(audio, "b d n -> d (b n)")
+
+    audio = audio.to(torch.float32).div(torch.max(torch.abs(audio))).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
+
+    torchaudio.save("output.wav", audio, sample_rate)
+
+    return "output.wav"
+
+def create_diffusion_prior_ui(model_config):
+    with gr.Blocks() as ui:
+        input_audio = gr.Audio(label="Input audio")
+        output_audio = gr.Audio(label="Output audio", interactive=False)
+        # Sampler params
+        with gr.Row():
+            steps_slider = gr.Slider(minimum=1, maximum=500, step=1, value=100, label="Steps")
+            sampler_type_dropdown = gr.Dropdown(["dpmpp-2m-sde", "dpmpp-3m-sde", "k-heun", "k-lms", "k-dpmpp-2s-ancestral", "k-dpm-2", "k-dpm-fast"], label="Sampler type", value="dpmpp-2m-sde")
+            sigma_min_slider = gr.Slider(minimum=0.0, maximum=2.0, step=0.01, value=0.03, label="Sigma min")
+            sigma_max_slider = gr.Slider(minimum=0.0, maximum=200.0, step=0.1, value=80, label="Sigma max")
+        process_button = gr.Button("Process", variant='primary', scale=1)
+        process_button.click(fn=diffusion_prior_process, inputs=[input_audio, steps_slider, sampler_type_dropdown, sigma_min_slider, sigma_max_slider], outputs=output_audio, api_name="process")    
+
+    return ui
+
+def create_lm_ui(model_config):
+    with gr.Blocks() as ui:
+        output_audio = gr.Audio(label="Output audio", interactive=False)
+        audio_spectrogram_output = gr.Gallery(label="Output spectrogram", show_label=False)
+
+        # Sampling params
+        with gr.Row():
+            temperature_slider = gr.Slider(minimum=0, maximum=5, step=0.01, value=1.0, label="Temperature")
+            top_p_slider = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.95, label="Top p")
+            top_k_slider = gr.Slider(minimum=0, maximum=100, step=1, value=0, label="Top k")
+
+        generate_button = gr.Button("Generate", variant='primary', scale=1)
+        generate_button.click(
+            fn=generate_lm, 
+            inputs=[
+                temperature_slider, 
+                top_p_slider, 
+                top_k_slider
+            ], 
+            outputs=[output_audio, audio_spectrogram_output],
+            api_name="generate"
+        )
+
+    return ui
+
+def create_ui(model_config_path=None, ckpt_path=None, pretrained_name=None, pretransform_ckpt_path=None, model_half=False):
+
+    assert (pretrained_name is not None) ^ (model_config_path is not None and ckpt_path is not None), "Must specify either pretrained name or provide a model config and checkpoint, but not both"
+
+    if model_config_path is not None:
+        # Load config from json file
+        with open(model_config_path) as f:
+            model_config = json.load(f)
+    else:
+        model_config = None
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    _, model_config = load_model(model_config, ckpt_path, pretrained_name=pretrained_name, pretransform_ckpt_path=pretransform_ckpt_path, model_half=model_half, device=device)
+    
+    model_type = model_config["model_type"]
+
+    if model_type == "diffusion_cond":
+        ui = create_txt2audio_ui(model_config)
+    elif model_type == "diffusion_uncond":
+        ui = create_diffusion_uncond_ui(model_config)
+    elif model_type == "autoencoder" or model_type == "diffusion_autoencoder":
+        ui = create_autoencoder_ui(model_config)
+    elif model_type == "diffusion_prior":
+        ui = create_diffusion_prior_ui(model_config)
+    elif model_type == "lm":
+        ui = create_lm_ui(model_config)
+        
+    return ui

stable_audio_tools/interface/testing.py

@@ -0,0 +1,409 @@
+import gc
+import numpy as np
+import json 
+import torch
+import torchaudio
+import os
+import re
+
+from aeiou.viz import audio_spectrogram_image
+from einops import rearrange
+from safetensors.torch import load_file
+from torch.nn import functional as F
+from torchaudio import transforms as T
+
+from ..inference.generation import generate_diffusion_cond, generate_diffusion_uncond
+from ..models.factory import create_model_from_config
+from ..models.pretrained import get_pretrained_model
+from ..models.utils import load_ckpt_state_dict
+from ..inference.utils import prepare_audio
+from ..training.utils import copy_state_dict
+
+
+model = None
+sample_rate = 44100
+sample_size = 524288
+
+def load_model(model_config=None, model_ckpt_path=None, pretrained_name=None, pretransform_ckpt_path=None, device="cuda", model_half=False):
+    global model, sample_rate, sample_size
+    
+    if pretrained_name is not None:
+        print(f"Loading pretrained model {pretrained_name}")
+        model, model_config = get_pretrained_model(pretrained_name)
+
+    elif model_config is not None and model_ckpt_path is not None:
+        print(f"Creating model from config")
+        model = create_model_from_config(model_config)
+
+        print(f"Loading model checkpoint from {model_ckpt_path}")
+        # Load checkpoint
+        copy_state_dict(model, load_ckpt_state_dict(model_ckpt_path))
+        #model.load_state_dict(load_ckpt_state_dict(model_ckpt_path))
+
+    sample_rate = model_config["sample_rate"]
+    sample_size = model_config["sample_size"]
+
+    if pretransform_ckpt_path is not None:
+        print(f"Loading pretransform checkpoint from {pretransform_ckpt_path}")
+        model.pretransform.load_state_dict(load_ckpt_state_dict(pretransform_ckpt_path), strict=False)
+        print(f"Done loading pretransform")
+
+    model.to(device).eval().requires_grad_(False)
+
+    if model_half:
+        model.to(torch.float16)
+        
+    print(f"Done loading model")
+
+    return model, model_config
+
+def generate_cond_with_path(
+        prompt,
+        negative_prompt=None,
+        seconds_start=0,
+        seconds_total=30,
+        latitude = 0.0,
+        longitude = 0.0,
+        temperature = 0.0,
+        humidity = 0.0,
+        wind_speed = 0.0,
+        pressure = 0.0,
+        minutes_of_day = 0.0,
+        day_of_year = 0.0,
+        cfg_scale=6.0,
+        steps=250,
+        preview_every=None,
+        seed=-1,
+        sampler_type="dpmpp-2m-sde",
+        sigma_min=0.03,
+        sigma_max=50,
+        cfg_rescale=0.4,
+        use_init=False,
+        init_audio=None,
+        init_noise_level=1.0,
+        mask_cropfrom=None,
+        mask_pastefrom=None,
+        mask_pasteto=None,
+        mask_maskstart=None,
+        mask_maskend=None,
+        mask_softnessL=None,
+        mask_softnessR=None,
+        mask_marination=None,
+        batch_size=1,
+        destination_folder=None,
+        file_name=None    
+    ):
+
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+    gc.collect()
+
+    print(f"Prompt: {prompt}")
+
+    global preview_images
+    preview_images = []
+    if preview_every == 0:
+        preview_every = None
+
+    # Return fake stereo audio
+    conditioning = [{"prompt": prompt, "latitude": latitude, "longitude": longitude, "temperature": temperature, "humidity": humidity, "wind_speed": wind_speed, "pressure": pressure, "minutes_of_day": minutes_of_day,"day_of_year": day_of_year, "seconds_start":seconds_start, "seconds_total": seconds_total }] * batch_size
+
+    if negative_prompt:
+        negative_conditioning = [{"prompt": negative_prompt, "latitude": latitude, "longitude": longitude, "temperature": temperature, "humidity": humidity, "wind_speed": wind_speed, "pressure": pressure, "minutes_of_day": minutes_of_day,"day_of_year": day_of_year, "seconds_start":seconds_start, "seconds_total": seconds_total}] * batch_size
+    else:
+        negative_conditioning = None
+        
+    #Get the device from the model
+    device = next(model.parameters()).device
+
+    seed = int(seed)
+
+    if not use_init:
+        init_audio = None
+    
+    input_sample_size = sample_size
+
+    if init_audio is not None:
+        in_sr, init_audio = init_audio
+        # Turn into torch tensor, converting from int16 to float32
+        init_audio = torch.from_numpy(init_audio).float().div(32767)
+        
+        if init_audio.dim() == 1:
+            init_audio = init_audio.unsqueeze(0) # [1, n]
+        elif init_audio.dim() == 2:
+            init_audio = init_audio.transpose(0, 1) # [n, 2] -> [2, n]
+
+        if in_sr != sample_rate:
+            resample_tf = T.Resample(in_sr, sample_rate).to(init_audio.device)
+            init_audio = resample_tf(init_audio)
+
+        audio_length = init_audio.shape[-1]
+
+        if audio_length > sample_size:
+
+            input_sample_size = audio_length + (model.min_input_length - (audio_length % model.min_input_length)) % model.min_input_length
+
+        init_audio = (sample_rate, init_audio)
+
+    def progress_callback(callback_info):
+        global preview_images
+        denoised = callback_info["denoised"]
+        current_step = callback_info["i"]
+        sigma = callback_info["sigma"]
+
+        if (current_step - 1) % preview_every == 0:
+            if model.pretransform is not None:
+                denoised = model.pretransform.decode(denoised)
+            denoised = rearrange(denoised, "b d n -> d (b n)")
+            denoised = denoised.clamp(-1, 1).mul(32767).to(torch.int16).cpu()
+            audio_spectrogram = audio_spectrogram_image(denoised, sample_rate=sample_rate)
+            preview_images.append((audio_spectrogram, f"Step {current_step} sigma={sigma:.3f})"))
+
+    # If inpainting, send mask args
+    # This will definitely change in the future
+    if mask_cropfrom is not None: 
+        mask_args = {
+            "cropfrom": mask_cropfrom,
+            "pastefrom": mask_pastefrom,
+            "pasteto": mask_pasteto,
+            "maskstart": mask_maskstart,
+            "maskend": mask_maskend,
+            "softnessL": mask_softnessL,
+            "softnessR": mask_softnessR,
+            "marination": mask_marination,
+        }
+    else:
+        mask_args = None 
+
+    # Do the audio generation
+    audio = generate_diffusion_cond(
+        model, 
+        conditioning=conditioning,
+        negative_conditioning=negative_conditioning,
+        steps=steps,
+        cfg_scale=cfg_scale,
+        batch_size=batch_size,
+        sample_size=input_sample_size,
+        sample_rate=sample_rate,
+        seed=seed,
+        device=device,
+        sampler_type=sampler_type,
+        sigma_min=sigma_min,
+        sigma_max=sigma_max,
+        init_audio=init_audio,
+        init_noise_level=init_noise_level,
+        mask_args = mask_args,
+        callback = progress_callback if preview_every is not None else None,
+        scale_phi = cfg_rescale
+    )
+
+    # Convert to WAV file
+    audio = rearrange(audio, "b d n -> d (b n)")
+    audio = audio.to(torch.float32).div(torch.max(torch.abs(audio))).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
+    #save to the desired folder with the required filename and add the .wav extension
+    
+    if destination_folder is not None and file_name is not None:
+        torchaudio.save(f"{destination_folder}/{file_name}.wav", audio, sample_rate)
+        
+        
+
+    # Let's look at a nice spectrogram too
+    # audio_spectrogram = audio_spectrogram_image(audio, sample_rate=sample_rate)
+
+    # return ("output.wav", [audio_spectrogram, *preview_images])
+
+
+
+def generate_lm(
+        temperature=1.0,
+        top_p=0.95,
+        top_k=0,    
+        batch_size=1,
+        ):
+    
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+    gc.collect()
+
+    #Get the device from the model
+    device = next(model.parameters()).device
+
+    audio = model.generate_audio(
+        batch_size=batch_size,
+        max_gen_len = sample_size//model.pretransform.downsampling_ratio,
+        conditioning=None,
+        temp=temperature,
+        top_p=top_p,
+        top_k=top_k,
+        use_cache=True
+    )
+
+    audio = rearrange(audio, "b d n -> d (b n)")
+
+    audio = audio.to(torch.float32).div(torch.max(torch.abs(audio))).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
+
+    torchaudio.save("output.wav", audio, sample_rate)
+
+    audio_spectrogram = audio_spectrogram_image(audio, sample_rate=sample_rate)
+
+    return ("output.wav", [audio_spectrogram])
+
+
+
+
+def autoencoder_process(audio, latent_noise, n_quantizers):
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+    gc.collect()
+
+    #Get the device from the model
+    device = next(model.parameters()).device
+
+    in_sr, audio = audio
+
+    audio = torch.from_numpy(audio).float().div(32767).to(device)
+
+    if audio.dim() == 1:
+        audio = audio.unsqueeze(0)
+    else:
+        audio = audio.transpose(0, 1)
+
+    audio = model.preprocess_audio_for_encoder(audio, in_sr)
+    # Note: If you need to do chunked encoding, to reduce VRAM, 
+    # then add these arguments to encode_audio and decode_audio: chunked=True, overlap=32, chunk_size=128
+    # To turn it off, do chunked=False
+    # Optimal overlap and chunk_size values will depend on the model. 
+    # See encode_audio & decode_audio in autoencoders.py for more info
+    # Get dtype of model
+    dtype = next(model.parameters()).dtype
+
+    audio = audio.to(dtype)
+
+    if n_quantizers > 0:
+        latents = model.encode_audio(audio, chunked=False, n_quantizers=n_quantizers)
+    else:
+        latents = model.encode_audio(audio, chunked=False)
+
+    if latent_noise > 0:
+        latents = latents + torch.randn_like(latents) * latent_noise
+
+    audio = model.decode_audio(latents, chunked=False)
+
+    audio = rearrange(audio, "b d n -> d (b n)")
+
+    audio = audio.to(torch.float32).clamp(-1, 1).mul(32767).to(torch.int16).cpu()
+
+    torchaudio.save("output.wav", audio, sample_rate)
+
+    return "output.wav"
+
+def load_and_generate(model_path, json_dir, output_dir):
+    """Load JSON files and generate audio for each set of conditions."""
+    # List all files in the json_dir
+    files = os.listdir(json_dir)
+    
+    # Filter for JSON files
+    json_files = [file for file in files if file.endswith('.json')]
+    
+    if not json_files:
+        print(f"No JSON files found in {json_dir}. Please check the directory path and file permissions.")
+        return
+
+    for json_filename in json_files:
+        json_file_path = os.path.join(json_dir, json_filename)
+        
+        try:
+            with open(json_file_path, 'r') as file:
+                data = json.load(file)
+        except Exception as e:
+            print(f"Failed to read or parse {json_file_path}: {e}")
+            continue
+        
+        # Print the JSON path
+        print(json_file_path)
+        
+        # Extract conditions from JSON
+        conditions = {
+            'birdSpecies': data['birdSpecies'],
+            'latitude': data['coord']['lat'],
+            'longitude': data['coord']['lon'],
+            'temperature': data['main']['temp'],
+            'humidity': data['main']['humidity'],
+            'pressure': data['main']['pressure'],
+            'wind_speed': data['wind']['speed'],
+            'day_of_year': data['dayOfYear'],
+            'minutes_of_day': data['minutesOfDay']
+        }
+        
+        # Extract base filename components
+        step_number = re.search(r'step=(\d+)', model_path).group(1)
+        bird_species = conditions['birdSpecies'].replace(' ', '_')
+        base_filename = f"{bird_species}_{os.path.splitext(json_filename)[0]}_{step_number}_cfg_scale_"
+        
+
+        
+        #An array of cfg scale values to test
+        cfg_scales = [1.8, 2.5, 4.0, 5.0, 12.0]
+        
+        # Generate audio we do this 4 times with a loop
+        for scale in cfg_scales:
+            generate_cond_with_path(prompt = "",
+            negative_prompt="",
+            seconds_start=0,
+            seconds_total=22,
+            latitude = conditions['latitude'],
+            longitude = conditions['longitude'],
+            temperature = conditions['temperature'],
+            humidity = conditions['humidity'],
+            wind_speed = conditions['wind_speed'],
+            pressure = conditions['pressure'],
+            minutes_of_day = conditions['minutes_of_day'],
+            day_of_year = conditions['day_of_year'],
+            cfg_scale=scale,
+            steps=250,
+            preview_every=None,
+            seed=-1,
+            sampler_type="dpmpp-2m-sde",
+            sigma_min=0.03,
+            sigma_max=50,
+            cfg_rescale=0.4,
+            use_init=False,
+            init_audio=None,
+            init_noise_level=1.0,
+            mask_cropfrom=None,
+            mask_pastefrom=None,
+            mask_pasteto=None,
+            mask_maskstart=None,
+            mask_maskend=None,
+            mask_softnessL=None,
+            mask_softnessR=None,
+            mask_marination=None,
+            batch_size=1,
+            destination_folder=output_dir,
+            file_name=base_filename + str(scale))
+            
+            
+def runTests(model_config_path=None, ckpt_path=None, pretrained_name=None, pretransform_ckpt_path=None, model_half=False, json_dir=None, output_dir=None):
+    assert (pretrained_name is not None) ^ (model_config_path is not None and ckpt_path is not None), "Must specify either pretrained name or provide a model config and checkpoint, but not both"
+
+    if model_config_path is not None:
+        # Load config from json file
+        with open(model_config_path) as f:
+            model_config = json.load(f)
+    else:
+        model_config = None
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    _, model_config = load_model(model_config, ckpt_path, pretrained_name=pretrained_name, pretransform_ckpt_path=pretransform_ckpt_path, model_half=model_half, device=device)
+
+        # Ensure output directory exists-    os.makedirs(args.output_dir, exist_ok=True)
+
+    # Process all JSON files and generate audio
+    load_and_generate(ckpt_path, json_dir, output_dir)  
+
+
+
+
+            
+
+ 

stable_audio_tools/models/__init__.py

@@ -0,0 +1 @@
+from .factory import create_model_from_config, create_model_from_config_path

