cavargas10 commited on
Commit
87b3c4b
·
verified ·
1 Parent(s): 590f44b

Upload 11 files

Browse files
trellis/models/__init__.py ADDED
@@ -0,0 +1,96 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import importlib
2
+
3
+ __attributes = {
4
+ 'SparseStructureEncoder': 'sparse_structure_vae',
5
+ 'SparseStructureDecoder': 'sparse_structure_vae',
6
+
7
+ 'SparseStructureFlowModel': 'sparse_structure_flow',
8
+
9
+ 'SLatEncoder': 'structured_latent_vae',
10
+ 'SLatGaussianDecoder': 'structured_latent_vae',
11
+ 'SLatRadianceFieldDecoder': 'structured_latent_vae',
12
+ 'SLatMeshDecoder': 'structured_latent_vae',
13
+ 'ElasticSLatEncoder': 'structured_latent_vae',
14
+ 'ElasticSLatGaussianDecoder': 'structured_latent_vae',
15
+ 'ElasticSLatRadianceFieldDecoder': 'structured_latent_vae',
16
+ 'ElasticSLatMeshDecoder': 'structured_latent_vae',
17
+
18
+ 'SLatFlowModel': 'structured_latent_flow',
19
+ 'ElasticSLatFlowModel': 'structured_latent_flow',
20
+ }
21
+
22
+ __submodules = []
23
+
24
+ __all__ = list(__attributes.keys()) + __submodules
25
+
26
+ def __getattr__(name):
27
+ if name not in globals():
28
+ if name in __attributes:
29
+ module_name = __attributes[name]
30
+ module = importlib.import_module(f".{module_name}", __name__)
31
+ globals()[name] = getattr(module, name)
32
+ elif name in __submodules:
33
+ module = importlib.import_module(f".{name}", __name__)
34
+ globals()[name] = module
35
+ else:
36
+ raise AttributeError(f"module {__name__} has no attribute {name}")
37
+ return globals()[name]
38
+
39
+
40
+ def from_pretrained(path: str, **kwargs):
41
+ """
42
+ Load a model from a pretrained checkpoint.
43
+
44
+ Args:
45
+ path: The path to the checkpoint. Can be either local path or a Hugging Face model name.
46
+ NOTE: config file and model file should take the name f'{path}.json' and f'{path}.safetensors' respectively.
47
+ **kwargs: Additional arguments for the model constructor.
48
+ """
49
+ import os
50
+ import json
51
+ from safetensors.torch import load_file
52
+ is_local = os.path.exists(f"{path}.json") and os.path.exists(f"{path}.safetensors")
53
+
54
+ if is_local:
55
+ config_file = f"{path}.json"
56
+ model_file = f"{path}.safetensors"
57
+ else:
58
+ from huggingface_hub import hf_hub_download
59
+ path_parts = path.split('/')
60
+ repo_id = f'{path_parts[0]}/{path_parts[1]}'
61
+ model_name = '/'.join(path_parts[2:])
62
+ config_file = hf_hub_download(repo_id, f"{model_name}.json")
63
+ model_file = hf_hub_download(repo_id, f"{model_name}.safetensors")
64
+
65
+ with open(config_file, 'r') as f:
66
+ config = json.load(f)
67
+ model = __getattr__(config['name'])(**config['args'], **kwargs)
68
+ model.load_state_dict(load_file(model_file))
69
+
70
+ return model
71
+
72
+
73
+ # For Pylance
74
+ if __name__ == '__main__':
75
+ from .sparse_structure_vae import (
76
+ SparseStructureEncoder,
77
+ SparseStructureDecoder,
78
+ )
79
+
80
+ from .sparse_structure_flow import SparseStructureFlowModel
81
+
82
+ from .structured_latent_vae import (
83
+ SLatEncoder,