stable_audio_tools/models/adp.py

@@ -0,0 +1,1588 @@
+# Copied and modified from https://github.com/archinetai/audio-diffusion-pytorch/blob/v0.0.94/audio_diffusion_pytorch/modules.py under MIT License
+# License can be found in LICENSES/LICENSE_ADP.txt
+
+import math
+from inspect import isfunction
+from math import ceil, floor, log, pi, log2
+from typing import Any, Callable, Dict, List, Optional, Sequence, Tuple, TypeVar, Union
+from packaging import version
+
+import torch
+import torch.nn as nn
+from einops import rearrange, reduce, repeat
+from einops.layers.torch import Rearrange
+from einops_exts import rearrange_many
+from torch import Tensor, einsum
+from torch.backends.cuda import sdp_kernel
+from torch.nn import functional as F
+from dac.nn.layers import Snake1d
+
+"""
+Utils
+"""
+
+
+class ConditionedSequential(nn.Module):
+    def __init__(self, *modules):
+        super().__init__()
+        self.module_list = nn.ModuleList(*modules)
+
+    def forward(self, x: Tensor, mapping: Optional[Tensor] = None):
+        for module in self.module_list:
+            x = module(x, mapping)
+        return x
+
+T = TypeVar("T")
+
+def default(val: Optional[T], d: Union[Callable[..., T], T]) -> T:
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+def exists(val: Optional[T]) -> T:
+    return val is not None
+
+def closest_power_2(x: float) -> int:
+    exponent = log2(x)
+    distance_fn = lambda z: abs(x - 2 ** z)  # noqa
+    exponent_closest = min((floor(exponent), ceil(exponent)), key=distance_fn)
+    return 2 ** int(exponent_closest)
+
+def group_dict_by_prefix(prefix: str, d: Dict) -> Tuple[Dict, Dict]:
+    return_dicts: Tuple[Dict, Dict] = ({}, {})
+    for key in d.keys():
+        no_prefix = int(not key.startswith(prefix))
+        return_dicts[no_prefix][key] = d[key]
+    return return_dicts
+
+def groupby(prefix: str, d: Dict, keep_prefix: bool = False) -> Tuple[Dict, Dict]:
+    kwargs_with_prefix, kwargs = group_dict_by_prefix(prefix, d)
+    if keep_prefix:
+        return kwargs_with_prefix, kwargs
+    kwargs_no_prefix = {k[len(prefix) :]: v for k, v in kwargs_with_prefix.items()}
+    return kwargs_no_prefix, kwargs
+
+"""
+Convolutional Blocks
+"""
+import typing as tp
+
+# Copied from https://github.com/facebookresearch/audiocraft/blob/main/audiocraft/modules/conv.py under MIT License
+# License available in LICENSES/LICENSE_META.txt
+
+def get_extra_padding_for_conv1d(x: torch.Tensor, kernel_size: int, stride: int,
+                                 padding_total: int = 0) -> int:
+    """See `pad_for_conv1d`."""
+    length = x.shape[-1]
+    n_frames = (length - kernel_size + padding_total) / stride + 1
+    ideal_length = (math.ceil(n_frames) - 1) * stride + (kernel_size - padding_total)
+    return ideal_length - length
+
+
+def pad_for_conv1d(x: torch.Tensor, kernel_size: int, stride: int, padding_total: int = 0):
+    """Pad for a convolution to make sure that the last window is full.
+    Extra padding is added at the end. This is required to ensure that we can rebuild
+    an output of the same length, as otherwise, even with padding, some time steps
+    might get removed.
+    For instance, with total padding = 4, kernel size = 4, stride = 2:
+        0 0 1 2 3 4 5 0 0   # (0s are padding)
+        1   2   3           # (output frames of a convolution, last 0 is never used)
+        0 0 1 2 3 4 5 0     # (output of tr. conv., but pos. 5 is going to get removed as padding)
+            1 2 3 4         # once you removed padding, we are missing one time step !
+    """
+    extra_padding = get_extra_padding_for_conv1d(x, kernel_size, stride, padding_total)
+    return F.pad(x, (0, extra_padding))
+
+
+def pad1d(x: torch.Tensor, paddings: tp.Tuple[int, int], mode: str = 'constant', value: float = 0.):
+    """Tiny wrapper around F.pad, just to allow for reflect padding on small input.
+    If this is the case, we insert extra 0 padding to the right before the reflection happen.
+    """
+    length = x.shape[-1]
+    padding_left, padding_right = paddings
+    assert padding_left >= 0 and padding_right >= 0, (padding_left, padding_right)
+    if mode == 'reflect':
+        max_pad = max(padding_left, padding_right)
+        extra_pad = 0
+        if length <= max_pad:
+            extra_pad = max_pad - length + 1
+            x = F.pad(x, (0, extra_pad))
+        padded = F.pad(x, paddings, mode, value)
+        end = padded.shape[-1] - extra_pad
+        return padded[..., :end]
+    else:
+        return F.pad(x, paddings, mode, value)
+
+
+def unpad1d(x: torch.Tensor, paddings: tp.Tuple[int, int]):
+    """Remove padding from x, handling properly zero padding. Only for 1d!"""
+    padding_left, padding_right = paddings
+    assert padding_left >= 0 and padding_right >= 0, (padding_left, padding_right)
+    assert (padding_left + padding_right) <= x.shape[-1]
+    end = x.shape[-1] - padding_right
+    return x[..., padding_left: end]
+
+
+class Conv1d(nn.Conv1d):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+    
+    def forward(self, x: Tensor, causal=False) -> Tensor:
+        kernel_size = self.kernel_size[0]
+        stride = self.stride[0]
+        dilation = self.dilation[0]
+        kernel_size = (kernel_size - 1) * dilation + 1  # effective kernel size with dilations
+        padding_total = kernel_size - stride
+        extra_padding = get_extra_padding_for_conv1d(x, kernel_size, stride, padding_total)
+        if causal:
+            # Left padding for causal
+            x = pad1d(x, (padding_total, extra_padding))
+        else:
+            # Asymmetric padding required for odd strides
+            padding_right = padding_total // 2
+            padding_left = padding_total - padding_right
+            x = pad1d(x, (padding_left, padding_right + extra_padding))
+        return super().forward(x)
+        
+class ConvTranspose1d(nn.ConvTranspose1d):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def forward(self, x: Tensor, causal=False) -> Tensor:
+        kernel_size = self.kernel_size[0]
+        stride = self.stride[0]
+        padding_total = kernel_size - stride
+
+        y = super().forward(x)
+
+        # We will only trim fixed padding. Extra padding from `pad_for_conv1d` would be
+        # removed at the very end, when keeping only the right length for the output,
+        # as removing it here would require also passing the length at the matching layer
+        # in the encoder.
+        if causal:
+            padding_right = ceil(padding_total)
+            padding_left = padding_total - padding_right
+            y = unpad1d(y, (padding_left, padding_right))
+        else:
+            # Asymmetric padding required for odd strides
+            padding_right = padding_total // 2
+            padding_left = padding_total - padding_right
+            y = unpad1d(y, (padding_left, padding_right))
+        return y
+    
+
+def Downsample1d(
+    in_channels: int, out_channels: int, factor: int, kernel_multiplier: int = 2
+) -> nn.Module:
+    assert kernel_multiplier % 2 == 0, "Kernel multiplier must be even"
+
+    return Conv1d(
+        in_channels=in_channels,
+        out_channels=out_channels,
+        kernel_size=factor * kernel_multiplier + 1,
+        stride=factor
+    )
+
+
+def Upsample1d(
+    in_channels: int, out_channels: int, factor: int, use_nearest: bool = False
+) -> nn.Module:
+
+    if factor == 1:
+        return Conv1d(
+            in_channels=in_channels, out_channels=out_channels, kernel_size=3
+        )
+
+    if use_nearest:
+        return nn.Sequential(
+            nn.Upsample(scale_factor=factor, mode="nearest"),
+            Conv1d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=3
+            ),
+        )
+    else:
+        return ConvTranspose1d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=factor * 2,
+            stride=factor
+        )
+
+
+class ConvBlock1d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        *,
+        kernel_size: int = 3,
+        stride: int = 1,
+        dilation: int = 1,
+        num_groups: int = 8,
+        use_norm: bool = True,
+        use_snake: bool = False
+    ) -> None:
+        super().__init__()
+
+        self.groupnorm = (
+            nn.GroupNorm(num_groups=num_groups, num_channels=in_channels)
+            if use_norm
+            else nn.Identity()
+        )
+
+        if use_snake:
+            self.activation = Snake1d(in_channels)
+        else:
+            self.activation = nn.SiLU() 
+
+        self.project = Conv1d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            dilation=dilation,
+        )
+
+    def forward(
+        self, x: Tensor, scale_shift: Optional[Tuple[Tensor, Tensor]] = None, causal=False
+    ) -> Tensor:
+        x = self.groupnorm(x)
+        if exists(scale_shift):
+            scale, shift = scale_shift
+            x = x * (scale + 1) + shift
+        x = self.activation(x)
+        return self.project(x, causal=causal)
+
+
+class MappingToScaleShift(nn.Module):
+    def __init__(
+        self,
+        features: int,
+        channels: int,
+    ):
+        super().__init__()
+
+        self.to_scale_shift = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(in_features=features, out_features=channels * 2),
+        )
+
+    def forward(self, mapping: Tensor) -> Tuple[Tensor, Tensor]:
+        scale_shift = self.to_scale_shift(mapping)
+        scale_shift = rearrange(scale_shift, "b c -> b c 1")
+        scale, shift = scale_shift.chunk(2, dim=1)
+        return scale, shift
+
+
+class ResnetBlock1d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        *,
+        kernel_size: int = 3,
+        stride: int = 1,
+        dilation: int = 1,
+        use_norm: bool = True,
+        use_snake: bool = False,
+        num_groups: int = 8,
+        context_mapping_features: Optional[int] = None,
+    ) -> None:
+        super().__init__()
+
+        self.use_mapping = exists(context_mapping_features)
+
+        self.block1 = ConvBlock1d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            dilation=dilation,
+            use_norm=use_norm,
+            num_groups=num_groups,
+            use_snake=use_snake
+        )
+
+        if self.use_mapping:
+            assert exists(context_mapping_features)
+            self.to_scale_shift = MappingToScaleShift(
+                features=context_mapping_features, channels=out_channels
+            )
+
+        self.block2 = ConvBlock1d(
+            in_channels=out_channels,
+            out_channels=out_channels,
+            use_norm=use_norm,
+            num_groups=num_groups,
+            use_snake=use_snake
+        )
+
+        self.to_out = (
+            Conv1d(in_channels=in_channels, out_channels=out_channels, kernel_size=1)
+            if in_channels != out_channels
+            else nn.Identity()
+        )
+
+    def forward(self, x: Tensor, mapping: Optional[Tensor] = None, causal=False) -> Tensor:
+        assert_message = "context mapping required if context_mapping_features > 0"
+        assert not (self.use_mapping ^ exists(mapping)), assert_message
+
+        h = self.block1(x, causal=causal)
+
+        scale_shift = None
+        if self.use_mapping:
+            scale_shift = self.to_scale_shift(mapping)
+
+        h = self.block2(h, scale_shift=scale_shift, causal=causal)
+
+        return h + self.to_out(x)
+
+
+class Patcher(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        patch_size: int,
+        context_mapping_features: Optional[int] = None,
+        use_snake: bool = False,
+    ):
+        super().__init__()
+        assert_message = f"out_channels must be divisible by patch_size ({patch_size})"
+        assert out_channels % patch_size == 0, assert_message
+        self.patch_size = patch_size
+
+        self.block = ResnetBlock1d(
+            in_channels=in_channels,
+            out_channels=out_channels // patch_size,
+            num_groups=1,
+            context_mapping_features=context_mapping_features,
+            use_snake=use_snake
+        )
+
+    def forward(self, x: Tensor, mapping: Optional[Tensor] = None, causal=False) -> Tensor:
+        x = self.block(x, mapping, causal=causal)
+        x = rearrange(x, "b c (l p) -> b (c p) l", p=self.patch_size)
+        return x
+
+
+class Unpatcher(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        patch_size: int,
+        context_mapping_features: Optional[int] = None,
+        use_snake: bool = False
+    ):
+        super().__init__()
+        assert_message = f"in_channels must be divisible by patch_size ({patch_size})"
+        assert in_channels % patch_size == 0, assert_message
+        self.patch_size = patch_size
+
+        self.block = ResnetBlock1d(
+            in_channels=in_channels // patch_size,
+            out_channels=out_channels,
+            num_groups=1,
+            context_mapping_features=context_mapping_features,
+            use_snake=use_snake
+        )
+
+    def forward(self, x: Tensor, mapping: Optional[Tensor] = None, causal=False) -> Tensor:
+        x = rearrange(x, " b (c p) l -> b c (l p) ", p=self.patch_size)
+        x = self.block(x, mapping, causal=causal)
+        return x
+
+
+"""
+Attention Components
+"""
+def FeedForward(features: int, multiplier: int) -> nn.Module:
+    mid_features = features * multiplier
+    return nn.Sequential(
+        nn.Linear(in_features=features, out_features=mid_features),
+        nn.GELU(),
+        nn.Linear(in_features=mid_features, out_features=features),
+    )
+
+def add_mask(sim: Tensor, mask: Tensor) -> Tensor:
+    b, ndim = sim.shape[0], mask.ndim
+    if ndim == 3:
+        mask = rearrange(mask, "b n m -> b 1 n m")
+    if ndim == 2:
+        mask = repeat(mask, "n m -> b 1 n m", b=b)
+    max_neg_value = -torch.finfo(sim.dtype).max
+    sim = sim.masked_fill(~mask, max_neg_value)
+    return sim
+
+def causal_mask(q: Tensor, k: Tensor) -> Tensor:
+    b, i, j, device = q.shape[0], q.shape[-2], k.shape[-2], q.device
+    mask = ~torch.ones((i, j), dtype=torch.bool, device=device).triu(j - i + 1)
+    mask = repeat(mask, "n m -> b n m", b=b)
+    return mask
+
+class AttentionBase(nn.Module):
+    def __init__(
+        self,
+        features: int,
+        *,
+        head_features: int,
+        num_heads: int,
+        out_features: Optional[int] = None,
+    ):
+        super().__init__()
+        self.scale = head_features**-0.5
+        self.num_heads = num_heads
+        mid_features = head_features * num_heads
+        out_features = default(out_features, features)
+
+        self.to_out = nn.Linear(
+            in_features=mid_features, out_features=out_features
+        )
+
+        self.use_flash = False
+
+        if not self.use_flash:
+            return
+
+        device_properties = torch.cuda.get_device_properties(torch.device('cuda'))
+
+        if device_properties.major == 8 and device_properties.minor == 0:
+            # Use flash attention for A100 GPUs
+            self.sdp_kernel_config = (False, True, True)
+        else:
+            # Don't use flash attention for other GPUs
+            self.sdp_kernel_config = (False, True, True)
+
+    def forward(
+        self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor] = None, is_causal: bool = False
+    ) -> Tensor:
+        # Split heads
+        q, k, v = rearrange_many((q, k, v), "b n (h d) -> b h n d", h=self.num_heads)
+
+        if not self.use_flash:
+            if is_causal and not mask:
+                # Mask out future tokens for causal attention
+                mask = causal_mask(q, k)
+
+            # Compute similarity matrix and add eventual mask
+            sim = einsum("... n d, ... m d -> ... n m", q, k) * self.scale
+            sim = add_mask(sim, mask) if exists(mask) else sim
+
+            # Get attention matrix with softmax
+            attn = sim.softmax(dim=-1, dtype=torch.float32)
+
+            # Compute values
+            out = einsum("... n m, ... m d -> ... n d", attn, v)
+        else:
+            with sdp_kernel(*self.sdp_kernel_config):
+                out = F.scaled_dot_product_attention(q, k, v, attn_mask=mask, is_causal=is_causal)
+
+        out = rearrange(out, "b h n d -> b n (h d)")
+        return self.to_out(out)
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        features: int,
+        *,
+        head_features: int,
+        num_heads: int,
+        out_features: Optional[int] = None,
+        context_features: Optional[int] = None,
+        causal: bool = False,
+    ):
+        super().__init__()
+        self.context_features = context_features
+        self.causal = causal
+        mid_features = head_features * num_heads
+        context_features = default(context_features, features)
+
+        self.norm = nn.LayerNorm(features)
+        self.norm_context = nn.LayerNorm(context_features)
+        self.to_q = nn.Linear(
+            in_features=features, out_features=mid_features, bias=False
+        )
+        self.to_kv = nn.Linear(
+            in_features=context_features, out_features=mid_features * 2, bias=False
+        )
+        self.attention = AttentionBase(
+            features,
+            num_heads=num_heads,
+            head_features=head_features,
+            out_features=out_features,
+        )
+
+    def forward(
+        self,
+        x: Tensor, # [b, n, c]
+        context: Optional[Tensor] = None, # [b, m, d]
+        context_mask: Optional[Tensor] = None,  # [b, m], false is masked,
+        causal: Optional[bool] = False,
+    ) -> Tensor:
+        assert_message = "You must provide a context when using context_features"
+        assert not self.context_features or exists(context), assert_message
+        # Use context if provided
+        context = default(context, x)
+        # Normalize then compute q from input and k,v from context
+        x, context = self.norm(x), self.norm_context(context)
+
+        q, k, v = (self.to_q(x), *torch.chunk(self.to_kv(context), chunks=2, dim=-1))
+
+        if exists(context_mask):
+            # Mask out cross-attention for padding tokens
+            mask = repeat(context_mask, "b m -> b m d", d=v.shape[-1])
+            k, v = k * mask, v * mask
+
+        # Compute and return attention
+        return self.attention(q, k, v, is_causal=self.causal or causal)
+
+
+def FeedForward(features: int, multiplier: int) -> nn.Module:
+    mid_features = features * multiplier
+    return nn.Sequential(
+        nn.Linear(in_features=features, out_features=mid_features),
+        nn.GELU(),
+        nn.Linear(in_features=mid_features, out_features=features),