84
+ SLatGaussianDecoder,
85
+ SLatRadianceFieldDecoder,
86
+ SLatMeshDecoder,
87
+ ElasticSLatEncoder,
88
+ ElasticSLatGaussianDecoder,
89
+ ElasticSLatRadianceFieldDecoder,
90
+ ElasticSLatMeshDecoder,
91
+ )
92
+
93
+ from .structured_latent_flow import (
94
+ SLatFlowModel,
95
+ ElasticSLatFlowModel,
96
+ )
trellis/models/sparse_elastic_mixin.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from contextlib import contextmanager
2
+ from typing import *
3
+ import math
4
+ from ..modules import sparse as sp
5
+ from ..utils.elastic_utils import ElasticModuleMixin
6
+
7
+
8
+ class SparseTransformerElasticMixin(ElasticModuleMixin):
9
+ def _get_input_size(self, x: sp.SparseTensor, *args, **kwargs):
10
+ return x.feats.shape[0]
11
+
12
+ @contextmanager
13
+ def with_mem_ratio(self, mem_ratio=1.0):
14
+ if mem_ratio == 1.0:
15
+ yield 1.0
16
+ return
17
+ num_blocks = len(self.blocks)
18
+ num_checkpoint_blocks = min(math.ceil((1 - mem_ratio) * num_blocks) + 1, num_blocks)
19
+ exact_mem_ratio = 1 - (num_checkpoint_blocks - 1) / num_blocks
20
+ for i in range(num_blocks):
21
+ self.blocks[i].use_checkpoint = i < num_checkpoint_blocks
22
+ yield exact_mem_ratio
23
+ for i in range(num_blocks):
24
+ self.blocks[i].use_checkpoint = False
trellis/models/sparse_structure_flow.py ADDED
@@ -0,0 +1,200 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import *
2
+ import torch
3
+ import torch.nn as nn
4
+ import torch.nn.functional as F
5
+ import numpy as np
6
+ from ..modules.utils import convert_module_to_f16, convert_module_to_f32
7
+ from ..modules.transformer import AbsolutePositionEmbedder, ModulatedTransformerCrossBlock
8
+ from ..modules.spatial import patchify, unpatchify
9
+
10
+
11
+ class TimestepEmbedder(nn.Module):
12
+ """
13
+ Embeds scalar timesteps into vector representations.
14
+ """
15
+ def __init__(self, hidden_size, frequency_embedding_size=256):
16
+ super().__init__()
17
+ self.mlp = nn.Sequential(
18
+ nn.Linear(frequency_embedding_size, hidden_size, bias=True),
19
+ nn.SiLU(),
20
+ nn.Linear(hidden_size, hidden_size, bias=True),
21
+ )
22
+ self.frequency_embedding_size = frequency_embedding_size
23
+
24
+ @staticmethod
25
+ def timestep_embedding(t, dim, max_period=10000):
26
+ """
27
+ Create sinusoidal timestep embeddings.
28
+
29
+ Args:
30
+ t: a 1-D Tensor of N indices, one per batch element.
31
+ These may be fractional.
32
+ dim: the dimension of the output.
33
+ max_period: controls the minimum frequency of the embeddings.
34
+
35
+ Returns:
36
+ an (N, D) Tensor of positional embeddings.
37
+ """
38
+ # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
39
+ half = dim // 2
40
+ freqs = torch.exp(
41
+ -np.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
42
+ ).to(device=t.device)
43
+ args = t[:, None].float() * freqs[None]
44
+ embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
45
+ if dim % 2:
46
+ embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
47
+ return embedding
48
+
49
+ def forward(self, t):
50
+ t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
51
+ t_emb = self.mlp(t_freq)
52
+ return t_emb
53
+
54
+
55
+ class SparseStructureFlowModel(nn.Module):
56
+ def __init__(
57
+ self,
58
+ resolution: int,
59
+ in_channels: int,
60
+ model_channels: int,
61
+ cond_channels: int,
62
+ out_channels: int,
63
+ num_blocks: int,
64
+ num_heads: Optional[int] = None,
65
+ num_head_channels: Optional[int] = 64,
66
+ mlp_ratio: float = 4,
67
+ patch_size: int = 2,
68
+ pe_mode: Literal["ape", "rope"] = "ape",
69
+ use_fp16: bool = False,
70
+ use_checkpoint: bool = False,
71
+ share_mod: bool = False,
72
+ qk_rms_norm: bool = False,
73
+ qk_rms_norm_cross: bool = False,
74
+ ):
75
+ super().__init__()
76
+ self.resolution = resolution
77
+ self.in_channels = in_channels
78
+ self.model_channels = model_channels
79
+ self.cond_channels = cond_channels
80
+ self.out_channels = out_channels
81
+ self.num_blocks = num_blocks
82
+ self.num_heads = num_heads or model_channels // num_head_channels
83
+ self.mlp_ratio = mlp_ratio
84
+ self.patch_size = patch_size
85
+ self.pe_mode = pe_mode
86
+ self.use_fp16 = use_fp16
87
+ self.use_checkpoint = use_checkpoint
88
+ self.share_mod = share_mod
89
+ self.qk_rms_norm = qk_rms_norm
90
+ self.qk_rms_norm_cross = qk_rms_norm_cross
91
+ self.dtype = torch.float16 if use_fp16 else torch.float32
92
+
93
+ self.t_embedder = TimestepEmbedder(model_channels)
94
+ if share_mod:
95
+ self.adaLN_modulation = nn.Sequential(
96
+ nn.SiLU(),
97
+ nn.Linear(model_channels, 6 * model_channels, bias=True)
98
+ )
99
+
100
+ if pe_mode == "ape":
101
+ pos_embedder = AbsolutePositionEmbedder(model_channels, 3)
102
+ coords = torch.meshgrid(*[torch.arange(res, device=self.device) for res in [resolution // patch_size] * 3], indexing='ij')
103
+ coords = torch.stack(coords, dim=-1).reshape(-1, 3)
104
+ pos_emb = pos_embedder(coords)
105
+ self.register_buffer("pos_emb", pos_emb)
106
+
107
+ self.input_layer = nn.Linear(in_channels * patch_size**3, model_channels)
108
+
109
+ self.blocks = nn.ModuleList([
110
+ ModulatedTransformerCrossBlock(
111
+ model_channels,
112
+ cond_channels,
113
+ num_heads=self.num_heads,
114
+ mlp_ratio=self.mlp_ratio,
115
+ attn_mode='full',
116
+ use_checkpoint=self.use_checkpoint,
117
+ use_rope=(pe_mode == "rope"),
118
+ share_mod=share_mod,
119
+ qk_rms_norm=self.qk_rms_norm,
120
+ qk_rms_norm_cross=self.qk_rms_norm_cross,
121
+ )
122
+ for _ in range(num_blocks)
123
+ ])
124
+
125
+ self.out_layer = nn.Linear(model_channels, out_channels * patch_size**3)
126
+
127
+ self.initialize_weights()
128
+ if use_fp16:
129
+ self.convert_to_fp16()
130
+
131
+ @property
132
+ def device(self) -> torch.device:
133
+ """
134
+ Return the device of the model.
135
+ """
136
+ return next(self.parameters()).device
137
+
138
+ def convert_to_fp16(self) -> None:
139
+ """
140
+ Convert the torso of the model to float16.
141
+ """
142
+ self.blocks.apply(convert_module_to_f16)
143
+
144
+ def convert_to_fp32(self) -> None:
145
+ """
146
+ Convert the torso of the model to float32.