+    )
+
+"""
+Transformer Blocks
+"""
+
+
+class TransformerBlock(nn.Module):
+    def __init__(
+        self,
+        features: int,
+        num_heads: int,
+        head_features: int,
+        multiplier: int,
+        context_features: Optional[int] = None,
+    ):
+        super().__init__()
+
+        self.use_cross_attention = exists(context_features) and context_features > 0
+
+        self.attention = Attention(
+            features=features,
+            num_heads=num_heads,
+            head_features=head_features
+        )
+
+        if self.use_cross_attention:
+            self.cross_attention = Attention(
+                features=features,
+                num_heads=num_heads,
+                head_features=head_features,
+                context_features=context_features
+            )
+
+        self.feed_forward = FeedForward(features=features, multiplier=multiplier)
+
+    def forward(self, x: Tensor, *, context: Optional[Tensor] = None, context_mask: Optional[Tensor] = None, causal: Optional[bool] = False) -> Tensor:
+        x = self.attention(x, causal=causal) + x
+        if self.use_cross_attention:
+            x = self.cross_attention(x, context=context, context_mask=context_mask) + x
+        x = self.feed_forward(x) + x
+        return x
+
+
+"""
+Transformers
+"""
+
+
+class Transformer1d(nn.Module):
+    def __init__(
+        self,
+        num_layers: int,
+        channels: int,
+        num_heads: int,
+        head_features: int,
+        multiplier: int,
+        context_features: Optional[int] = None,
+    ):
+        super().__init__()
+
+        self.to_in = nn.Sequential(
+            nn.GroupNorm(num_groups=32, num_channels=channels, eps=1e-6, affine=True),
+            Conv1d(
+                in_channels=channels,
+                out_channels=channels,
+                kernel_size=1,
+            ),
+            Rearrange("b c t -> b t c"),
+        )
+
+        self.blocks = nn.ModuleList(
+            [
+                TransformerBlock(
+                    features=channels,
+                    head_features=head_features,
+                    num_heads=num_heads,
+                    multiplier=multiplier,
+                    context_features=context_features,
+                )
+                for i in range(num_layers)
+            ]
+        )
+
+        self.to_out = nn.Sequential(
+            Rearrange("b t c -> b c t"),
+            Conv1d(
+                in_channels=channels,
+                out_channels=channels,
+                kernel_size=1,
+            ),
+        )
+
+    def forward(self, x: Tensor, *, context: Optional[Tensor] = None, context_mask: Optional[Tensor] = None, causal=False) -> Tensor:
+        x = self.to_in(x)
+        for block in self.blocks:
+            x = block(x, context=context, context_mask=context_mask, causal=causal)
+        x = self.to_out(x)
+        return x
+
+
+"""
+Time Embeddings
+"""
+
+
+class SinusoidalEmbedding(nn.Module):
+    def __init__(self, dim: int):
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, x: Tensor) -> Tensor:
+        device, half_dim = x.device, self.dim // 2
+        emb = torch.tensor(log(10000) / (half_dim - 1), device=device)
+        emb = torch.exp(torch.arange(half_dim, device=device) * -emb)
+        emb = rearrange(x, "i -> i 1") * rearrange(emb, "j -> 1 j")
+        return torch.cat((emb.sin(), emb.cos()), dim=-1)
+
+
+class LearnedPositionalEmbedding(nn.Module):
+    """Used for continuous time"""
+
+    def __init__(self, dim: int):
+        super().__init__()
+        assert (dim % 2) == 0
+        half_dim = dim // 2
+        self.weights = nn.Parameter(torch.randn(half_dim))
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = rearrange(x, "b -> b 1")
+        freqs = x * rearrange(self.weights, "d -> 1 d") * 2 * pi
+        fouriered = torch.cat((freqs.sin(), freqs.cos()), dim=-1)
+        fouriered = torch.cat((x, fouriered), dim=-1)
+        return fouriered
+
+
+def TimePositionalEmbedding(dim: int, out_features: int) -> nn.Module:
+    return nn.Sequential(
+        LearnedPositionalEmbedding(dim),
+        nn.Linear(in_features=dim + 1, out_features=out_features),
+    )
+
+
+"""
+Encoder/Decoder Components
+"""
+
+
+class DownsampleBlock1d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        *,
+        factor: int,
+        num_groups: int,
+        num_layers: int,
+        kernel_multiplier: int = 2,
+        use_pre_downsample: bool = True,
+        use_skip: bool = False,
+        use_snake: bool = False,
+        extract_channels: int = 0,
+        context_channels: int = 0,
+        num_transformer_blocks: int = 0,
+        attention_heads: Optional[int] = None,
+        attention_features: Optional[int] = None,
+        attention_multiplier: Optional[int] = None,
+        context_mapping_features: Optional[int] = None,
+        context_embedding_features: Optional[int] = None,
+    ):
+        super().__init__()
+        self.use_pre_downsample = use_pre_downsample
+        self.use_skip = use_skip
+        self.use_transformer = num_transformer_blocks > 0
+        self.use_extract = extract_channels > 0
+        self.use_context = context_channels > 0
+
+        channels = out_channels if use_pre_downsample else in_channels
+
+        self.downsample = Downsample1d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            factor=factor,
+            kernel_multiplier=kernel_multiplier,
+        )
+
+        self.blocks = nn.ModuleList(
+            [
+                ResnetBlock1d(
+                    in_channels=channels + context_channels if i == 0 else channels,
+                    out_channels=channels,
+                    num_groups=num_groups,
+                    context_mapping_features=context_mapping_features,
+                    use_snake=use_snake
+                )
+                for i in range(num_layers)
+            ]
+        )
+
+        if self.use_transformer:
+            assert (
+                (exists(attention_heads) or exists(attention_features))
+                and exists(attention_multiplier)
+            )
+
+            if attention_features is None and attention_heads is not None:
+                attention_features = channels // attention_heads
+
+            if attention_heads is None and attention_features is not None:
+                attention_heads = channels // attention_features
+
+            self.transformer = Transformer1d(
+                num_layers=num_transformer_blocks,
+                channels=channels,
+                num_heads=attention_heads,
+                head_features=attention_features,
+                multiplier=attention_multiplier,
+                context_features=context_embedding_features
+            )
+
+        if self.use_extract:
+            num_extract_groups = min(num_groups, extract_channels)
+            self.to_extracted = ResnetBlock1d(
+                in_channels=out_channels,
+                out_channels=extract_channels,
+                num_groups=num_extract_groups,
+                use_snake=use_snake
+            )
+
+    def forward(
+        self,
+        x: Tensor,
+        *,
+        mapping: Optional[Tensor] = None,
+        channels: Optional[Tensor] = None,
+        embedding: Optional[Tensor] = None,
+        embedding_mask: Optional[Tensor] = None,
+        causal: Optional[bool] = False
+    ) -> Union[Tuple[Tensor, List[Tensor]], Tensor]:
+
+        if self.use_pre_downsample:
+            x = self.downsample(x)
+
+        if self.use_context and exists(channels):
+            x = torch.cat([x, channels], dim=1)
+
+        skips = []
+        for block in self.blocks:
+            x = block(x, mapping=mapping, causal=causal)
+            skips += [x] if self.use_skip else []
+
+        if self.use_transformer:
+            x = self.transformer(x, context=embedding, context_mask=embedding_mask, causal=causal)
+            skips += [x] if self.use_skip else []
+
+        if not self.use_pre_downsample:
+            x = self.downsample(x)
+
+        if self.use_extract:
+            extracted = self.to_extracted(x)
+            return x, extracted
+
+        return (x, skips) if self.use_skip else x
+
+
+class UpsampleBlock1d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        *,
+        factor: int,
+        num_layers: int,
+        num_groups: int,
+        use_nearest: bool = False,
+        use_pre_upsample: bool = False,
+        use_skip: bool = False,
+        use_snake: bool = False,
+        skip_channels: int = 0,
+        use_skip_scale: bool = False,
+        extract_channels: int = 0,
+        num_transformer_blocks: int = 0,
+        attention_heads: Optional[int] = None,
+        attention_features: Optional[int] = None,
+        attention_multiplier: Optional[int] = None,
+        context_mapping_features: Optional[int] = None,
+        context_embedding_features: Optional[int] = None,
+    ):
+        super().__init__()
+
+        self.use_extract = extract_channels > 0
+        self.use_pre_upsample = use_pre_upsample
+        self.use_transformer = num_transformer_blocks > 0
+        self.use_skip = use_skip
+        self.skip_scale = 2 ** -0.5 if use_skip_scale else 1.0
+
+        channels = out_channels if use_pre_upsample else in_channels
+
+        self.blocks = nn.ModuleList(
+            [
+                ResnetBlock1d(
+                    in_channels=channels + skip_channels,
+                    out_channels=channels,
+                    num_groups=num_groups,
+                    context_mapping_features=context_mapping_features,
+                    use_snake=use_snake
+                )
+                for _ in range(num_layers)
+            ]
+        )
+
+        if self.use_transformer:
+            assert (
+                (exists(attention_heads) or exists(attention_features))
+                and exists(attention_multiplier)
+            )
+
+            if attention_features is None and attention_heads is not None:
+                attention_features = channels // attention_heads
+
+            if attention_heads is None and attention_features is not None:
+                attention_heads = channels // attention_features
+
+            self.transformer = Transformer1d(
+                num_layers=num_transformer_blocks,
+                channels=channels,
+                num_heads=attention_heads,
+                head_features=attention_features,
+                multiplier=attention_multiplier,
+                context_features=context_embedding_features,
+            )
+
+        self.upsample = Upsample1d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            factor=factor,
+            use_nearest=use_nearest,
+        )
+
+        if self.use_extract:
+            num_extract_groups = min(num_groups, extract_channels)
+            self.to_extracted = ResnetBlock1d(
+                in_channels=out_channels,
+                out_channels=extract_channels,
+                num_groups=num_extract_groups,
+                use_snake=use_snake
+            )
+
+    def add_skip(self, x: Tensor, skip: Tensor) -> Tensor:
+        return torch.cat([x, skip * self.skip_scale], dim=1)
+
+    def forward(
+        self,
+        x: Tensor,
+        *,
+        skips: Optional[List[Tensor]] = None,
+        mapping: Optional[Tensor] = None,
+        embedding: Optional[Tensor] = None,
+        embedding_mask: Optional[Tensor] = None,
+        causal: Optional[bool] = False
+    ) -> Union[Tuple[Tensor, Tensor], Tensor]:
+
+        if self.use_pre_upsample:
+            x = self.upsample(x)
+
+        for block in self.blocks:
+            x = self.add_skip(x, skip=skips.pop()) if exists(skips) else x
+            x = block(x, mapping=mapping, causal=causal)
+
+        if self.use_transformer:
+            x = self.transformer(x, context=embedding, context_mask=embedding_mask, causal=causal)
+
+        if not self.use_pre_upsample:
+            x = self.upsample(x)
+
+        if self.use_extract:
+            extracted = self.to_extracted(x)
+            return x, extracted
+
+        return x
+
+
+class BottleneckBlock1d(nn.Module):
+    def __init__(
+        self,
+        channels: int,
+        *,
+        num_groups: int,
+        num_transformer_blocks: int = 0,
+        attention_heads: Optional[int] = None,
+        attention_features: Optional[int] = None,
+        attention_multiplier: Optional[int] = None,
+        context_mapping_features: Optional[int] = None,
+        context_embedding_features: Optional[int] = None,
+        use_snake: bool = False,
+    ):
+        super().__init__()
+        self.use_transformer = num_transformer_blocks > 0
+
+        self.pre_block = ResnetBlock1d(
+            in_channels=channels,
+            out_channels=channels,
+            num_groups=num_groups,
+            context_mapping_features=context_mapping_features,
+            use_snake=use_snake
+        )
+
+        if self.use_transformer:
+            assert (
+                (exists(attention_heads) or exists(attention_features))
+                and exists(attention_multiplier)
+            )
+
+            if attention_features is None and attention_heads is not None:
+                attention_features = channels // attention_heads
+
+            if attention_heads is None and attention_features is not None:
+                attention_heads = channels // attention_features
+
+            self.transformer = Transformer1d(
+                num_layers=num_transformer_blocks,
+                channels=channels,
+                num_heads=attention_heads,
+                head_features=attention_features,
+                multiplier=attention_multiplier,
+                context_features=context_embedding_features,
+            )
+
+        self.post_block = ResnetBlock1d(
+            in_channels=channels,
+            out_channels=channels,
+            num_groups=num_groups,
+            context_mapping_features=context_mapping_features,
+            use_snake=use_snake
+        )
+
+    def forward(
+        self,
+        x: Tensor,
+        *,
+        mapping: Optional[Tensor] = None,
+        embedding: Optional[Tensor] = None,
+        embedding_mask: Optional[Tensor] = None,
+        causal: Optional[bool] = False
+    ) -> Tensor:
+        x = self.pre_block(x, mapping=mapping, causal=causal)
+        if self.use_transformer:
+            x = self.transformer(x, context=embedding, context_mask=embedding_mask, causal=causal)
+        x = self.post_block(x, mapping=mapping, causal=causal)
+        return x
+
+
+"""
+UNet
+"""
+
+
+class UNet1d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        channels: int,
+        multipliers: Sequence[int],
+        factors: Sequence[int],
+        num_blocks: Sequence[int],
+        attentions: Sequence[int],
+        patch_size: int = 1,
+        resnet_groups: int = 8,
+        use_context_time: bool = True,
+        kernel_multiplier_downsample: int = 2,
+        use_nearest_upsample: bool = False,
+        use_skip_scale: bool = True,
+        use_snake: bool = False,
+        use_stft: bool = False,
+        use_stft_context: bool = False,
+        out_channels: Optional[int] = None,
+        context_features: Optional[int] = None,
+        context_features_multiplier: int = 4,
+        context_channels: Optional[Sequence[int]] = None,
+        context_embedding_features: Optional[int] = None,
+        **kwargs,
+    ):
+        super().__init__()
+        out_channels = default(out_channels, in_channels)
+        context_channels = list(default(context_channels, []))
+        num_layers = len(multipliers) - 1
+        use_context_features = exists(context_features)
+        use_context_channels = len(context_channels) > 0
+        context_mapping_features = None
+
+        attention_kwargs, kwargs = groupby("attention_", kwargs, keep_prefix=True)
+
+        self.num_layers = num_layers
+        self.use_context_time = use_context_time
+        self.use_context_features = use_context_features
+        self.use_context_channels = use_context_channels
+        self.use_stft = use_stft
+        self.use_stft_context = use_stft_context
+
+        self.context_features = context_features
+        context_channels_pad_length = num_layers + 1 - len(context_channels)
+        context_channels = context_channels + [0] * context_channels_pad_length
+        self.context_channels = context_channels
+        self.context_embedding_features = context_embedding_features
+
+        if use_context_channels:
+            has_context = [c > 0 for c in context_channels]
+            self.has_context = has_context
+            self.channels_ids = [sum(has_context[:i]) for i in range(len(has_context))]
+
+        assert (
+            len(factors) == num_layers
+            and len(attentions) >= num_layers
+            and len(num_blocks) == num_layers
+        )
+
+        if use_context_time or use_context_features:
+            context_mapping_features = channels * context_features_multiplier
+
+            self.to_mapping = nn.Sequential(
+                nn.Linear(context_mapping_features, context_mapping_features),
+                nn.GELU(),
+                nn.Linear(context_mapping_features, context_mapping_features),
+                nn.GELU(),
+            )
+
+        if use_context_time:
+            assert exists(context_mapping_features)
+            self.to_time = nn.Sequential(
+                TimePositionalEmbedding(
+                    dim=channels, out_features=context_mapping_features
+                ),
+                nn.GELU(),
+            )
+
+        if use_context_features:
+            assert exists(context_features) and exists(context_mapping_features)
+            self.to_features = nn.Sequential(
+                nn.Linear(
+                    in_features=context_features, out_features=context_mapping_features
+                ),
+                nn.GELU(),
+            )
+
+        if use_stft:
+            stft_kwargs, kwargs = groupby("stft_", kwargs)
+            assert "num_fft" in stft_kwargs, "stft_num_fft required if use_stft=True"
+            stft_channels = (stft_kwargs["num_fft"] // 2 + 1) * 2
+            in_channels *= stft_channels
+            out_channels *= stft_channels
+            context_channels[0] *= stft_channels if use_stft_context else 1
+            assert exists(in_channels) and exists(out_channels)
+            self.stft = STFT(**stft_kwargs)
+
+        assert not kwargs, f"Unknown arguments: {', '.join(list(kwargs.keys()))}"
+
+        self.to_in = Patcher(
+            in_channels=in_channels + context_channels[0],
+            out_channels=channels * multipliers[0],
+            patch_size=patch_size,
+            context_mapping_features=context_mapping_features,
+            use_snake=use_snake
+        )
+
+        self.downsamples = nn.ModuleList(
+            [
+                DownsampleBlock1d(
+                    in_channels=channels * multipliers[i],
+                    out_channels=channels * multipliers[i + 1],
+                    context_mapping_features=context_mapping_features,
+                    context_channels=context_channels[i + 1],
+                    context_embedding_features=context_embedding_features,
+                    num_layers=num_blocks[i],
+                    factor=factors[i],
+                    kernel_multiplier=kernel_multiplier_downsample,
+                    num_groups=resnet_groups,
+                    use_pre_downsample=True,
+                    use_skip=True,
+                    use_snake=use_snake,
+                    num_transformer_blocks=attentions[i],
+                    **attention_kwargs,
+                )
+                for i in range(num_layers)
+            ]
+        )
+
+        self.bottleneck = BottleneckBlock1d(
+            channels=channels * multipliers[-1],
+            context_mapping_features=context_mapping_features,
+            context_embedding_features=context_embedding_features,
+            num_groups=resnet_groups,
+            num_transformer_blocks=attentions[-1],
+            use_snake=use_snake,
+            **attention_kwargs,
+        )
+
+        self.upsamples = nn.ModuleList(
+            [
+                UpsampleBlock1d(
+                    in_channels=channels * multipliers[i + 1],
+                    out_channels=channels * multipliers[i],
+                    context_mapping_features=context_mapping_features,
+                    context_embedding_features=context_embedding_features,
+                    num_layers=num_blocks[i] + (1 if attentions[i] else 0),
+                    factor=factors[i],
+                    use_nearest=use_nearest_upsample,
+                    num_groups=resnet_groups,
+                    use_skip_scale=use_skip_scale,
+                    use_pre_upsample=False,
+                    use_skip=True,
+                    use_snake=use_snake,
+                    skip_channels=channels * multipliers[i + 1],
+                    num_transformer_blocks=attentions[i],
+                    **attention_kwargs,
+                )
+                for i in reversed(range(num_layers))
+            ]
+        )
+
+        self.to_out = Unpatcher(
+            in_channels=channels * multipliers[0],
+            out_channels=out_channels,
+            patch_size=patch_size,
+            context_mapping_features=context_mapping_features,
+            use_snake=use_snake
+        )
+
+    def get_channels(
+        self, channels_list: Optional[Sequence[Tensor]] = None, layer: int = 0
+    ) -> Optional[Tensor]:
+        """Gets context channels at `layer` and checks that shape is correct"""
+        use_context_channels = self.use_context_channels and self.has_context[layer]
+        if not use_context_channels:
+            return None
+        assert exists(channels_list), "Missing context"
+        # Get channels index (skipping zero channel contexts)
+        channels_id = self.channels_ids[layer]
+        # Get channels
+        channels = channels_list[channels_id]
+        message = f"Missing context for layer {layer} at index {channels_id}"
+        assert exists(channels), message
+        # Check channels
+        num_channels = self.context_channels[layer]
+        message = f"Expected context with {num_channels} channels at idx {channels_id}"
+        assert channels.shape[1] == num_channels, message
+        # STFT channels if requested
+        channels = self.stft.encode1d(channels) if self.use_stft_context else channels  # type: ignore # noqa
+        return channels
+
+    def get_mapping(
+        self, time: Optional[Tensor] = None, features: Optional[Tensor] = None
+    ) -> Optional[Tensor]:
+        """Combines context time features and features into mapping"""
+        items, mapping = [], None
+        # Compute time features
+        if self.use_context_time:
+            assert_message = "use_context_time=True but no time features provided"
+            assert exists(time), assert_message
+            items += [self.to_time(time)]
+        # Compute features
+        if self.use_context_features:
+            assert_message = "context_features exists but no features provided"
+            assert exists(features), assert_message
+            items += [self.to_features(features)]
+        # Compute joint mapping
+        if self.use_context_time or self.use_context_features:
+            mapping = reduce(torch.stack(items), "n b m -> b m", "sum")
+            mapping = self.to_mapping(mapping)
+        return mapping
+
+    def forward(
+        self,
+        x: Tensor,
+        time: Optional[Tensor] = None,
+        *,
+        features: Optional[Tensor] = None,
+        channels_list: Optional[Sequence[Tensor]] = None,
+        embedding: Optional[Tensor] = None,
+        embedding_mask: Optional[Tensor] = None,
+        causal: Optional[bool] = False,
+    ) -> Tensor:
+        channels = self.get_channels(channels_list, layer=0)
+        # Apply stft if required
+        x = self.stft.encode1d(x) if self.use_stft else x  # type: ignore
+        # Concat context channels at layer 0 if provided
+        x = torch.cat([x, channels], dim=1) if exists(channels) else x
+        # Compute mapping from time and features
+        mapping = self.get_mapping(time, features)
+        x = self.to_in(x, mapping, causal=causal)
+        skips_list = [x]
+
+        for i, downsample in enumerate(self.downsamples):
+            channels = self.get_channels(channels_list, layer=i + 1)
+            x, skips = downsample(
+                x, mapping=mapping, channels=channels, embedding=embedding, embedding_mask=embedding_mask, causal=causal
+            )
+            skips_list += [skips]
+
+        x = self.bottleneck(x, mapping=mapping, embedding=embedding, embedding_mask=embedding_mask, causal=causal)
+
+        for i, upsample in enumerate(self.upsamples):
+            skips = skips_list.pop()
+            x = upsample(x, skips=skips, mapping=mapping, embedding=embedding, embedding_mask=embedding_mask, causal=causal)
+
+        x += skips_list.pop()
+        x = self.to_out(x, mapping, causal=causal)
+        x = self.stft.decode1d(x) if self.use_stft else x
+
+        return x
+
+
+""" Conditioning Modules """
+
+
+class FixedEmbedding(nn.Module):
+    def __init__(self, max_length: int, features: int):
+        super().__init__()
+        self.max_length = max_length
+        self.embedding = nn.Embedding(max_length, features)
+
+    def forward(self, x: Tensor) -> Tensor:
+        batch_size, length, device = *x.shape[0:2], x.device
+        assert_message = "Input sequence length must be <= max_length"
+        assert length <= self.max_length, assert_message
+        position = torch.arange(length, device=device)
+        fixed_embedding = self.embedding(position)
+        fixed_embedding = repeat(fixed_embedding, "n d -> b n d", b=batch_size)
+        return fixed_embedding
+
+
+def rand_bool(shape: Any, proba: float, device: Any = None) -> Tensor:
+    if proba == 1:
+        return torch.ones(shape, device=device, dtype=torch.bool)
+    elif proba == 0:
+        return torch.zeros(shape, device=device, dtype=torch.bool)
+    else:
+        return torch.bernoulli(torch.full(shape, proba, device=device)).to(torch.bool)
+
+
+class UNetCFG1d(UNet1d):
+
+    """UNet1d with Classifier-Free Guidance"""
+
+    def __init__(
+        self,
+        context_embedding_max_length: int,
+        context_embedding_features: int,
+        use_xattn_time: bool = False,
+        **kwargs,
+    ):
+        super().__init__(
+            context_embedding_features=context_embedding_features, **kwargs
+        )
+
+        self.use_xattn_time = use_xattn_time
+
+        if use_xattn_time:
+            assert exists(context_embedding_features)
+            self.to_time_embedding = nn.Sequential(
+                TimePositionalEmbedding(
+                    dim=kwargs["channels"], out_features=context_embedding_features
+                ),
+                nn.GELU(),
+            )
+
+            context_embedding_max_length += 1   # Add one for time embedding
+
+        self.fixed_embedding = FixedEmbedding(
+            max_length=context_embedding_max_length, features=context_embedding_features
+        )
+
+    def forward(  # type: ignore
+        self,
+        x: Tensor,
+        time: Tensor,
+        *,
+        embedding: Tensor,
+        embedding_mask: Optional[Tensor] = None,
+        embedding_scale: float = 1.0,
+        embedding_mask_proba: float = 0.0,
+        batch_cfg: bool = False,
+        rescale_cfg: bool = False,
+        scale_phi: float = 0.4,
+        negative_embedding: Optional[Tensor] = None,
+        negative_embedding_mask: Optional[Tensor] = None,
+        **kwargs,
+    ) -> Tensor:
+        b, device = embedding.shape[0], embedding.device
+
+        if self.use_xattn_time:
+            embedding = torch.cat([embedding, self.to_time_embedding(time).unsqueeze(1)], dim=1)
+
+            if embedding_mask is not None:
+                embedding_mask = torch.cat([embedding_mask, torch.ones((b, 1), device=device)], dim=1)
+
+        fixed_embedding = self.fixed_embedding(embedding)
+
+        if embedding_mask_proba > 0.0:
+            # Randomly mask embedding
+            batch_mask = rand_bool(
+                shape=(b, 1, 1), proba=embedding_mask_proba, device=device
+            )
+            embedding = torch.where(batch_mask, fixed_embedding, embedding)
+
+        if embedding_scale != 1.0:
+            if batch_cfg:
+                batch_x = torch.cat([x, x], dim=0)
+                batch_time = torch.cat([time, time], dim=0)
+
+                if negative_embedding is not None:
+                    if negative_embedding_mask is not None:
+                        negative_embedding_mask = negative_embedding_mask.to(torch.bool).unsqueeze(2)
+
+                        negative_embedding = torch.where(negative_embedding_mask, negative_embedding, fixed_embedding)
+                    
+                    batch_embed = torch.cat([embedding, negative_embedding], dim=0)
+
+                else:
+                    batch_embed = torch.cat([embedding, fixed_embedding], dim=0)
+
+                batch_mask = None
+                if embedding_mask is not None:
+                    batch_mask = torch.cat([embedding_mask, embedding_mask], dim=0)
+
+                batch_features = None
+                features = kwargs.pop("features", None)
+                if self.use_context_features:
+                    batch_features = torch.cat([features, features], dim=0)
+
+                batch_channels = None
+                channels_list = kwargs.pop("channels_list", None)
+                if self.use_context_channels:
+                    batch_channels = []
+                    for channels in channels_list:
+                        batch_channels += [torch.cat([channels, channels], dim=0)]
+
+                # Compute both normal and fixed embedding outputs
+                batch_out = super().forward(batch_x, batch_time, embedding=batch_embed, embedding_mask=batch_mask, features=batch_features, channels_list=batch_channels, **kwargs)
+                out, out_masked = batch_out.chunk(2, dim=0)
+           
+            else:
+                # Compute both normal and fixed embedding outputs
+                out = super().forward(x, time, embedding=embedding, embedding_mask=embedding_mask, **kwargs)
+                out_masked = super().forward(x, time, embedding=fixed_embedding, embedding_mask=embedding_mask, **kwargs)
+
+            out_cfg = out_masked + (out - out_masked) * embedding_scale
+
+            if rescale_cfg:
+
+                out_std = out.std(dim=1, keepdim=True)
+                out_cfg_std = out_cfg.std(dim=1, keepdim=True)
+
+                return scale_phi * (out_cfg * (out_std/out_cfg_std)) + (1-scale_phi) * out_cfg
+
+            else:
+
+                return out_cfg
+                
+        else:
+            return super().forward(x, time, embedding=embedding, embedding_mask=embedding_mask, **kwargs)
+
+
+class UNetNCCA1d(UNet1d):
+
+    """UNet1d with Noise Channel Conditioning Augmentation"""
+
+    def __init__(self, context_features: int, **kwargs):
+        super().__init__(context_features=context_features, **kwargs)
+        self.embedder = NumberEmbedder(features=context_features)
+
+    def expand(self, x: Any, shape: Tuple[int, ...]) -> Tensor:
+        x = x if torch.is_tensor(x) else torch.tensor(x)
+        return x.expand(shape)
+
+    def forward(  # type: ignore
+        self,
+        x: Tensor,
+        time: Tensor,
+        *,
+        channels_list: Sequence[Tensor],
+        channels_augmentation: Union[
+            bool, Sequence[bool], Sequence[Sequence[bool]], Tensor
+        ] = False,
+        channels_scale: Union[
+            float, Sequence[float], Sequence[Sequence[float]], Tensor
+        ] = 0,
+        **kwargs,
+    ) -> Tensor:
+        b, n = x.shape[0], len(channels_list)
+        channels_augmentation = self.expand(channels_augmentation, shape=(b, n)).to(x)
+        channels_scale = self.expand(channels_scale, shape=(b, n)).to(x)
+
+        # Augmentation (for each channel list item)
+        for i in range(n):
+            scale = channels_scale[:, i] * channels_augmentation[:, i]
+            scale = rearrange(scale, "b -> b 1 1")
+            item = channels_list[i]
+            channels_list[i] = torch.randn_like(item) * scale + item * (1 - scale)  # type: ignore # noqa
+
+        # Scale embedding (sum reduction if more than one channel list item)
+        channels_scale_emb = self.embedder(channels_scale)
+        channels_scale_emb = reduce(channels_scale_emb, "b n d -> b d", "sum")
+
+        return super().forward(
+            x=x,
+            time=time,
+            channels_list=channels_list,
+            features=channels_scale_emb,
+            **kwargs,
+        )
+
+
+class UNetAll1d(UNetCFG1d, UNetNCCA1d):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def forward(self, *args, **kwargs):  # type: ignore
+        return UNetCFG1d.forward(self, *args, **kwargs)
+
+
+def XUNet1d(type: str = "base", **kwargs) -> UNet1d:
+    if type == "base":
+        return UNet1d(**kwargs)
+    elif type == "all":
+        return UNetAll1d(**kwargs)
+    elif type == "cfg":
+        return UNetCFG1d(**kwargs)
+    elif type == "ncca":
+        return UNetNCCA1d(**kwargs)
+    else:
+        raise ValueError(f"Unknown XUNet1d type: {type}")
+
+class NumberEmbedder(nn.Module):
+    def __init__(
+        self,
+        features: int,
+        dim: int = 256,
+    ):
+        super().__init__()
+        self.features = features
+        self.embedding = TimePositionalEmbedding(dim=dim, out_features=features)
+
+    def forward(self, x: Union[List[float], Tensor]) -> Tensor:
+        if not torch.is_tensor(x):
+            device = next(self.embedding.parameters()).device
+            x = torch.tensor(x, device=device)
+        assert isinstance(x, Tensor)
+        shape = x.shape
+        x = rearrange(x, "... -> (...)")
+        embedding = self.embedding(x)
+        x = embedding.view(*shape, self.features)
+        return x  # type: ignore
+
+
+"""
+Audio Transforms
+"""
+
+
+class STFT(nn.Module):
+    """Helper for torch stft and istft"""
+
+    def __init__(
+        self,
+        num_fft: int = 1023,
+        hop_length: int = 256,
+        window_length: Optional[int] = None,
+        length: Optional[int] = None,
+        use_complex: bool = False,
+    ):
+        super().__init__()
+        self.num_fft = num_fft
+        self.hop_length = default(hop_length, floor(num_fft // 4))
+        self.window_length = default(window_length, num_fft)
+        self.length = length
+        self.register_buffer("window", torch.hann_window(self.window_length))
+        self.use_complex = use_complex
+
+    def encode(self, wave: Tensor) -> Tuple[Tensor, Tensor]:
+        b = wave.shape[0]
+        wave = rearrange(wave, "b c t -> (b c) t")
+
+        stft = torch.stft(
+            wave,
+            n_fft=self.num_fft,
+            hop_length=self.hop_length,
+            win_length=self.window_length,
+            window=self.window,  # type: ignore
+            return_complex=True,
+            normalized=True,
+        )
+
+        if self.use_complex:
+            # Returns real and imaginary
+            stft_a, stft_b = stft.real, stft.imag
+        else:
+            # Returns magnitude and phase matrices
+            magnitude, phase = torch.abs(stft), torch.angle(stft)
+            stft_a, stft_b = magnitude, phase
+
+        return rearrange_many((stft_a, stft_b), "(b c) f l -> b c f l", b=b)
+
+    def decode(self, stft_a: Tensor, stft_b: Tensor) -> Tensor:
+        b, l = stft_a.shape[0], stft_a.shape[-1]  # noqa
+        length = closest_power_2(l * self.hop_length)
+
+        stft_a, stft_b = rearrange_many((stft_a, stft_b), "b c f l -> (b c) f l")
+
+        if self.use_complex:
+            real, imag = stft_a, stft_b
+        else:
+            magnitude, phase = stft_a, stft_b
+            real, imag = magnitude * torch.cos(phase), magnitude * torch.sin(phase)
+
+        stft = torch.stack([real, imag], dim=-1)
+
+        wave = torch.istft(
+            stft,
+            n_fft=self.num_fft,
+            hop_length=self.hop_length,
+            win_length=self.window_length,
+            window=self.window,  # type: ignore
+            length=default(self.length, length),
+            normalized=True,
+        )
+
+        return rearrange(wave, "(b c) t -> b c t", b=b)
+
+    def encode1d(
+        self, wave: Tensor, stacked: bool = True
+    ) -> Union[Tensor, Tuple[Tensor, Tensor]]:
+        stft_a, stft_b = self.encode(wave)
+        stft_a, stft_b = rearrange_many((stft_a, stft_b), "b c f l -> b (c f) l")
+        return torch.cat((stft_a, stft_b), dim=1) if stacked else (stft_a, stft_b)
+
+    def decode1d(self, stft_pair: Tensor) -> Tensor:
+        f = self.num_fft // 2 + 1
+        stft_a, stft_b = stft_pair.chunk(chunks=2, dim=1)
+        stft_a, stft_b = rearrange_many((stft_a, stft_b), "b (c f) l -> b c f l", f=f)
+        return self.decode(stft_a, stft_b)

stable_audio_tools/models/autoencoders.py

@@ -0,0 +1,800 @@
+import torch
+import math
+import numpy as np
+
+from torch import nn, sin, pow
+from torch.nn import functional as F
+from torchaudio import transforms as T
+from alias_free_torch import Activation1d
+from dac.nn.layers import WNConv1d, WNConvTranspose1d
+from typing import List, Literal, Dict, Any, Callable
+from einops import rearrange
+
+from ..inference.sampling import sample
+from ..inference.utils import prepare_audio
+from .blocks import SnakeBeta
+from .bottleneck import Bottleneck, DiscreteBottleneck
+from .diffusion import ConditionedDiffusionModel, DAU1DCondWrapper, UNet1DCondWrapper, DiTWrapper
+from .factory import create_pretransform_from_config, create_bottleneck_from_config
+from .pretransforms import Pretransform, AutoencoderPretransform
+
+def checkpoint(function, *args, **kwargs):
+    kwargs.setdefault("use_reentrant", False)
+    return torch.utils.checkpoint.checkpoint(function, *args, **kwargs)
+
+def get_activation(activation: Literal["elu", "snake", "none"], antialias=False, channels=None) -> nn.Module:
+    if activation == "elu":
+        act = nn.ELU()
+    elif activation == "snake":
+        act = SnakeBeta(channels)
+    elif activation == "none":
+        act = nn.Identity()
+    else:
+        raise ValueError(f"Unknown activation {activation}")
+    
+    if antialias:
+        act = Activation1d(act)
+    
+    return act
+
+class ResidualUnit(nn.Module):
+    def __init__(self, in_channels, out_channels, dilation, use_snake=False, antialias_activation=False):
+        super().__init__()
+        
+        self.dilation = dilation
+
+        act = get_activation("snake" if use_snake else "elu", antialias=antialias_activation, channels=out_channels)
+
+        padding = (dilation * (7-1)) // 2
+
+        self.layers = nn.Sequential(
+            act,
+            WNConv1d(in_channels=in_channels, out_channels=out_channels,
+                      kernel_size=7, dilation=dilation, padding=padding),
+            act,
+            WNConv1d(in_channels=out_channels, out_channels=out_channels,
+                      kernel_size=1)
+        )
+
+    def forward(self, x):
+        res = x
+        
+        # Disable checkpoint until tensor mismatch is fixed
+        #x = checkpoint(self.layers, x)
+        x = self.layers(x)
+
+
+class EncoderBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, stride, use_snake=False, antialias_activation=False):
+        super().__init__()
+
+        act = get_activation("snake" if use_snake else "elu", antialias=antialias_activation, channels=in_channels)
+
+        self.layers = nn.Sequential(
+            ResidualUnit(in_channels=in_channels,
+                         out_channels=in_channels, dilation=1, use_snake=use_snake),
+            ResidualUnit(in_channels=in_channels,
+                         out_channels=in_channels, dilation=3, use_snake=use_snake),
+            ResidualUnit(in_channels=in_channels,
+                         out_channels=in_channels, dilation=9, use_snake=use_snake),
+            act,
+            WNConv1d(in_channels=in_channels, out_channels=out_channels,
+                      kernel_size=2*stride, stride=stride, padding=math.ceil(stride/2)),
+        )
+
+    def forward(self, x):
+        return self.layers(x)
+
+class DecoderBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, stride, use_snake=False, antialias_activation=False, use_nearest_upsample=False):
+        super().__init__()
+
+        if use_nearest_upsample:
+            upsample_layer = nn.Sequential(
+                nn.Upsample(scale_factor=stride, mode="nearest"),
+                WNConv1d(in_channels=in_channels,
+                        out_channels=out_channels, 
+                        kernel_size=2*stride,
+                        stride=1,
+                        bias=False,
+                        padding='same')
+            )
+        else:
+            upsample_layer = WNConvTranspose1d(in_channels=in_channels,
+                               out_channels=out_channels,
+                               kernel_size=2*stride, stride=stride, padding=math.ceil(stride/2))
+
+        act = get_activation("snake" if use_snake else "elu", antialias=antialias_activation, channels=in_channels)
+
+        self.layers = nn.Sequential(
+            act,
+            upsample_layer,
+            ResidualUnit(in_channels=out_channels, out_channels=out_channels,
+                         dilation=1, use_snake=use_snake),
+            ResidualUnit(in_channels=out_channels, out_channels=out_channels,
+                         dilation=3, use_snake=use_snake),
+            ResidualUnit(in_channels=out_channels, out_channels=out_channels,
+                         dilation=9, use_snake=use_snake),
+        )
+
+    def forward(self, x):
+        return self.layers(x)
+
+class OobleckEncoder(nn.Module):
+    def __init__(self, 
+                 in_channels=2, 
+                 channels=128, 
+                 latent_dim=32, 
+                 c_mults = [1, 2, 4, 8], 
+                 strides = [2, 4, 8, 8],
+                 use_snake=False,
+                 antialias_activation=False
+        ):
+        super().__init__()
+          
+        c_mults = [1] + c_mults
+
+        self.depth = len(c_mults)
+
+        layers = [
+            WNConv1d(in_channels=in_channels, out_channels=c_mults[0] * channels, kernel_size=7, padding=3)
+        ]
+        
+        for i in range(self.depth-1):
+            layers += [EncoderBlock(in_channels=c_mults[i]*channels, out_channels=c_mults[i+1]*channels, stride=strides[i], use_snake=use_snake)]
+
+        layers += [
+            get_activation("snake" if use_snake else "elu", antialias=antialias_activation, channels=c_mults[-1] * channels),
+            WNConv1d(in_channels=c_mults[-1]*channels, out_channels=latent_dim, kernel_size=3, padding=1)
+        ]
+
+        self.layers = nn.Sequential(*layers)
+
+    def forward(self, x):
+        return self.layers(x)
+
+
+class OobleckDecoder(nn.Module):
+    def __init__(self, 
+                 out_channels=2, 
+                 channels=128, 
+                 latent_dim=32, 
+                 c_mults = [1, 2, 4, 8], 
+                 strides = [2, 4, 8, 8],
+                 use_snake=False,
+                 antialias_activation=False,
+                 use_nearest_upsample=False,
+                 final_tanh=True):
+        super().__init__()
+
+        c_mults = [1] + c_mults
+        
+        self.depth = len(c_mults)
+
+        layers = [
+            WNConv1d(in_channels=latent_dim, out_channels=c_mults[-1]*channels, kernel_size=7, padding=3),
+        ]
+        
+        for i in range(self.depth-1, 0, -1):
+            layers += [DecoderBlock(
+                in_channels=c_mults[i]*channels, 
+                out_channels=c_mults[i-1]*channels, 
+                stride=strides[i-1], 
+                use_snake=use_snake, 
+                antialias_activation=antialias_activation,
+                use_nearest_upsample=use_nearest_upsample
+                )
+            ]
+
+        layers += [
+            get_activation("snake" if use_snake else "elu", antialias=antialias_activation, channels=c_mults[0] * channels),
+            WNConv1d(in_channels=c_mults[0] * channels, out_channels=out_channels, kernel_size=7, padding=3, bias=False),
+            nn.Tanh() if final_tanh else nn.Identity()
+        ]
+
+        self.layers = nn.Sequential(*layers)
+
+    def forward(self, x):
+        return self.layers(x)
+
+class DACEncoderWrapper(nn.Module):
+    def __init__(self, in_channels=1, **kwargs):
+        super().__init__()
+
+        from dac.model.dac import Encoder as DACEncoder
+
+        latent_dim = kwargs.pop("latent_dim", None)
+
+        encoder_out_dim = kwargs["d_model"] * (2 ** len(kwargs["strides"]))
+        self.encoder = DACEncoder(d_latent=encoder_out_dim, **kwargs)
+        self.latent_dim = latent_dim
+
+        # Latent-dim support was added to DAC after this was first written, and implemented differently, so this is for backwards compatibility
+        self.proj_out = nn.Conv1d(self.encoder.enc_dim, latent_dim, kernel_size=1) if latent_dim is not None else nn.Identity()
+
+        if in_channels != 1:
+            self.encoder.block[0] = WNConv1d(in_channels, kwargs.get("d_model", 64), kernel_size=7, padding=3)
+
+    def forward(self, x):
+        x = self.encoder(x)
+        x = self.proj_out(x)
+        return x
+
+class DACDecoderWrapper(nn.Module):
+    def __init__(self, latent_dim, out_channels=1, **kwargs):
+        super().__init__()
+
+        from dac.model.dac import Decoder as DACDecoder
+
+        self.decoder = DACDecoder(**kwargs, input_channel = latent_dim, d_out=out_channels)
+
+        self.latent_dim = latent_dim
+
+    def forward(self, x):
+        return self.decoder(x)
+
+class AudioAutoencoder(nn.Module):
+    def __init__(
+        self,
+        encoder,
+        decoder,
+        latent_dim,
+        downsampling_ratio,
+        sample_rate,
+        io_channels=2,
+        bottleneck: Bottleneck = None,
+        pretransform: Pretransform = None,
+        in_channels = None,
+        out_channels = None,
+        soft_clip = False
+    ):
+        super().__init__()
+
+        self.downsampling_ratio = downsampling_ratio
+        self.sample_rate = sample_rate
+
+        self.latent_dim = latent_dim
+        self.io_channels = io_channels
+        self.in_channels = io_channels
+        self.out_channels = io_channels
+
+        self.min_length = self.downsampling_ratio
+
+        if in_channels is not None:
+            self.in_channels = in_channels
+
+        if out_channels is not None:
+            self.out_channels = out_channels
+
+        self.bottleneck = bottleneck
+
+        self.encoder = encoder
+
+        self.decoder = decoder
+
+        self.pretransform = pretransform
+
+        self.soft_clip = soft_clip
+ 
+        self.is_discrete = self.bottleneck is not None and self.bottleneck.is_discrete
+
+    def encode(self, audio, return_info=False, skip_pretransform=False, iterate_batch=False, **kwargs):
+
+        info = {}
+
+        if self.pretransform is not None and not skip_pretransform:
+            if self.pretransform.enable_grad:
+                if iterate_batch:
+                    audios = []
+                    for i in range(audio.shape[0]):
+                        audios.append(self.pretransform.encode(audio[i:i+1]))
+                    audio = torch.cat(audios, dim=0)
+                else:
+                    audio = self.pretransform.encode(audio)
+            else:
+                with torch.no_grad():
+                    if iterate_batch:
+                        audios = []
+                        for i in range(audio.shape[0]):
+                            audios.append(self.pretransform.encode(audio[i:i+1]))
+                        audio = torch.cat(audios, dim=0)
+                    else:
+                        audio = self.pretransform.encode(audio)
+
+        if self.encoder is not None:
+            if iterate_batch:
+                latents = []
+                for i in range(audio.shape[0]):
+                    latents.append(self.encoder(audio[i:i+1]))
+                latents = torch.cat(latents, dim=0)
+            else:
+                latents = self.encoder(audio)
+        else:
+            latents = audio
+
+        if self.bottleneck is not None:
+            # TODO: Add iterate batch logic, needs to merge the info dicts
+            latents, bottleneck_info = self.bottleneck.encode(latents, return_info=True, **kwargs)
+
+            info.update(bottleneck_info)
+        
+        if return_info:
+            return latents, info
+
+        return latents
+
+    def decode(self, latents, iterate_batch=False, **kwargs):
+
+        if self.bottleneck is not None:
+            if iterate_batch:
+                decoded = []
+                for i in range(latents.shape[0]):
+                    decoded.append(self.bottleneck.decode(latents[i:i+1]))
+                decoded = torch.cat(decoded, dim=0)
+            else:
+                latents = self.bottleneck.decode(latents)
+
+        if iterate_batch:
+            decoded = []
+            for i in range(latents.shape[0]):
+                decoded.append(self.decoder(latents[i:i+1]))
+            decoded = torch.cat(decoded, dim=0)
+        else:
+            decoded = self.decoder(latents, **kwargs)
+
+        if self.pretransform is not None:
+            if self.pretransform.enable_grad:
+                if iterate_batch:
+                    decodeds = []
+                    for i in range(decoded.shape[0]):
+                        decodeds.append(self.pretransform.decode(decoded[i:i+1]))
+                    decoded = torch.cat(decodeds, dim=0)
+                else:
+                    decoded = self.pretransform.decode(decoded)
+            else:
+                with torch.no_grad():
+                    if iterate_batch:
+                        decodeds = []
+                        for i in range(latents.shape[0]):
+                            decodeds.append(self.pretransform.decode(decoded[i:i+1]))
+                        decoded = torch.cat(decodeds, dim=0)
+                    else:
+                        decoded = self.pretransform.decode(decoded)
+
+        if self.soft_clip:
+            decoded = torch.tanh(decoded)
+        
+        return decoded
+          
+    def decode_tokens(self, tokens, **kwargs):
+        '''
+        Decode discrete tokens to audio
+        Only works with discrete autoencoders
+        '''
+
+        assert isinstance(self.bottleneck, DiscreteBottleneck), "decode_tokens only works with discrete autoencoders"
+
+        latents = self.bottleneck.decode_tokens(tokens, **kwargs)
+
+        return self.decode(latents, **kwargs)
+        
+    
+    def preprocess_audio_for_encoder(self, audio, in_sr):
+        '''
+        Preprocess single audio tensor (Channels x Length) to be compatible with the encoder.
+        If the model is mono, stereo audio will be converted to mono.
+        Audio will be silence-padded to be a multiple of the model's downsampling ratio.
+        Audio will be resampled to the model's sample rate. 
+        The output will have batch size 1 and be shape (1 x Channels x Length)
+        '''
+        return self.preprocess_audio_list_for_encoder([audio], [in_sr])
+
+    def preprocess_audio_list_for_encoder(self, audio_list, in_sr_list):
+        '''
+        Preprocess a [list] of audio (Channels x Length) into a batch tensor to be compatable with the encoder. 
+        The audio in that list can be of different lengths and channels. 
+        in_sr can be an integer or list. If it's an integer it will be assumed it is the input sample_rate for every audio.
+        All audio will be resampled to the model's sample rate. 
+        Audio will be silence-padded to the longest length, and further padded to be a multiple of the model's downsampling ratio. 
+        If the model is mono, all audio will be converted to mono. 
+        The output will be a tensor of shape (Batch x Channels x Length)
+        '''
+        batch_size = len(audio_list)
+        if isinstance(in_sr_list, int):
+            in_sr_list = [in_sr_list]*batch_size
+        assert len(in_sr_list) == batch_size, "list of sample rates must be the same length of audio_list"
+        new_audio = []
+        max_length = 0
+        # resample & find the max length
+        for i in range(batch_size):
+            audio = audio_list[i]
+            in_sr = in_sr_list[i]
+            if len(audio.shape) == 3 and audio.shape[0] == 1:
+                # batchsize 1 was given by accident. Just squeeze it.
+                audio = audio.squeeze(0)
+            elif len(audio.shape) == 1:
+                # Mono signal, channel dimension is missing, unsqueeze it in
+                audio = audio.unsqueeze(0)
+            assert len(audio.shape)==2, "Audio should be shape (Channels x Length) with no batch dimension" 
+            # Resample audio
+            if in_sr != self.sample_rate:
+                resample_tf = T.Resample(in_sr, self.sample_rate).to(audio.device)
+                audio = resample_tf(audio)
+            new_audio.append(audio)
+            if audio.shape[-1] > max_length:
+                max_length = audio.shape[-1]
+        # Pad every audio to the same length, multiple of model's downsampling ratio
+        padded_audio_length = max_length + (self.min_length - (max_length % self.min_length)) % self.min_length
+        for i in range(batch_size):
+            # Pad it & if necessary, mixdown/duplicate stereo/mono channels to support model
+            new_audio[i] = prepare_audio(new_audio[i], in_sr=in_sr, target_sr=in_sr, target_length=padded_audio_length, 
+                target_channels=self.in_channels, device=new_audio[i].device).squeeze(0)
+        # convert to tensor 
+        return torch.stack(new_audio) 
+
+    def encode_audio(self, audio, chunked=False, overlap=32, chunk_size=128, **kwargs):
+        '''
+        Encode audios into latents. Audios should already be preprocesed by preprocess_audio_for_encoder.
+        If chunked is True, split the audio into chunks of a given maximum size chunk_size, with given overlap.
+        Overlap and chunk_size params are both measured in number of latents (not audio samples) 
+        # and therefore you likely could use the same values with decode_audio. 
+        A overlap of zero will cause discontinuity artefacts. Overlap should be => receptive field size. 
+        Every autoencoder will have a different receptive field size, and thus ideal overlap.
+        You can determine it empirically by diffing unchunked vs chunked output and looking at maximum diff.
+        The final chunk may have a longer overlap in order to keep chunk_size consistent for all chunks.
+        Smaller chunk_size uses less memory, but more compute.
+        The chunk_size vs memory tradeoff isn't linear, and possibly depends on the GPU and CUDA version
+        For example, on a A6000 chunk_size 128 is overall faster than 256 and 512 even though it has more chunks
+        '''
+        if not chunked:
+            # default behavior. Encode the entire audio in parallel
+            return self.encode(audio, **kwargs)
+        else:
+            # CHUNKED ENCODING
+            # samples_per_latent is just the downsampling ratio (which is also the upsampling ratio)
+            samples_per_latent = self.downsampling_ratio
+            total_size = audio.shape[2] # in samples
+            batch_size = audio.shape[0]
+            chunk_size *= samples_per_latent # converting metric in latents to samples
+            overlap *= samples_per_latent # converting metric in latents to samples
+            hop_size = chunk_size - overlap
+            chunks = []
+            for i in range(0, total_size - chunk_size + 1, hop_size):
+                chunk = audio[:,:,i:i+chunk_size]
+                chunks.append(chunk)
+            if i+chunk_size != total_size:
+                # Final chunk
+                chunk = audio[:,:,-chunk_size:]
+                chunks.append(chunk)
+            chunks = torch.stack(chunks)
+            num_chunks = chunks.shape[0]
+            # Note: y_size might be a different value from the latent length used in diffusion training
+            # because we can encode audio of varying lengths
+            # However, the audio should've been padded to a multiple of samples_per_latent by now.
+            y_size = total_size // samples_per_latent
+            # Create an empty latent, we will populate it with chunks as we encode them
+            y_final = torch.zeros((batch_size,self.latent_dim,y_size)).to(audio.device)
+            for i in range(num_chunks):
+                x_chunk = chunks[i,:]
+                # encode the chunk
+                y_chunk = self.encode(x_chunk)
+                # figure out where to put the audio along the time domain
+                if i == num_chunks-1:
+                    # final chunk always goes at the end
+                    t_end = y_size
+                    t_start = t_end - y_chunk.shape[2]
+                else:
+                    t_start = i * hop_size // samples_per_latent
+                    t_end = t_start + chunk_size // samples_per_latent
+                #  remove the edges of the overlaps
+                ol = overlap//samples_per_latent//2
+                chunk_start = 0
+                chunk_end = y_chunk.shape[2]
+                if i > 0:
+                    # no overlap for the start of the first chunk
+                    t_start += ol
+                    chunk_start += ol
+                if i < num_chunks-1:
+                    # no overlap for the end of the last chunk
+                    t_end -= ol
+                    chunk_end -= ol
+                # paste the chunked audio into our y_final output audio
+                y_final[:,:,t_start:t_end] = y_chunk[:,:,chunk_start:chunk_end]
+            return y_final
+    
+    def decode_audio(self, latents, chunked=False, overlap=32, chunk_size=128, **kwargs):
+        '''
+        Decode latents to audio. 
+        If chunked is True, split the latents into chunks of a given maximum size chunk_size, with given overlap, both of which are measured in number of latents. 
+        A overlap of zero will cause discontinuity artefacts. Overlap should be => receptive field size. 
+        Every autoencoder will have a different receptive field size, and thus ideal overlap.
+        You can determine it empirically by diffing unchunked vs chunked audio and looking at maximum diff.
+        The final chunk may have a longer overlap in order to keep chunk_size consistent for all chunks.
+        Smaller chunk_size uses less memory, but more compute.
+        The chunk_size vs memory tradeoff isn't linear, and possibly depends on the GPU and CUDA version
+        For example, on a A6000 chunk_size 128 is overall faster than 256 and 512 even though it has more chunks
+        '''
+        if not chunked:
+            # default behavior. Decode the entire latent in parallel
+            return self.decode(latents, **kwargs)
+        else:
+            # chunked decoding
+            hop_size = chunk_size - overlap
+            total_size = latents.shape[2]
+            batch_size = latents.shape[0]
+            chunks = []
+            for i in range(0, total_size - chunk_size + 1, hop_size):
+                chunk = latents[:,:,i:i+chunk_size]
+                chunks.append(chunk)
+            if i+chunk_size != total_size:
+                # Final chunk
+                chunk = latents[:,:,-chunk_size:]
+                chunks.append(chunk)
+            chunks = torch.stack(chunks)
+            num_chunks = chunks.shape[0]
+            # samples_per_latent is just the downsampling ratio
+            samples_per_latent = self.downsampling_ratio
+            # Create an empty waveform, we will populate it with chunks as decode them
+            y_size = total_size * samples_per_latent
+            y_final = torch.zeros((batch_size,self.out_channels,y_size)).to(latents.device)
+            for i in range(num_chunks):
+                x_chunk = chunks[i,:]
+                # decode the chunk
+                y_chunk = self.decode(x_chunk)
+                # figure out where to put the audio along the time domain
+                if i == num_chunks-1:
+                    # final chunk always goes at the end
+                    t_end = y_size
+                    t_start = t_end - y_chunk.shape[2]
+                else:
+                    t_start = i * hop_size * samples_per_latent
+                    t_end = t_start + chunk_size * samples_per_latent
+                #  remove the edges of the overlaps
+                ol = (overlap//2) * samples_per_latent
+                chunk_start = 0
+                chunk_end = y_chunk.shape[2]
+                if i > 0:
+                    # no overlap for the start of the first chunk
+                    t_start += ol
+                    chunk_start += ol
+                if i < num_chunks-1:
+                    # no overlap for the end of the last chunk
+                    t_end -= ol
+                    chunk_end -= ol
+                # paste the chunked audio into our y_final output audio
+                y_final[:,:,t_start:t_end] = y_chunk[:,:,chunk_start:chunk_end]
+            return y_final
+
+    
+class DiffusionAutoencoder(AudioAutoencoder):
+    def __init__(
+        self,
+        diffusion: ConditionedDiffusionModel,
+        diffusion_downsampling_ratio,
+        *args,
+        **kwargs
+    ):
+        super().__init__(*args, **kwargs)
+
+        self.diffusion = diffusion
+
+        self.min_length = self.downsampling_ratio * diffusion_downsampling_ratio
+
+        if self.encoder is not None:
+            # Shrink the initial encoder parameters to avoid saturated latents
+            with torch.no_grad():
+                for param in self.encoder.parameters():
+                    param *= 0.5
+
+    def decode(self, latents, steps=100):
+
+        upsampled_length = latents.shape[2] * self.downsampling_ratio
+
+        if self.bottleneck is not None:
+            latents = self.bottleneck.decode(latents)
+
+        if self.decoder is not None:
+            latents = self.decode(latents)
+    
+        # Upsample latents to match diffusion length
+        if latents.shape[2] != upsampled_length:
+            latents = F.interpolate(latents, size=upsampled_length, mode='nearest')
+
+        noise = torch.randn(latents.shape[0], self.io_channels, upsampled_length, device=latents.device)
+        decoded = sample(self.diffusion, noise, steps, 0, input_concat_cond=latents)
+
+        if self.pretransform is not None:
+            if self.pretransform.enable_grad:
+                decoded = self.pretransform.decode(decoded)
+            else:
+                with torch.no_grad():
+                    decoded = self.pretransform.decode(decoded)
+
+        return decoded
+        
+# AE factories
+
+def create_encoder_from_config(encoder_config: Dict[str, Any]):
+    encoder_type = encoder_config.get("type", None)
+    assert encoder_type is not None, "Encoder type must be specified"
+
+    if encoder_type == "oobleck":
+        encoder = OobleckEncoder(
+            **encoder_config["config"]
+        )
+    
+    elif encoder_type == "seanet":
+        from encodec.modules import SEANetEncoder
+        seanet_encoder_config = encoder_config["config"]
+
+        #SEANet encoder expects strides in reverse order
+        seanet_encoder_config["ratios"] = list(reversed(seanet_encoder_config.get("ratios", [2, 2, 2, 2, 2])))
+        encoder = SEANetEncoder(
+            **seanet_encoder_config
+        )
+    elif encoder_type == "dac":
+        dac_config = encoder_config["config"]
+
+        encoder = DACEncoderWrapper(**dac_config)
+    elif encoder_type == "local_attn":
+        from .local_attention import TransformerEncoder1D
+
+        local_attn_config = encoder_config["config"]
+
+        encoder = TransformerEncoder1D(
+            **local_attn_config
+        )
+    else:
+        raise ValueError(f"Unknown encoder type {encoder_type}")
+    
+    requires_grad = encoder_config.get("requires_grad", True)
+    if not requires_grad:
+        for param in encoder.parameters():
+            param.requires_grad = False
+
+    return encoder
+
+def create_decoder_from_config(decoder_config: Dict[str, Any]):
+    decoder_type = decoder_config.get("type", None)
+    assert decoder_type is not None, "Decoder type must be specified"
+
+    if decoder_type == "oobleck":
+        decoder = OobleckDecoder(
+            **decoder_config["config"]
+        )
+    elif decoder_type == "seanet":
+        from encodec.modules import SEANetDecoder
+
+        decoder = SEANetDecoder(
+            **decoder_config["config"]
+        )
+    elif decoder_type == "dac":
+        dac_config = decoder_config["config"]
+
+        decoder = DACDecoderWrapper(**dac_config)
+    elif decoder_type == "local_attn":
+        from .local_attention import TransformerDecoder1D
+
+        local_attn_config = decoder_config["config"]
+
+        decoder = TransformerDecoder1D(
+            **local_attn_config
+        )
+    else:
+        raise ValueError(f"Unknown decoder type {decoder_type}")
+    
+    requires_grad = decoder_config.get("requires_grad", True)
+    if not requires_grad:
+        for param in decoder.parameters():
+            param.requires_grad = False
+
+    return decoder
+
+def create_autoencoder_from_config(config: Dict[str, Any]):
+    
+    ae_config = config["model"]
+
+    encoder = create_encoder_from_config(ae_config["encoder"])
+    decoder = create_decoder_from_config(ae_config["decoder"])
+
+    bottleneck = ae_config.get("bottleneck", None)
+
+    latent_dim = ae_config.get("latent_dim", None)
+    assert latent_dim is not None, "latent_dim must be specified in model config"
+    downsampling_ratio = ae_config.get("downsampling_ratio", None)
+    assert downsampling_ratio is not None, "downsampling_ratio must be specified in model config"
+    io_channels = ae_config.get("io_channels", None)
+    assert io_channels is not None, "io_channels must be specified in model config"
+    sample_rate = config.get("sample_rate", None)
+    assert sample_rate is not None, "sample_rate must be specified in model config"
+
+    in_channels = ae_config.get("in_channels", None)
+    out_channels = ae_config.get("out_channels", None)
+
+    pretransform = ae_config.get("pretransform", None)
+
+    if pretransform is not None:
+        pretransform = create_pretransform_from_config(pretransform, sample_rate)
+
+    if bottleneck is not None:
+        bottleneck = create_bottleneck_from_config(bottleneck)
+
+    soft_clip = ae_config["decoder"].get("soft_clip", False)
+
+    return AudioAutoencoder(
+        encoder,
+        decoder,
+        io_channels=io_channels,
+        latent_dim=latent_dim,
+        downsampling_ratio=downsampling_ratio,
+        sample_rate=sample_rate,
+        bottleneck=bottleneck,
+        pretransform=pretransform,
+        in_channels=in_channels,
+        out_channels=out_channels,
+        soft_clip=soft_clip
+    )
+
+def create_diffAE_from_config(config: Dict[str, Any]):
+    
+    diffae_config = config["model"]
+
+    if "encoder" in diffae_config:
+        encoder = create_encoder_from_config(diffae_config["encoder"])
+    else:
+        encoder = None
+
+    if "decoder" in diffae_config:
+        decoder = create_decoder_from_config(diffae_config["decoder"])
+    else:
+        decoder = None
+
+    diffusion_model_type = diffae_config["diffusion"]["type"]
+
+    if diffusion_model_type == "DAU1d":
+        diffusion = DAU1DCondWrapper(**diffae_config["diffusion"]["config"])
+    elif diffusion_model_type == "adp_1d":
+        diffusion = UNet1DCondWrapper(**diffae_config["diffusion"]["config"])
+    elif diffusion_model_type == "dit":
+        diffusion = DiTWrapper(**diffae_config["diffusion"]["config"])
+
+    latent_dim = diffae_config.get("latent_dim", None)
+    assert latent_dim is not None, "latent_dim must be specified in model config"
+    downsampling_ratio = diffae_config.get("downsampling_ratio", None)
+    assert downsampling_ratio is not None, "downsampling_ratio must be specified in model config"
+    io_channels = diffae_config.get("io_channels", None)
+    assert io_channels is not None, "io_channels must be specified in model config"
+    sample_rate = config.get("sample_rate", None)
+    assert sample_rate is not None, "sample_rate must be specified in model config"
+
+    bottleneck = diffae_config.get("bottleneck", None)
+
+    pretransform = diffae_config.get("pretransform", None)
+
+    if pretransform is not None:
+        pretransform = create_pretransform_from_config(pretransform, sample_rate)
+
+    if bottleneck is not None:
+        bottleneck = create_bottleneck_from_config(bottleneck)
+
+    diffusion_downsampling_ratio = None,
+
+    if diffusion_model_type == "DAU1d":
+        diffusion_downsampling_ratio = np.prod(diffae_config["diffusion"]["config"]["strides"])
+    elif diffusion_model_type == "adp_1d":
+        diffusion_downsampling_ratio = np.prod(diffae_config["diffusion"]["config"]["factors"])
+    elif diffusion_model_type == "dit":
+        diffusion_downsampling_ratio = 1
+
+    return DiffusionAutoencoder(
+        encoder=encoder,
+        decoder=decoder,
+        diffusion=diffusion,
+        io_channels=io_channels,
+        sample_rate=sample_rate,
+        latent_dim=latent_dim,
+        downsampling_ratio=downsampling_ratio,
+        diffusion_downsampling_ratio=diffusion_downsampling_ratio,
+        bottleneck=bottleneck,
+        pretransform=pretransform
+    )

stable_audio_tools/models/blocks.py

@@ -0,0 +1,339 @@
+from functools import reduce
+import math
+import numpy as np
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from torch.backends.cuda import sdp_kernel
+from packaging import version
+
+from dac.nn.layers import Snake1d
+
+class ResidualBlock(nn.Module):
+    def __init__(self, main, skip=None):
+        super().__init__()
+        self.main = nn.Sequential(*main)
+        self.skip = skip if skip else nn.Identity()
+
+    def forward(self, input):
+        return self.main(input) + self.skip(input)
+
+class ResConvBlock(ResidualBlock):
+    def __init__(self, c_in, c_mid, c_out, is_last=False, kernel_size=5, conv_bias=True, use_snake=False):
+        skip = None if c_in == c_out else nn.Conv1d(c_in, c_out, 1, bias=False)
+        super().__init__([
+            nn.Conv1d(c_in, c_mid, kernel_size, padding=kernel_size//2, bias=conv_bias),
+            nn.GroupNorm(1, c_mid),
+            Snake1d(c_mid) if use_snake else nn.GELU(),
+            nn.Conv1d(c_mid, c_out, kernel_size, padding=kernel_size//2, bias=conv_bias),
+            nn.GroupNorm(1, c_out) if not is_last else nn.Identity(),
+            (Snake1d(c_out) if use_snake else nn.GELU()) if not is_last else nn.Identity(),
+        ], skip)
+
+class SelfAttention1d(nn.Module):
+    def __init__(self, c_in, n_head=1, dropout_rate=0.):
+        super().__init__()
+        assert c_in % n_head == 0
+        self.norm = nn.GroupNorm(1, c_in)
+        self.n_head = n_head
+        self.qkv_proj = nn.Conv1d(c_in, c_in * 3, 1)
+        self.out_proj = nn.Conv1d(c_in, c_in, 1)
+        self.dropout = nn.Dropout(dropout_rate, inplace=True)
+
+        self.use_flash = False
+
+        if not self.use_flash:
+            return
+
+        device_properties = torch.cuda.get_device_properties(torch.device('cuda'))
+
+        if device_properties.major == 8 and device_properties.minor == 0:
+            # Use flash attention for A100 GPUs
+            self.sdp_kernel_config = (False, True, True)
+        else:
+            # Don't use flash attention for other GPUs
+            self.sdp_kernel_config = (False, True, True)
+
+    def forward(self, input):
+        n, c, s = input.shape
+        qkv = self.qkv_proj(self.norm(input))
+        qkv = qkv.view(
+            [n, self.n_head * 3, c // self.n_head, s]).transpose(2, 3)
+        q, k, v = qkv.chunk(3, dim=1)
+        scale = k.shape[3]**-0.25
+
+        if self.use_flash:
+            with sdp_kernel(*self.sdp_kernel_config):
+                y = F.scaled_dot_product_attention(q, k, v, is_causal=False).contiguous().view([n, c, s])
+        else:
+            att = ((q * scale) @ (k.transpose(2, 3) * scale)).softmax(3)
+            y = (att @ v).transpose(2, 3).contiguous().view([n, c, s])
+
+
+        return input + self.dropout(self.out_proj(y))
+
+class SkipBlock(nn.Module):
+    def __init__(self, *main):
+        super().__init__()
+        self.main = nn.Sequential(*main)
+
+    def forward(self, input):
+        return torch.cat([self.main(input), input], dim=1)
+
+class FourierFeatures(nn.Module):
+    def __init__(self, in_features, out_features, std=1.):
+        super().__init__()
+        assert out_features % 2 == 0
+        self.weight = nn.Parameter(torch.randn(
+            [out_features // 2, in_features]) * std)
+
+    def forward(self, input):
+        f = 2 * math.pi * input @ self.weight.T
+        return torch.cat([f.cos(), f.sin()], dim=-1)
+
+def expand_to_planes(input, shape):
+    return input[..., None].repeat([1, 1, shape[2]])
+
+_kernels = {
+    'linear':
+        [1 / 8, 3 / 8, 3 / 8, 1 / 8],
+    'cubic': 
+        [-0.01171875, -0.03515625, 0.11328125, 0.43359375,
+        0.43359375, 0.11328125, -0.03515625, -0.01171875],
+    'lanczos3': 
+        [0.003689131001010537, 0.015056144446134567, -0.03399861603975296,
+        -0.066637322306633, 0.13550527393817902, 0.44638532400131226,
+        0.44638532400131226, 0.13550527393817902, -0.066637322306633,
+        -0.03399861603975296, 0.015056144446134567, 0.003689131001010537]
+}
+
+class Downsample1d(nn.Module):
+    def __init__(self, kernel='linear', pad_mode='reflect', channels_last=False):
+        super().__init__()
+        self.pad_mode = pad_mode
+        kernel_1d = torch.tensor(_kernels[kernel])
+        self.pad = kernel_1d.shape[0] // 2 - 1
+        self.register_buffer('kernel', kernel_1d)
+        self.channels_last = channels_last
+    
+    def forward(self, x):
+        if self.channels_last:
+            x = x.permute(0, 2, 1)
+        x = F.pad(x, (self.pad,) * 2, self.pad_mode)
+        weight = x.new_zeros([x.shape[1], x.shape[1], self.kernel.shape[0]])
+        indices = torch.arange(x.shape[1], device=x.device)
+        weight[indices, indices] = self.kernel.to(weight)
+        x = F.conv1d(x, weight, stride=2)
+        if self.channels_last:
+            x = x.permute(0, 2, 1)
+        return x
+
+
+class Upsample1d(nn.Module):
+    def __init__(self, kernel='linear', pad_mode='reflect', channels_last=False):
+        super().__init__()
+        self.pad_mode = pad_mode
+        kernel_1d = torch.tensor(_kernels[kernel]) * 2
+        self.pad = kernel_1d.shape[0] // 2 - 1
+        self.register_buffer('kernel', kernel_1d)
+        self.channels_last = channels_last
+    
+    def forward(self, x):
+        if self.channels_last:
+            x = x.permute(0, 2, 1)
+        x = F.pad(x, ((self.pad + 1) // 2,) * 2, self.pad_mode)
+        weight = x.new_zeros([x.shape[1], x.shape[1], self.kernel.shape[0]])
+        indices = torch.arange(x.shape[1], device=x.device)
+        weight[indices, indices] = self.kernel.to(weight)
+        x = F.conv_transpose1d(x, weight, stride=2, padding=self.pad * 2 + 1)
+        if self.channels_last:
+            x = x.permute(0, 2, 1)
+        return x
+        
+def Downsample1d_2(
+    in_channels: int, out_channels: int, factor: int, kernel_multiplier: int = 2
+) -> nn.Module:
+    assert kernel_multiplier % 2 == 0, "Kernel multiplier must be even"
+
+    return nn.Conv1d(
+        in_channels=in_channels,
+        out_channels=out_channels,
+        kernel_size=factor * kernel_multiplier + 1,
+        stride=factor,
+        padding=factor * (kernel_multiplier // 2),
+    )
+
+
+def Upsample1d_2(
+    in_channels: int, out_channels: int, factor: int, use_nearest: bool = False
+) -> nn.Module:
+
+    if factor == 1:
+        return nn.Conv1d(
+            in_channels=in_channels, out_channels=out_channels, kernel_size=3, padding=1
+        )
+
+    if use_nearest:
+        return nn.Sequential(
+            nn.Upsample(scale_factor=factor, mode="nearest"),
+            nn.Conv1d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=3,
+                padding=1,
+            ),
+        )
+    else:
+        return nn.ConvTranspose1d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=factor * 2,
+            stride=factor,
+            padding=factor // 2 + factor % 2,
+            output_padding=factor % 2,
+        )
+
+def zero_init(layer):
+    nn.init.zeros_(layer.weight)
+    if layer.bias is not None:
+        nn.init.zeros_(layer.bias)
+    return layer
+
+def rms_norm(x, scale, eps):
+    dtype = reduce(torch.promote_types, (x.dtype, scale.dtype, torch.float32))
+    mean_sq = torch.mean(x.to(dtype)**2, dim=-1, keepdim=True)
+    scale = scale.to(dtype) * torch.rsqrt(mean_sq + eps)
+    return x * scale.to(x.dtype)
+
+rms_norm = torch.compile(rms_norm)
+
+class AdaRMSNorm(nn.Module):
+    def __init__(self, features, cond_features, eps=1e-6):
+        super().__init__()
+        self.eps = eps
+        self.linear = zero_init(nn.Linear(cond_features, features, bias=False))
+  
+    def extra_repr(self):
+        return f"eps={self.eps},"
+
+    def forward(self, x, cond):
+        return rms_norm(x, self.linear(cond)[:, None, :] + 1, self.eps)
+    
+def normalize(x, eps=1e-4):
+    dim = list(range(1, x.ndim))
+    n = torch.linalg.vector_norm(x, dim=dim, keepdim=True)
+    alpha = np.sqrt(n.numel() / x.numel())
+    return x / torch.add(eps, n, alpha=alpha)
+
+class ForcedWNConv1d(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size=1):
+        super().__init__()
+        self.weight = nn.Parameter(torch.randn([out_channels, in_channels, kernel_size]))
+
+    def forward(self, x):
+        if self.training:
+            with torch.no_grad():
+                self.weight.copy_(normalize(self.weight))
+        
+        fan_in = self.weight[0].numel()
+
+        w = normalize(self.weight) / math.sqrt(fan_in)
+
+        return F.conv1d(x, w, padding='same')
+        
+# Kernels
+
+use_compile = True
+
+def compile(function, *args, **kwargs):
+    if not use_compile:
+        return function
+    try:
+        return torch.compile(function, *args, **kwargs)
+    except RuntimeError:
+        return function
+
+
+@compile
+def linear_geglu(x, weight, bias=None):
+    x = x @ weight.mT
+    if bias is not None:
+        x = x + bias
+    x, gate = x.chunk(2, dim=-1)
+    return x * F.gelu(gate)
+
+
+@compile
+def rms_norm(x, scale, eps):
+    dtype = reduce(torch.promote_types, (x.dtype, scale.dtype, torch.float32))
+    mean_sq = torch.mean(x.to(dtype)**2, dim=-1, keepdim=True)
+    scale = scale.to(dtype) * torch.rsqrt(mean_sq + eps)
+    return x * scale.to(x.dtype)
+
+# Layers
+
+class LinearGEGLU(nn.Linear):
+    def __init__(self, in_features, out_features, bias=True):
+        super().__init__(in_features, out_features * 2, bias=bias)
+        self.out_features = out_features
+
+    def forward(self, x):
+        return linear_geglu(x, self.weight, self.bias)
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, shape, fix_scale = False, eps=1e-6):
+        super().__init__()
+        self.eps = eps
+
+        if fix_scale:
+            self.register_buffer("scale", torch.ones(shape))
+        else:
+            self.scale = nn.Parameter(torch.ones(shape))
+
+    def extra_repr(self):
+        return f"shape={tuple(self.scale.shape)}, eps={self.eps}"
+
+    def forward(self, x):
+        return rms_norm(x, self.scale, self.eps)    
+
+def snake_beta(x, alpha, beta):
+    return x + (1.0 / (beta + 0.000000001)) * pow(torch.sin(x * alpha), 2)
+
+try:
+    snake_beta = torch.compile(snake_beta)
+except RuntimeError:
+    pass
+
+# Adapted from https://github.com/NVIDIA/BigVGAN/blob/main/activations.py under MIT license
+# License available in LICENSES/LICENSE_NVIDIA.txt
+class SnakeBeta(nn.Module):
+
+    def __init__(self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=True):
+        super(SnakeBeta, self).__init__()
+        self.in_features = in_features
+
+        # initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale: # log scale alphas initialized to zeros
+            self.alpha = nn.Parameter(torch.zeros(in_features) * alpha)
+            self.beta = nn.Parameter(torch.zeros(in_features) * alpha)
+        else: # linear scale alphas initialized to ones
+            self.alpha = nn.Parameter(torch.ones(in_features) * alpha)
+            self.beta = nn.Parameter(torch.ones(in_features) * alpha)
+
+        self.alpha.requires_grad = alpha_trainable
+        self.beta.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        alpha = self.alpha.unsqueeze(0).unsqueeze(-1) # line up with x to [B, C, T]
+        beta = self.beta.unsqueeze(0).unsqueeze(-1)
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+            beta = torch.exp(beta)
+        x = snake_beta(x, alpha, beta)
+
+        return x

stable_audio_tools/models/bottleneck.py

@@ -0,0 +1,326 @@
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from einops import rearrange
+from vector_quantize_pytorch import ResidualVQ, FSQ
+from dac.nn.quantize import ResidualVectorQuantize as DACResidualVQ
+
+class Bottleneck(nn.Module):
+    def __init__(self, is_discrete: bool = False):
+        super().__init__()
+
+        self.is_discrete = is_discrete
+
+    def encode(self, x, return_info=False, **kwargs):
+        raise NotImplementedError
+
+    def decode(self, x):
+        raise NotImplementedError
+
+class DiscreteBottleneck(Bottleneck):
+    def __init__(self, num_quantizers, codebook_size, tokens_id):
+        super().__init__(is_discrete=True)
+
+        self.num_quantizers = num_quantizers
+        self.codebook_size = codebook_size
+        self.tokens_id = tokens_id
+
+    def decode_tokens(self, codes, **kwargs):
+        raise NotImplementedError
+    
+class TanhBottleneck(Bottleneck):
+    def __init__(self):
+        super().__init__(is_discrete=False)
+        self.tanh = nn.Tanh()
+
+    def encode(self, x, return_info=False):
+        info = {}
+
+        x = torch.tanh(x)
+
+        if return_info:
+            return x, info
+        else:
+            return x
+
+    def decode(self, x):
+        return x
+
+def vae_sample(mean, scale):
+        stdev = nn.functional.softplus(scale) + 1e-4
+        var = stdev * stdev
+        logvar = torch.log(var)
+        latents = torch.randn_like(mean) * stdev + mean
+
+        kl = (mean * mean + var - logvar - 1).sum(1).mean()
+
+        return latents, kl
+
+class VAEBottleneck(Bottleneck):
+    def __init__(self):
+        super().__init__(is_discrete=False)
+
+    def encode(self, x, return_info=False, **kwargs):
+        info = {}
+
+        mean, scale = x.chunk(2, dim=1)
+
+        x, kl = vae_sample(mean, scale)
+
+        info["kl"] = kl
+
+        if return_info:
+            return x, info
+        else:
+            return x
+
+    def decode(self, x):
+        return x
+
+def compute_mean_kernel(x, y):
+        kernel_input = (x[:, None] - y[None]).pow(2).mean(2) / x.shape[-1]
+        return torch.exp(-kernel_input).mean()
+
+def compute_mmd(latents):
+    latents_reshaped = latents.permute(0, 2, 1).reshape(-1, latents.shape[1])
+    noise = torch.randn_like(latents_reshaped)
+
+    latents_kernel = compute_mean_kernel(latents_reshaped, latents_reshaped)
+    noise_kernel = compute_mean_kernel(noise, noise)
+    latents_noise_kernel = compute_mean_kernel(latents_reshaped, noise)
+    
+    mmd = latents_kernel + noise_kernel - 2 * latents_noise_kernel
+    return mmd.mean()
+
+class WassersteinBottleneck(Bottleneck):
+    def __init__(self, noise_augment_dim: int = 0):
+        super().__init__(is_discrete=False)
+
+        self.noise_augment_dim = noise_augment_dim
+    
+    def encode(self, x, return_info=False):
+        info = {}
+
+        if self.training and return_info:
+            mmd = compute_mmd(x)
+            info["mmd"] = mmd
+        
+        if return_info:
+            return x, info
+        
+        return x
+
+    def decode(self, x):
+
+        if self.noise_augment_dim > 0:
+            noise = torch.randn(x.shape[0], self.noise_augment_dim,
+                                x.shape[-1]).type_as(x)
+            x = torch.cat([x, noise], dim=1)
+
+        return x
+
+class L2Bottleneck(Bottleneck):
+    def __init__(self):
+        super().__init__(is_discrete=False)
+    
+    def encode(self, x, return_info=False):
+        info = {}
+
+        x = F.normalize(x, dim=1)
+
+        if return_info:
+            return x, info
+        else:
+            return x
+        
+    def decode(self, x):
+        return F.normalize(x, dim=1)
+        
+class RVQBottleneck(DiscreteBottleneck):
+    def __init__(self, **quantizer_kwargs):
+        super().__init__(num_quantizers = quantizer_kwargs["num_quantizers"], codebook_size = quantizer_kwargs["codebook_size"], tokens_id = "quantizer_indices")
+        self.quantizer = ResidualVQ(**quantizer_kwargs)
+        self.num_quantizers = quantizer_kwargs["num_quantizers"]
+
+    def encode(self, x, return_info=False, **kwargs):
+        info = {}
+
+        x = rearrange(x, "b c n -> b n c")
+        x, indices, loss = self.quantizer(x)
+        x = rearrange(x, "b n c -> b c n")
+
+        info["quantizer_indices"] = indices
+        info["quantizer_loss"] = loss.mean()
+
+        if return_info:
+            return x, info
+        else:
+            return x
+        
+    def decode(self, x):
+        return x
+    
+    def decode_tokens(self, codes, **kwargs):
+        latents = self.quantizer.get_outputs_from_indices(codes)
+
+        return self.decode(latents, **kwargs)
+    
+class RVQVAEBottleneck(DiscreteBottleneck):
+    def __init__(self, **quantizer_kwargs):
+        super().__init__(num_quantizers = quantizer_kwargs["num_quantizers"], codebook_size = quantizer_kwargs["codebook_size"], tokens_id = "quantizer_indices")
+        self.quantizer = ResidualVQ(**quantizer_kwargs)
+        self.num_quantizers = quantizer_kwargs["num_quantizers"]
+
+    def encode(self, x, return_info=False):
+        info = {}
+
+        x, kl = vae_sample(*x.chunk(2, dim=1))
+
+        info["kl"] = kl
+
+        x = rearrange(x, "b c n -> b n c")
+        x, indices, loss = self.quantizer(x)
+        x = rearrange(x, "b n c -> b c n")
+
+        info["quantizer_indices"] = indices
+        info["quantizer_loss"] = loss.mean()
+
+        if return_info:
+            return x, info
+        else:
+            return x
+        
+    def decode(self, x):
+        return x
+    
+    def decode_tokens(self, codes, **kwargs):
+        latents = self.quantizer.get_outputs_from_indices(codes)
+
+        return self.decode(latents, **kwargs)
+
+class DACRVQBottleneck(DiscreteBottleneck):
+    def __init__(self, quantize_on_decode=False, **quantizer_kwargs):
+        super().__init__(num_quantizers = quantizer_kwargs["n_codebooks"], codebook_size = quantizer_kwargs["codebook_size"], tokens_id = "codes")
+        self.quantizer = DACResidualVQ(**quantizer_kwargs)
+        self.num_quantizers = quantizer_kwargs["n_codebooks"]
+        self.quantize_on_decode = quantize_on_decode
+
+    def encode(self, x, return_info=False, **kwargs):
+        info = {}
+
+        info["pre_quantizer"] = x
+
+        if self.quantize_on_decode:
+            return x, info if return_info else x
+
+        z, codes, latents, commitment_loss, codebook_loss = self.quantizer(x, **kwargs)
+
+        output = {
+            "z": z,
+            "codes": codes,
+            "latents": latents,
+            "vq/commitment_loss": commitment_loss,
+            "vq/codebook_loss": codebook_loss,
+        }
+
+        output["vq/commitment_loss"] /= self.num_quantizers
+        output["vq/codebook_loss"] /= self.num_quantizers
+
+        info.update(output)
+
+        if return_info:
+            return output["z"], info
+        
+        return output["z"]
+    
+    def decode(self, x):
+
+        if self.quantize_on_decode:
+            x = self.quantizer(x)[0]
+
+        return x
+    
+    def decode_tokens(self, codes, **kwargs):
+        latents, _, _ = self.quantizer.from_codes(codes)
+
+        return self.decode(latents, **kwargs)
+
+class DACRVQVAEBottleneck(DiscreteBottleneck):
+    def __init__(self, quantize_on_decode=False, **quantizer_kwargs):
+        super().__init__(num_quantizers = quantizer_kwargs["n_codebooks"], codebook_size = quantizer_kwargs["codebook_size"], tokens_id = "codes")
+        self.quantizer = DACResidualVQ(**quantizer_kwargs)
+        self.num_quantizers = quantizer_kwargs["n_codebooks"]
+        self.quantize_on_decode = quantize_on_decode
+
+    def encode(self, x, return_info=False, n_quantizers: int = None):
+        info = {}
+
+        mean, scale = x.chunk(2, dim=1)
+
+        x, kl = vae_sample(mean, scale)
+
+        info["pre_quantizer"] = x
+        info["kl"] = kl
+
+        if self.quantize_on_decode:
+            return x, info if return_info else x
+
+        z, codes, latents, commitment_loss, codebook_loss = self.quantizer(x, n_quantizers=n_quantizers)
+
+        output = {
+            "z": z,
+            "codes": codes,
+            "latents": latents,
+            "vq/commitment_loss": commitment_loss,
+            "vq/codebook_loss": codebook_loss,
+        }
+
+        output["vq/commitment_loss"] /= self.num_quantizers
+        output["vq/codebook_loss"] /= self.num_quantizers
+
+        info.update(output)
+
+        if return_info:
+            return output["z"], info
+        
+        return output["z"]
+    
+    def decode(self, x):
+
+        if self.quantize_on_decode:
+            x = self.quantizer(x)[0]
+
+        return x
+
+    def decode_tokens(self, codes, **kwargs):
+        latents, _, _ = self.quantizer.from_codes(codes)
+
+        return self.decode(latents, **kwargs)
+    
+class FSQBottleneck(DiscreteBottleneck):
+    def __init__(self, dim, levels):
+        super().__init__(num_quantizers = 1, codebook_size = levels ** dim, tokens_id = "quantizer_indices")
+        self.quantizer = FSQ(levels=[levels] * dim)
+
+    def encode(self, x, return_info=False):
+        info = {}
+
+        x = rearrange(x, "b c n -> b n c")
+        x, indices = self.quantizer(x)
+        x = rearrange(x, "b n c -> b c n")
+
+        info["quantizer_indices"] = indices
+
+        if return_info:
+            return x, info
+        else:
+            return x
+        
+    def decode(self, x):
+        return x
+    
+    def decode_tokens(self, tokens, **kwargs):
+        latents = self.quantizer.indices_to_codes(tokens)
+
+        return self.decode(latents, **kwargs)

stable_audio_tools/models/conditioners.py

@@ -0,0 +1,558 @@
+#Heavily influenced by https://github.com/facebookresearch/audiocraft/blob/main/audiocraft/modules/conditioners.py
+
+import torch
+import logging, warnings
+import string
+import typing as tp
+import gc
+
+from .adp import NumberEmbedder
+from ..inference.utils import set_audio_channels
+from .factory import create_pretransform_from_config
+from .pretransforms import Pretransform
+from ..training.utils import copy_state_dict
+from .utils import load_ckpt_state_dict
+
+from torch import nn
+
+class Conditioner(nn.Module):
+    def __init__(
+            self,
+            dim: int,
+            output_dim: int,
+            project_out: bool = False,
+            ):
+        
+        super().__init__()
+
+        self.dim = dim
+        self.output_dim = output_dim
+        self.proj_out = nn.Linear(dim, output_dim) if (dim != output_dim or project_out) else nn.Identity()
+
+    def forward(self, x: tp.Any) -> tp.Any:
+        raise NotImplementedError()
+    
+class IntConditioner(Conditioner):
+    def __init__(self, 
+                output_dim: int,
+                min_val: int=0,
+                max_val: int=512
+                ):
+        super().__init__(output_dim, output_dim)
+
+        self.min_val = min_val
+        self.max_val = max_val
+        self.int_embedder = nn.Embedding(max_val - min_val + 1, output_dim).requires_grad_(True)
+
+    def forward(self, ints: tp.List[int], device=None) -> tp.Any:
+            
+            #self.int_embedder.to(device)
+    
+            ints = torch.tensor(ints).to(device)
+            ints = ints.clamp(self.min_val, self.max_val)
+    
+            int_embeds = self.int_embedder(ints).unsqueeze(1)
+    
+            return [int_embeds, torch.ones(int_embeds.shape[0], 1).to(device)]
+
+class NumberConditioner(Conditioner):
+    '''
+        Conditioner that takes a list of floats, normalizes them for a given range, and returns a list of embeddings
+    '''
+    def __init__(self, 
+                output_dim: int,
+                min_val: float=0,
+                max_val: float=1
+                ):
+        super().__init__(output_dim, output_dim)
+
+        self.min_val = min_val
+        self.max_val = max_val
+
+        self.embedder = NumberEmbedder(features=output_dim)
+
+    def forward(self, floats: tp.List[float], device=None) -> tp.Any:
+    
+            # Cast the inputs to floats
+            floats = [float(x) for x in floats]
+
+            floats = torch.tensor(floats).to(device)
+
+            floats = floats.clamp(self.min_val, self.max_val)
+    
+            normalized_floats = (floats - self.min_val) / (self.max_val - self.min_val)
+
+            # Cast floats to same type as embedder
+            embedder_dtype = next(self.embedder.parameters()).dtype
+            normalized_floats = normalized_floats.to(embedder_dtype)
+
+            float_embeds = self.embedder(normalized_floats).unsqueeze(1)
+    
+            return [float_embeds, torch.ones(float_embeds.shape[0], 1).to(device)]
+
+class CLAPTextConditioner(Conditioner):
+    def __init__(self, 
+                 output_dim: int, 
+                 clap_ckpt_path,
+                 use_text_features = False,
+                 feature_layer_ix: int = -1,
+                 audio_model_type="HTSAT-base", 
+                 enable_fusion=True,
+                 project_out: bool = False,
+                 finetune: bool = False):
+        super().__init__(768 if use_text_features else 512, output_dim, project_out=project_out)
+
+        self.use_text_features = use_text_features
+        self.feature_layer_ix = feature_layer_ix
+        self.finetune = finetune
+
+        # Suppress logging from transformers
+        previous_level = logging.root.manager.disable
+        logging.disable(logging.ERROR)
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            try:
+                import laion_clap
+                from laion_clap.clap_module.factory import load_state_dict as clap_load_state_dict
+                
+                model = laion_clap.CLAP_Module(enable_fusion=enable_fusion, amodel=audio_model_type, device='cpu')
+
+                if self.finetune:
+                    self.model = model
+                else: 
+                    self.__dict__["model"] = model
+
+                state_dict = clap_load_state_dict(clap_ckpt_path)
+                self.model.model.load_state_dict(state_dict, strict=False)
+
+                if self.finetune:
+                    self.model.model.text_branch.requires_grad_(True)
+                    self.model.model.text_branch.train()
+                else:
+                    self.model.model.text_branch.requires_grad_(False)
+                    self.model.model.text_branch.eval()
+
+            finally:
+                logging.disable(previous_level)
+
+        del self.model.model.audio_branch
+
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def get_clap_features(self, prompts, layer_ix=-2, device: tp.Any = "cuda"):
+        prompt_tokens = self.model.tokenizer(prompts)
+        attention_mask = prompt_tokens["attention_mask"].to(device=device, non_blocking=True)
+        prompt_features = self.model.model.text_branch(
+            input_ids=prompt_tokens["input_ids"].to(device=device, non_blocking=True),
+            attention_mask=attention_mask,
+            output_hidden_states=True
+        )["hidden_states"][layer_ix]
+
+        return prompt_features, attention_mask
+
+    def forward(self, texts: tp.List[str], device: tp.Any = "cuda") -> tp.Any:
+        self.model.to(device)
+
+        if self.use_text_features:
+            if len(texts) == 1:
+                text_features, text_attention_mask = self.get_clap_features([texts[0], ""], layer_ix=self.feature_layer_ix, device=device)
+                text_features = text_features[:1, ...]
+                text_attention_mask = text_attention_mask[:1, ...]
+            else:
+                text_features, text_attention_mask = self.get_clap_features(texts, layer_ix=self.feature_layer_ix, device=device)
+            return [self.proj_out(text_features), text_attention_mask]
+
+        # Fix for CLAP bug when only one text is passed
+        if len(texts) == 1:
+            text_embedding = self.model.get_text_embedding([texts[0], ""], use_tensor=True)[:1, ...]
+        else:
+            text_embedding = self.model.get_text_embedding(texts, use_tensor=True)
+
+        text_embedding = text_embedding.unsqueeze(1).to(device)
+
+        return [self.proj_out(text_embedding), torch.ones(text_embedding.shape[0], 1).to(device)]
+
+class CLAPAudioConditioner(Conditioner):
+    def __init__(self, 
+                 output_dim: int, 
+                 clap_ckpt_path,
+                 audio_model_type="HTSAT-base", 
+                 enable_fusion=True,
+                 project_out: bool = False):
+        super().__init__(512, output_dim, project_out=project_out)
+
+        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+        # Suppress logging from transformers
+        previous_level = logging.root.manager.disable
+        logging.disable(logging.ERROR)
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            try:
+                import laion_clap
+                from laion_clap.clap_module.factory import load_state_dict as clap_load_state_dict
+                
+                model = laion_clap.CLAP_Module(enable_fusion=enable_fusion, amodel=audio_model_type, device='cpu')
+
+                if self.finetune:
+                    self.model = model
+                else: 
+                    self.__dict__["model"] = model
+
+                state_dict = clap_load_state_dict(clap_ckpt_path)
+                self.model.model.load_state_dict(state_dict, strict=False)
+
+                if self.finetune:
+                    self.model.model.audio_branch.requires_grad_(True)
+                    self.model.model.audio_branch.train()
+                else:
+                    self.model.model.audio_branch.requires_grad_(False)
+                    self.model.model.audio_branch.eval()
+
+            finally:
+                logging.disable(previous_level)
+
+        del self.model.model.text_branch
+
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def forward(self, audios: tp.Union[torch.Tensor, tp.List[torch.Tensor], tp.Tuple[torch.Tensor]] , device: tp.Any = "cuda") -> tp.Any:
+
+        self.model.to(device)
+
+        if isinstance(audios, list) or isinstance(audios, tuple):
+            audios = torch.cat(audios, dim=0)
+
+        # Convert to mono
+        mono_audios = audios.mean(dim=1)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            audio_embedding = self.model.get_audio_embedding_from_data(mono_audios.float(), use_tensor=True)
+
+        audio_embedding = audio_embedding.unsqueeze(1).to(device)
+
+        return [self.proj_out(audio_embedding), torch.ones(audio_embedding.shape[0], 1).to(device)]
+
+class T5Conditioner(Conditioner):
+
+    T5_MODELS = ["t5-small", "t5-base", "t5-large", "t5-3b", "t5-11b",
+              "google/flan-t5-small", "google/flan-t5-base", "google/flan-t5-large",
+              "google/flan-t5-xl", "google/flan-t5-xxl"]
+    
+    T5_MODEL_DIMS = {
+        "t5-small": 512,
+        "t5-base": 768,
+        "t5-large": 1024,
+        "t5-3b": 1024,
+        "t5-11b": 1024,
+        "t5-xl": 2048,
+        "t5-xxl": 4096,
+        "google/flan-t5-small": 512,
+        "google/flan-t5-base": 768,
+        "google/flan-t5-large": 1024,
+        "google/flan-t5-3b": 1024,
+        "google/flan-t5-11b": 1024,
+        "google/flan-t5-xl": 2048,
+        "google/flan-t5-xxl": 4096,
+    }
+
+    def __init__(
+            self,
+            output_dim: int,
+            t5_model_name: str = "t5-base",
+            max_length: str = 128,
+            enable_grad: bool = False,
+            project_out: bool = False,
+    ):
+        assert t5_model_name in self.T5_MODELS, f"Unknown T5 model name: {t5_model_name}"
+        super().__init__(self.T5_MODEL_DIMS[t5_model_name], output_dim, project_out=project_out)
+        
+        from transformers import T5EncoderModel, AutoTokenizer
+
+        self.max_length = max_length
+        self.enable_grad = enable_grad
+
+        # Suppress logging from transformers
+        previous_level = logging.root.manager.disable
+        logging.disable(logging.ERROR)
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            try:
+                # self.tokenizer = T5Tokenizer.from_pretrained(t5_model_name, model_max_length = max_length)
+                # model = T5EncoderModel.from_pretrained(t5_model_name, max_length=max_length).train(enable_grad).requires_grad_(enable_grad)
+                self.tokenizer = AutoTokenizer.from_pretrained(t5_model_name)
+                model = T5EncoderModel.from_pretrained(t5_model_name).train(enable_grad).requires_grad_(enable_grad).to(torch.float16)
+            finally:
+                logging.disable(previous_level)
+            
+        if self.enable_grad:
+            self.model = model
+        else: 
+            self.__dict__["model"] = model
+
+
+    def forward(self, texts: tp.List[str], device: tp.Union[torch.device, str]) -> tp.Tuple[torch.Tensor, torch.Tensor]:
+        
+        self.model.to(device)
+        self.proj_out.to(device)
+
+        encoded = self.tokenizer(
+            texts,
+            truncation=True,
+            max_length=self.max_length,
+            padding="max_length",
+            return_tensors="pt",
+        )
+
+        input_ids = encoded["input_ids"].to(device)
+        attention_mask = encoded["attention_mask"].to(device).to(torch.bool)
+
+        self.model.eval()
+            
+        with torch.cuda.amp.autocast(dtype=torch.float16) and torch.set_grad_enabled(self.enable_grad):
+            embeddings = self.model(
+                input_ids=input_ids, attention_mask=attention_mask
+            )["last_hidden_state"]    
+            
+        embeddings = self.proj_out(embeddings.float())
+
+        embeddings = embeddings * attention_mask.unsqueeze(-1).float()
+
+        return embeddings, attention_mask
+    
+class PhonemeConditioner(Conditioner):
+    """
+    A conditioner that turns text into phonemes and embeds them using a lookup table
+    Only works for English text
+
+    Args:
+        output_dim: the dimension of the output embeddings
+        max_length: the maximum number of phonemes to embed
+        project_out: whether to add another linear projection to the output embeddings
+    """
+
+    def __init__(
+            self,
+            output_dim: int,
+            max_length: int = 1024,
+            project_out: bool = False,
+    ):
+        super().__init__(output_dim, output_dim, project_out=project_out)
+        
+        from g2p_en import G2p
+
+        self.max_length = max_length
+
+        self.g2p = G2p()
+
+        # Reserving 0 for padding, 1 for ignored
+        self.phoneme_embedder = nn.Embedding(len(self.g2p.phonemes) + 2, output_dim)
+
+    def forward(self, texts: tp.List[str], device: tp.Union[torch.device, str]) -> tp.Tuple[torch.Tensor, torch.Tensor]:
+        
+        self.phoneme_embedder.to(device)
+        self.proj_out.to(device)
+
+        batch_phonemes = [self.g2p(text) for text in texts] # shape [batch_size, length]
+        
+        phoneme_ignore = [" ", *string.punctuation]
+
+        # Remove ignored phonemes and cut to max length
+        batch_phonemes = [[p if p not in phoneme_ignore else "_" for p in phonemes] for phonemes in batch_phonemes]
+
+        # Convert to ids
+        phoneme_ids = [[self.g2p.p2idx[p] + 2 if p in self.g2p.p2idx else 1 for p in phonemes] for phonemes in batch_phonemes]
+
+        #Pad to match longest and make a mask tensor for the padding
+        longest = max([len(ids) for ids in phoneme_ids])
+        phoneme_ids = [ids + [0] * (longest - len(ids)) for ids in phoneme_ids]
+        
+        phoneme_ids = torch.tensor(phoneme_ids).to(device)
+
+        # Convert to embeddings
+        phoneme_embeds = self.phoneme_embedder(phoneme_ids)
+        
+        phoneme_embeds = self.proj_out(phoneme_embeds)
+
+        return phoneme_embeds, torch.ones(phoneme_embeds.shape[0], phoneme_embeds.shape[1]).to(device)
+  
+class TokenizerLUTConditioner(Conditioner):
+    """
+    A conditioner that embeds text using a lookup table on a pretrained tokenizer's vocabulary
+
+    Args:
+        tokenizer_name: the name of the tokenizer from the Hugging Face transformers library
+        output_dim: the dimension of the output embeddings
+        max_length: the maximum length of the text to embed
+        project_out: whether to add another linear projection to the output embeddings
+    """
+
+    def __init__(
+            self,
+            tokenizer_name: str, # Name of a tokenizer from the Hugging Face transformers library
+            output_dim: int,
+            max_length: int = 1024,
+            project_out: bool = False,
+    ):
+        super().__init__(output_dim, output_dim, project_out=project_out)
+        
+        from transformers import AutoTokenizer
+
+         # Suppress logging from transformers
+        previous_level = logging.root.manager.disable
+        logging.disable(logging.ERROR)
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore")
+            try:
+                self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
+            finally:
+                logging.disable(previous_level)
+
+        self.max_length = max_length
+
+        self.token_embedder = nn.Embedding(len(self.tokenizer), output_dim)
+
+    def forward(self, texts: tp.List[str], device: tp.Union[torch.device, str]) -> tp.Tuple[torch.Tensor, torch.Tensor]:
+        self.proj_out.to(device)
+
+        encoded = self.tokenizer(
+            texts,
+            truncation=True,
+            max_length=self.max_length,
+            padding="max_length",
+            return_tensors="pt",
+        )
+
+        input_ids = encoded["input_ids"].to(device)
+        attention_mask = encoded["attention_mask"].to(device).to(torch.bool)
+    
+        embeddings = self.token_embedder(input_ids)
+            
+        embeddings = self.proj_out(embeddings)
+
+        embeddings = embeddings * attention_mask.unsqueeze(-1).float()
+
+        return embeddings, attention_mask
+
+class PretransformConditioner(Conditioner):
+    """
+    A conditioner that uses a pretransform's encoder for conditioning
+
+    Args:
+        pretransform: an instantiated pretransform to use for conditioning
+        output_dim: the dimension of the output embeddings
+    """
+    def __init__(self, pretransform: Pretransform, output_dim: int):
+        super().__init__(pretransform.encoded_channels, output_dim)
+
+        self.pretransform = pretransform
+
+    def forward(self, audio: tp.Union[torch.Tensor, tp.List[torch.Tensor], tp.Tuple[torch.Tensor]], device: tp.Union[torch.device, str]) -> tp.Tuple[torch.Tensor, torch.Tensor]:
+
+        self.pretransform.to(device)
+        self.proj_out.to(device)
+
+        if isinstance(audio, list) or isinstance(audio, tuple):
+            audio = torch.cat(audio, dim=0)
+
+        # Convert audio to pretransform input channels
+        audio = set_audio_channels(audio, self.pretransform.io_channels)
+        
+        latents = self.pretransform.encode(audio)
+
+        latents = self.proj_out(latents)
+
+        return [latents, torch.ones(latents.shape[0], latents.shape[2]).to(latents.device)]
+
+class MultiConditioner(nn.Module):
+    """
+    A module that applies multiple conditioners to an input dictionary based on the keys
+
+    Args:
+        conditioners: a dictionary of conditioners with keys corresponding to the keys of the conditioning input dictionary (e.g. "prompt")
+        default_keys: a dictionary of default keys to use if the key is not in the input dictionary (e.g. {"prompt_t5": "prompt"})
+    """
+    def __init__(self, conditioners: tp.Dict[str, Conditioner], default_keys: tp.Dict[str, str] = {}):
+        super().__init__()
+
+        self.conditioners = nn.ModuleDict(conditioners)
+        self.default_keys = default_keys
+
+    def forward(self, batch_metadata: tp.List[tp.Dict[str, tp.Any]], device: tp.Union[torch.device, str]) -> tp.Dict[str, tp.Any]:
+        output = {}
+
+        for key, conditioner in self.conditioners.items():
+            condition_key = key
+
+            conditioner_inputs = []
+
+            for x in batch_metadata:
+
+                if condition_key not in x:
+                    if condition_key in self.default_keys:
+                        condition_key = self.default_keys[condition_key]
+                    else:
+                        raise ValueError(f"Conditioner key {condition_key} not found in batch metadata")
+
+                #Unwrap the condition info if it's a single-element list or tuple, this is to support collation functions that wrap everything in a list
+                if isinstance(x[condition_key], list) or isinstance(x[condition_key], tuple) and len(x[condition_key]) == 1:
+                    conditioner_inputs.append(x[condition_key][0])
+                else:
+                    conditioner_inputs.append(x[condition_key])
+            
+            output[key] = conditioner(conditioner_inputs, device)
+
+        return output
+    
+def create_multi_conditioner_from_conditioning_config(config: tp.Dict[str, tp.Any]) -> MultiConditioner:
+    """
+    Create a MultiConditioner from a conditioning config dictionary
+
+    Args:
+        config: the conditioning config dictionary
+        device: the device to put the conditioners on
+    """
+    conditioners = {}
+    cond_dim = config["cond_dim"]
+    
+    default_keys = config.get("default_keys", {})
+
+    for conditioner_info in config["configs"]:
+        id = conditioner_info["id"]
+
+        conditioner_type = conditioner_info["type"]
+
+        conditioner_config = {"output_dim": cond_dim}
+        
+        conditioner_config.update(conditioner_info["config"])
+
+        if conditioner_type == "t5":
+            conditioners[id] = T5Conditioner(**conditioner_config)
+        elif conditioner_type == "clap_text":
+            conditioners[id] = CLAPTextConditioner(**conditioner_config)
+        elif conditioner_type == "clap_audio":
+            conditioners[id] = CLAPAudioConditioner(**conditioner_config)
+        elif conditioner_type == "int":
+            conditioners[id] = IntConditioner(**conditioner_config)
+        elif conditioner_type == "number":
+            conditioners[id] = NumberConditioner(**conditioner_config)
+        elif conditioner_type == "phoneme":
+            conditioners[id] = PhonemeConditioner(**conditioner_config)
+        elif conditioner_type == "lut":
+            conditioners[id] = TokenizerLUTConditioner(**conditioner_config)
+        elif conditioner_type == "pretransform":
+            sample_rate = conditioner_config.pop("sample_rate", None)
+            assert sample_rate is not None, "Sample rate must be specified for pretransform conditioners"
+
+            pretransform = create_pretransform_from_config(conditioner_config.pop("pretransform_config"), sample_rate=sample_rate)
+
+            if conditioner_config.get("pretransform_ckpt_path", None) is not None:
+                pretransform.load_state_dict(load_ckpt_state_dict(conditioner_config.pop("pretransform_ckpt_path")))
+
+            conditioners[id] = PretransformConditioner(pretransform, **conditioner_config)
+        else:
+            raise ValueError(f"Unknown conditioner type: {conditioner_type}")
+
+    return MultiConditioner(conditioners, default_keys=default_keys)

stable_audio_tools/models/diffusion.py

@@ -0,0 +1,678 @@
+import torch
+from torch import nn
+from torch.nn import functional as F
+from functools import partial, reduce
+import numpy as np
+import typing as tp
+
+from .blocks import ResConvBlock, FourierFeatures, Upsample1d, Upsample1d_2, Downsample1d, Downsample1d_2, SelfAttention1d, SkipBlock, expand_to_planes
+from .conditioners import MultiConditioner, create_multi_conditioner_from_conditioning_config
+from .dit import DiffusionTransformer
+from .factory import create_pretransform_from_config
+from .pretransforms import Pretransform
+from ..inference.generation import generate_diffusion_cond
+
+from .adp import UNetCFG1d, UNet1d
+
+from time import time
+
+class Profiler:
+
+    def __init__(self):
+        self.ticks = [[time(), None]]
+
+    def tick(self, msg):
+        self.ticks.append([time(), msg])
+
+    def __repr__(self):
+        rep = 80 * "=" + "\n"
+        for i in range(1, len(self.ticks)):
+            msg = self.ticks[i][1]
+            ellapsed = self.ticks[i][0] - self.ticks[i - 1][0]
+            rep += msg + f": {ellapsed*1000:.2f}ms\n"
+        rep += 80 * "=" + "\n\n\n"
+        return rep
+
+class DiffusionModel(nn.Module):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def forward(self, x, t, **kwargs):
+        raise NotImplementedError()
+
+class DiffusionModelWrapper(nn.Module):
+    def __init__(
+                self,
+                model: DiffusionModel,
+                io_channels,
+                sample_size,
+                sample_rate,
+                min_input_length,
+                pretransform: tp.Optional[Pretransform] = None,
+    ):
+        super().__init__()
+        self.io_channels = io_channels
+        self.sample_size = sample_size
+        self.sample_rate = sample_rate
+        self.min_input_length = min_input_length
+
+        self.model = model
+
+        if pretransform is not None:
+            self.pretransform = pretransform
+        else:
+            self.pretransform = None
+
+    def forward(self, x, t, **kwargs):
+        return self.model(x, t, **kwargs)
+
+class ConditionedDiffusionModel(nn.Module):
+    def __init__(self,
+                *args,
+                supports_cross_attention: bool = False,
+                supports_input_concat: bool = False,
+                supports_global_cond: bool = False,
+                supports_prepend_cond: bool = False,
+                **kwargs):
+        super().__init__(*args, **kwargs)
+        self.supports_cross_attention = supports_cross_attention
+        self.supports_input_concat = supports_input_concat
+        self.supports_global_cond = supports_global_cond
+        self.supports_prepend_cond = supports_prepend_cond
+
+    def forward(self,
+                x: torch.Tensor,
+                t: torch.Tensor,
+                cross_attn_cond: torch.Tensor = None,
+                cross_attn_mask: torch.Tensor = None,
+                input_concat_cond: torch.Tensor = None,
+                global_embed: torch.Tensor = None,
+                prepend_cond: torch.Tensor = None,
+                prepend_cond_mask: torch.Tensor = None,
+                cfg_scale: float = 1.0,
+                cfg_dropout_prob: float = 0.0,
+                batch_cfg: bool = False,
+                rescale_cfg: bool = False,
+                **kwargs):
+        raise NotImplementedError()
+
+class ConditionedDiffusionModelWrapper(nn.Module):
+    """
+    A diffusion model that takes in conditioning
+    """
+    def __init__(
+            self,
+            model: ConditionedDiffusionModel,
+            conditioner: MultiConditioner,
+            io_channels,
+            sample_rate,
+            min_input_length: int,
+            pretransform: tp.Optional[Pretransform] = None,
+            cross_attn_cond_ids: tp.List[str] = [],
+            global_cond_ids: tp.List[str] = [],
+            input_concat_ids: tp.List[str] = [],
+            prepend_cond_ids: tp.List[str] = [],
+            ):
+        super().__init__()
+
+        self.model = model
+        self.conditioner = conditioner
+        self.io_channels = io_channels
+        self.sample_rate = sample_rate
+        self.pretransform = pretransform
+        self.cross_attn_cond_ids = cross_attn_cond_ids
+        self.global_cond_ids = global_cond_ids
+        self.input_concat_ids = input_concat_ids
+        self.prepend_cond_ids = prepend_cond_ids
+        self.min_input_length = min_input_length
+
+    def get_conditioning_inputs(self, cond: tp.Dict[str, tp.Any], negative=False):
+        cross_attention_input = None
+        cross_attention_masks = None
+        global_cond = None
+        input_concat_cond = None
+        prepend_cond = None
+        prepend_cond_mask = None
+
+        if len(self.cross_attn_cond_ids) > 0:
+            # Concatenate all cross-attention inputs over the sequence dimension
+            # Assumes that the cross-attention inputs are of shape (batch, seq, channels)
+            cross_attention_input = []
+            cross_attention_masks = []
+
+            for key in self.cross_attn_cond_ids:
+                cross_attn_in, cross_attn_mask = cond[key]
+
+                # Add sequence dimension if it's not there
+                if len(cross_attn_in.shape) == 2:
+                    cross_attn_in = cross_attn_in.unsqueeze(1)
+                    cross_attn_mask = cross_attn_mask.unsqueeze(1)
+
+                cross_attention_input.append(cross_attn_in)
+                cross_attention_masks.append(cross_attn_mask)
+
+            cross_attention_input = torch.cat(cross_attention_input, dim=1)
+            cross_attention_masks = torch.cat(cross_attention_masks, dim=1)
+
+        if len(self.global_cond_ids) > 0:
+            # Concatenate all global conditioning inputs over the channel dimension
+            # Assumes that the global conditioning inputs are of shape (batch, channels)
+            global_cond = torch.cat([cond[key][0] for key in self.global_cond_ids], dim=-1)
+            if len(global_cond.shape) == 3:
+                global_cond = global_cond.squeeze(1)
+
+        if len(self.input_concat_ids) > 0:
+            # Concatenate all input concat conditioning inputs over the channel dimension
+            # Assumes that the input concat conditioning inputs are of shape (batch, channels, seq)
+            input_concat_cond = torch.cat([cond[key][0] for key in self.input_concat_ids], dim=1)
+
+        if len(self.prepend_cond_ids) > 0:
+            # Concatenate all prepend conditioning inputs over the sequence dimension
+            # Assumes that the prepend conditioning inputs are of shape (batch, seq, channels)
+            prepend_cond = torch.cat([cond[key][0] for key in self.prepend_cond_ids], dim=1)
+            prepend_cond_mask = torch.cat([cond[key][1] for key in self.prepend_cond_ids], dim=1)
+
+        if negative:
+            return {
+                "negative_cross_attn_cond": cross_attention_input,
+                "negative_cross_attn_mask": cross_attention_masks,
+                "negative_global_cond": global_cond,
+                "negative_input_concat_cond": input_concat_cond
+            }
+        else:
+            return {
+                "cross_attn_cond": cross_attention_input,
+                "cross_attn_mask": cross_attention_masks,
+                "global_cond": global_cond,
+                "input_concat_cond": input_concat_cond,
+                "prepend_cond": prepend_cond,
+                "prepend_cond_mask": prepend_cond_mask
+            }
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor, cond: tp.Dict[str, tp.Any], **kwargs):
+        return self.model(x, t, **self.get_conditioning_inputs(cond), **kwargs)
+
+    def generate(self, *args, **kwargs):
+        return generate_diffusion_cond(self, *args, **kwargs)
+
+class UNetCFG1DWrapper(ConditionedDiffusionModel):
+    def __init__(
+        self,
+        *args,
+        **kwargs
+    ):
+        super().__init__(supports_cross_attention=True, supports_global_cond=True, supports_input_concat=True)
+
+        self.model = UNetCFG1d(*args, **kwargs)
+
+        with torch.no_grad():
+            for param in self.model.parameters():
+                param *= 0.5
+
+    def forward(self,
+                x,
+                t,
+                cross_attn_cond=None,
+                cross_attn_mask=None,
+                input_concat_cond=None,
+                global_cond=None,
+                cfg_scale=1.0,
+                cfg_dropout_prob: float = 0.0,
+                batch_cfg: bool = False,
+                rescale_cfg: bool = False,
+                negative_cross_attn_cond=None,
+                negative_cross_attn_mask=None,
+                negative_global_cond=None,
+                negative_input_concat_cond=None,
+                prepend_cond=None,
+                prepend_cond_mask=None,
+                **kwargs):
+        p = Profiler()
+
+        p.tick("start")
+
+        channels_list = None
+        if input_concat_cond is not None:
+            channels_list = [input_concat_cond]
+
+        outputs = self.model(
+            x,
+            t,
+            embedding=cross_attn_cond,
+            embedding_mask=cross_attn_mask,
+            features=global_cond,
+            channels_list=channels_list,
+            embedding_scale=cfg_scale,
+            embedding_mask_proba=cfg_dropout_prob,
+            batch_cfg=batch_cfg,
+            rescale_cfg=rescale_cfg,
+            negative_embedding=negative_cross_attn_cond,
+            negative_embedding_mask=negative_cross_attn_mask,
+            **kwargs)
+
+        p.tick("UNetCFG1D forward")
+
+        #print(f"Profiler: {p}")
+        return outputs
+
+class UNet1DCondWrapper(ConditionedDiffusionModel):
+    def __init__(
+        self,
+        *args,
+        **kwargs
+    ):
+        super().__init__(supports_cross_attention=False, supports_global_cond=True, supports_input_concat=True)
+
+        self.model = UNet1d(*args, **kwargs)
+
+        with torch.no_grad():
+            for param in self.model.parameters():
+                param *= 0.5
+
+    def forward(self,
+                x,
+                t,
+                input_concat_cond=None,
+                global_cond=None,
+                cross_attn_cond=None,
+                cross_attn_mask=None,
+                prepend_cond=None,
+                prepend_cond_mask=None,
+                cfg_scale=1.0,
+                cfg_dropout_prob: float = 0.0,
+                batch_cfg: bool = False,
+                rescale_cfg: bool = False,
+                negative_cross_attn_cond=None,
+                negative_cross_attn_mask=None,
+                negative_global_cond=None,
+                negative_input_concat_cond=None,
+                **kwargs):
+
+        channels_list = None
+        if input_concat_cond is not None:
+
+            # Interpolate input_concat_cond to the same length as x
+            if input_concat_cond.shape[2] != x.shape[2]:
+                input_concat_cond = F.interpolate(input_concat_cond, (x.shape[2], ), mode='nearest')
+
+            channels_list = [input_concat_cond]
+
+        outputs = self.model(
+            x,
+            t,
+            features=global_cond,
+            channels_list=channels_list,
+            **kwargs)
+
+        return outputs
+
+class UNet1DUncondWrapper(DiffusionModel):
+    def __init__(
+        self,
+        in_channels,
+        *args,
+        **kwargs
+    ):
+        super().__init__()
+
+        self.model = UNet1d(in_channels=in_channels, *args, **kwargs)
+
+        self.io_channels = in_channels
+
+        with torch.no_grad():
+            for param in self.model.parameters():
+                param *= 0.5
+
+    def forward(self, x, t, **kwargs):
+        return self.model(x, t, **kwargs)
+
+class DAU1DCondWrapper(ConditionedDiffusionModel):
+    def __init__(
+        self,
+        *args,
+        **kwargs
+    ):
+        super().__init__(supports_cross_attention=False, supports_global_cond=False, supports_input_concat=True)
+
+        self.model = DiffusionAttnUnet1D(*args, **kwargs)
+
+        with torch.no_grad():
+            for param in self.model.parameters():
+                param *= 0.5
+
+    def forward(self,
+                x,
+                t,
+                input_concat_cond=None,
+                cross_attn_cond=None,
+                cross_attn_mask=None,
+                global_cond=None,
+                cfg_scale=1.0,
+                cfg_dropout_prob: float = 0.0,
+                batch_cfg: bool = False,
+                rescale_cfg: bool = False,
+                negative_cross_attn_cond=None,
+                negative_cross_attn_mask=None,
+                negative_global_cond=None,
+                negative_input_concat_cond=None,
+                prepend_cond=None,
+                **kwargs):
+
+        return self.model(x, t, cond = input_concat_cond)
+
+class DiffusionAttnUnet1D(nn.Module):
+    def __init__(
+        self,
+        io_channels = 2,
+        depth=14,
+        n_attn_layers = 6,
+        channels = [128, 128, 256, 256] + [512] * 10,
+        cond_dim = 0,
+        cond_noise_aug = False,
+        kernel_size = 5,
+        learned_resample = False,
+        strides = [2] * 13,
+        conv_bias = True,
+        use_snake = False
+    ):
+        super().__init__()
+
+        self.cond_noise_aug = cond_noise_aug
+
+        self.io_channels = io_channels
+
+        if self.cond_noise_aug:
+            self.rng = torch.quasirandom.SobolEngine(1, scramble=True)
+
+        self.timestep_embed = FourierFeatures(1, 16)
+
+        attn_layer = depth - n_attn_layers
+
+        strides = [1] + strides
+
+        block = nn.Identity()
+
+        conv_block = partial(ResConvBlock, kernel_size=kernel_size, conv_bias = conv_bias, use_snake=use_snake)
+
+        for i in range(depth, 0, -1):
+            c = channels[i - 1]
+            stride = strides[i-1]
+            if stride > 2 and not learned_resample:
+                raise ValueError("Must have stride 2 without learned resampling")
+
+            if i > 1:
+                c_prev = channels[i - 2]
+                add_attn = i >= attn_layer and n_attn_layers > 0
+                block = SkipBlock(
+                    Downsample1d_2(c_prev, c_prev, stride) if (learned_resample or stride == 1) else Downsample1d("cubic"),
+                    conv_block(c_prev, c, c),
+                    SelfAttention1d(
+                        c, c // 32) if add_attn else nn.Identity(),
+                    conv_block(c, c, c),
+                    SelfAttention1d(
+                        c, c // 32) if add_attn else nn.Identity(),
+                    conv_block(c, c, c),
+                    SelfAttention1d(
+                        c, c // 32) if add_attn else nn.Identity(),
+                    block,
+                    conv_block(c * 2 if i != depth else c, c, c),
+                    SelfAttention1d(
+                        c, c // 32) if add_attn else nn.Identity(),
+                    conv_block(c, c, c),
+                    SelfAttention1d(
+                        c, c // 32) if add_attn else nn.Identity(),
+                    conv_block(c, c, c_prev),
+                    SelfAttention1d(c_prev, c_prev //
+                                    32) if add_attn else nn.Identity(),
+                    Upsample1d_2(c_prev, c_prev, stride) if learned_resample else Upsample1d(kernel="cubic")
+                )
+            else:
+                cond_embed_dim = 16 if not self.cond_noise_aug else 32
+                block = nn.Sequential(
+                    conv_block((io_channels + cond_dim) + cond_embed_dim, c, c),
+                    conv_block(c, c, c),
+                    conv_block(c, c, c),
+                    block,
+                    conv_block(c * 2, c, c),
+                    conv_block(c, c, c),
+                    conv_block(c, c, io_channels, is_last=True),
+                )
+        self.net = block
+
+        with torch.no_grad():
+            for param in self.net.parameters():
+                param *= 0.5
+
+    def forward(self, x, t, cond=None, cond_aug_scale=None):
+
+        timestep_embed = expand_to_planes(self.timestep_embed(t[:, None]), x.shape)
+
+        inputs = [x, timestep_embed]
+
+        if cond is not None:
+            if cond.shape[2] != x.shape[2]:
+                cond = F.interpolate(cond, (x.shape[2], ), mode='linear', align_corners=False)
+
+            if self.cond_noise_aug:
+                # Get a random number between 0 and 1, uniformly sampled
+                if cond_aug_scale is None:
+                    aug_level = self.rng.draw(cond.shape[0])[:, 0].to(cond)
+                else:
+                    aug_level = torch.tensor([cond_aug_scale]).repeat([cond.shape[0]]).to(cond)
+
+                # Add noise to the conditioning signal
+                cond = cond + torch.randn_like(cond) * aug_level[:, None, None]
+
+                # Get embedding for noise cond level, reusing timestamp_embed
+                aug_level_embed = expand_to_planes(self.timestep_embed(aug_level[:, None]), x.shape)
+
+                inputs.append(aug_level_embed)
+
+            inputs.append(cond)
+
+        outputs = self.net(torch.cat(inputs, dim=1))
+
+        return outputs
+
+class DiTWrapper(ConditionedDiffusionModel):
+    def __init__(
+        self,
+        *args,
+        **kwargs
+    ):
+        super().__init__(supports_cross_attention=True, supports_global_cond=False, supports_input_concat=False)
+
+        self.model = DiffusionTransformer(*args, **kwargs)
+
+        with torch.no_grad():
+            for param in self.model.parameters():
+                param *= 0.5
+
+    def forward(self,
+                x,
+                t,
+                cross_attn_cond=None,
+                cross_attn_mask=None,
+                negative_cross_attn_cond=None,
+                negative_cross_attn_mask=None,
+                input_concat_cond=None,
+                negative_input_concat_cond=None,
+                global_cond=None,
+                negative_global_cond=None,
+                prepend_cond=None,
+                prepend_cond_mask=None,
+                cfg_scale=1.0,
+                cfg_dropout_prob: float = 0.0,
+                batch_cfg: bool = True,
+                rescale_cfg: bool = False,
+                scale_phi: float = 0.0,
+                **kwargs):
+
+        assert batch_cfg, "batch_cfg must be True for DiTWrapper"
+        assert negative_input_concat_cond is None, "negative_input_concat_cond is not supported for DiTWrapper"
+
+        return self.model(
+            x,
+            t,
+            cross_attn_cond=cross_attn_cond,
+            cross_attn_cond_mask=cross_attn_mask,
+            negative_cross_attn_cond=negative_cross_attn_cond,
+            negative_cross_attn_mask=negative_cross_attn_mask,
+            input_concat_cond=input_concat_cond,
+            prepend_cond=prepend_cond,
+            prepend_cond_mask=prepend_cond_mask,
+            cfg_scale=cfg_scale,
+            cfg_dropout_prob=cfg_dropout_prob,
+            scale_phi=scale_phi,
+            global_embed=global_cond,
+            **kwargs)
+
+class DiTUncondWrapper(DiffusionModel):
+    def __init__(
+        self,
+        in_channels,
+        *args,
+        **kwargs
+    ):
+        super().__init__()
+
+        self.model = DiffusionTransformer(io_channels=in_channels, *args, **kwargs)
+
+        self.io_channels = in_channels
+
+        with torch.no_grad():
+            for param in self.model.parameters():
+                param *= 0.5
+
+    def forward(self, x, t, **kwargs):
+        return self.model(x, t, **kwargs)
+
+def create_diffusion_uncond_from_config(config: tp.Dict[str, tp.Any]):
+    diffusion_uncond_config = config["model"]
+
+    model_type = diffusion_uncond_config.get('type', None)
+
+    diffusion_config = diffusion_uncond_config.get('config', {})
+
+    assert model_type is not None, "Must specify model type in config"
+
+    pretransform = diffusion_uncond_config.get("pretransform", None)
+
+    sample_size = config.get("sample_size", None)
+    assert sample_size is not None, "Must specify sample size in config"
+
+    sample_rate = config.get("sample_rate", None)
+    assert sample_rate is not None, "Must specify sample rate in config"
+
+    if pretransform is not None:
+        pretransform = create_pretransform_from_config(pretransform, sample_rate)
+        min_input_length = pretransform.downsampling_ratio
+    else:
+        min_input_length = 1
+
+    if model_type == 'DAU1d':
+
+        model = DiffusionAttnUnet1D(
+            **diffusion_config
+        )
+    
+    elif model_type == "adp_uncond_1d":
+
+        model = UNet1DUncondWrapper(
+            **diffusion_config
+        )
+
+    elif model_type == "dit":
+        model = DiTUncondWrapper(
+            **diffusion_config
+        )
+
+    else:
+        raise NotImplementedError(f'Unknown model type: {model_type}')
+
+    return DiffusionModelWrapper(model,
+                                io_channels=model.io_channels,
+                                sample_size=sample_size,
+                                sample_rate=sample_rate,
+                                pretransform=pretransform,
+                                min_input_length=min_input_length)
+
+def create_diffusion_cond_from_config(config: tp.Dict[str, tp.Any]):
+
+    model_config = config["model"]
+
+    model_type = config["model_type"]
+
+    diffusion_config = model_config.get('diffusion', None)
+    assert diffusion_config is not None, "Must specify diffusion config"
+
+    diffusion_model_type = diffusion_config.get('type', None)
+    assert diffusion_model_type is not None, "Must specify diffusion model type"
+
+    diffusion_model_config = diffusion_config.get('config', None)
+    assert diffusion_model_config is not None, "Must specify diffusion model config"
+
+    if diffusion_model_type == 'adp_cfg_1d':
+        diffusion_model = UNetCFG1DWrapper(**diffusion_model_config)
+    elif diffusion_model_type == 'adp_1d':
+        diffusion_model = UNet1DCondWrapper(**diffusion_model_config)
+    elif diffusion_model_type == 'dit':
+        diffusion_model = DiTWrapper(**diffusion_model_config)
+
+    io_channels = model_config.get('io_channels', None)
+    assert io_channels is not None, "Must specify io_channels in model config"
+
+    sample_rate = config.get('sample_rate', None)
+    assert sample_rate is not None, "Must specify sample_rate in config"
+
+    conditioning_config = model_config.get('conditioning', None)
+
+    conditioner = None
+    if conditioning_config is not None:
+        conditioner = create_multi_conditioner_from_conditioning_config(conditioning_config)
+
+    cross_attention_ids = diffusion_config.get('cross_attention_cond_ids', [])
+    global_cond_ids = diffusion_config.get('global_cond_ids', [])
+    input_concat_ids = diffusion_config.get('input_concat_ids', [])
+    prepend_cond_ids = diffusion_config.get('prepend_cond_ids', [])
+
+    pretransform = model_config.get("pretransform", None)
+
+    if pretransform is not None:
+        pretransform = create_pretransform_from_config(pretransform, sample_rate)
+        min_input_length = pretransform.downsampling_ratio
+    else:
+        min_input_length = 1
+
+    if diffusion_model_type == "adp_cfg_1d" or diffusion_model_type == "adp_1d":
+        min_input_length *= np.prod(diffusion_model_config["factors"])
+    elif diffusion_model_type == "dit":
+        min_input_length *= diffusion_model.model.patch_size
+
+    # Get the proper wrapper class
+
+    if model_type == "diffusion_cond" or model_type == "diffusion_cond_inpaint":
+        wrapper_fn = ConditionedDiffusionModelWrapper
+    elif model_type == "diffusion_prior":
+        prior_type = model_config.get("prior_type", None)
+        assert prior_type is not None, "Must specify prior_type in diffusion prior model config"
+
+        if prior_type == "mono_stereo":
+            from .diffusion_prior import MonoToStereoDiffusionPrior
+            wrapper_fn = MonoToStereoDiffusionPrior
+        elif prior_type == "source_separation":
+            from .diffusion_prior import SourceSeparationDiffusionPrior
+            wrapper_fn = SourceSeparationDiffusionPrior
+
+    return wrapper_fn(
+        diffusion_model,
+        conditioner,
+        min_input_length=min_input_length,
+        sample_rate=sample_rate,
+        cross_attn_cond_ids=cross_attention_ids,
+        global_cond_ids=global_cond_ids,
+        input_concat_ids=input_concat_ids,
+        prepend_cond_ids=prepend_cond_ids,
+        pretransform=pretransform,
+        io_channels=io_channels
+    )