147
+ """
148
+ self.blocks.apply(convert_module_to_f32)
149
+
150
+ def initialize_weights(self) -> None:
151
+ # Initialize transformer layers:
152
+ def _basic_init(module):
153
+ if isinstance(module, nn.Linear):
154
+ torch.nn.init.xavier_uniform_(module.weight)
155
+ if module.bias is not None:
156
+ nn.init.constant_(module.bias, 0)
157
+ self.apply(_basic_init)
158
+
159
+ # Initialize timestep embedding MLP:
160
+ nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
161
+ nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
162
+
163
+ # Zero-out adaLN modulation layers in DiT blocks:
164
+ if self.share_mod:
165
+ nn.init.constant_(self.adaLN_modulation[-1].weight, 0)
166
+ nn.init.constant_(self.adaLN_modulation[-1].bias, 0)
167
+ else:
168
+ for block in self.blocks:
169
+ nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
170
+ nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
171
+
172
+ # Zero-out output layers:
173
+ nn.init.constant_(self.out_layer.weight, 0)
174
+ nn.init.constant_(self.out_layer.bias, 0)
175
+
176
+ def forward(self, x: torch.Tensor, t: torch.Tensor, cond: torch.Tensor) -> torch.Tensor:
177
+ assert [*x.shape] == [x.shape[0], self.in_channels, *[self.resolution] * 3], \
178
+ f"Input shape mismatch, got {x.shape}, expected {[x.shape[0], self.in_channels, *[self.resolution] * 3]}"
179
+
180
+ h = patchify(x, self.patch_size)
181
+ h = h.view(*h.shape[:2], -1).permute(0, 2, 1).contiguous()
182
+
183
+ h = self.input_layer(h)
184
+ h = h + self.pos_emb[None]
185
+ t_emb = self.t_embedder(t)
186
+ if self.share_mod:
187
+ t_emb = self.adaLN_modulation(t_emb)
188
+ t_emb = t_emb.type(self.dtype)
189
+ h = h.type(self.dtype)
190
+ cond = cond.type(self.dtype)
191
+ for block in self.blocks:
192
+ h = block(h, t_emb, cond)
193
+ h = h.type(x.dtype)
194
+ h = F.layer_norm(h, h.shape[-1:])
195
+ h = self.out_layer(h)
196
+
197
+ h = h.permute(0, 2, 1).view(h.shape[0], h.shape[2], *[self.resolution // self.patch_size] * 3)
198
+ h = unpatchify(h, self.patch_size).contiguous()
199
+
200
+ return h
trellis/models/sparse_structure_vae.py ADDED
@@ -0,0 +1,306 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import *
2
+ import torch
3
+ import torch.nn as nn
4
+ import torch.nn.functional as F
5
+ from ..modules.norm import GroupNorm32, ChannelLayerNorm32
6
+ from ..modules.spatial import pixel_shuffle_3d
7
+ from ..modules.utils import zero_module, convert_module_to_f16, convert_module_to_f32
8
+
9
+
10
+ def norm_layer(norm_type: str, *args, **kwargs) -> nn.Module:
11
+ """
12
+ Return a normalization layer.
13
+ """
14
+ if norm_type == "group":
15
+ return GroupNorm32(32, *args, **kwargs)
16
+ elif norm_type == "layer":
17
+ return ChannelLayerNorm32(*args, **kwargs)
18
+ else:
19
+ raise ValueError(f"Invalid norm type {norm_type}")
20
+
21
+
22
+ class ResBlock3d(nn.Module):
23
+ def __init__(
24
+ self,
25
+ channels: int,
26
+ out_channels: Optional[int] = None,
27
+ norm_type: Literal["group", "layer"] = "layer",
28
+ ):
29
+ super().__init__()
30
+ self.channels = channels
31
+ self.out_channels = out_channels or channels
32
+
33
+ self.norm1 = norm_layer(norm_type, channels)
34
+ self.norm2 = norm_layer(norm_type, self.out_channels)
35
+ self.conv1 = nn.Conv3d(channels, self.out_channels, 3, padding=1)
36
+ self.conv2 = zero_module(nn.Conv3d(self.out_channels, self.out_channels, 3, padding=1))
37
+ self.skip_connection = nn.Conv3d(channels, self.out_channels, 1) if channels != self.out_channels else nn.Identity()
38
+
39
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
40
+ h = self.norm1(x)
41
+ h = F.silu(h)
42
+ h = self.conv1(h)
43
+ h = self.norm2(h)
44
+ h = F.silu(h)
45
+ h = self.conv2(h)
46
+ h = h + self.skip_connection(x)
47
+ return h
48
+
49
+
50
+ class DownsampleBlock3d(nn.Module):
51
+ def __init__(
52
+ self,
53
+ in_channels: int,
54
+ out_channels: int,
55
+ mode: Literal["conv", "avgpool"] = "conv",
56
+ ):
57
+ assert mode in ["conv", "avgpool"], f"Invalid mode {mode}"
58
+
59
+ super().__init__()
60
+ self.in_channels = in_channels
61
+ self.out_channels = out_channels
62
+
63
+ if mode == "conv":
64
+ self.conv = nn.Conv3d(in_channels, out_channels, 2, stride=2)
65
+ elif mode == "avgpool":
66
+ assert in_channels == out_channels, "Pooling mode requires in_channels to be equal to out_channels"
67
+
68
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
69
+ if hasattr(self, "conv"):
70
+ return self.conv(x)
71
+ else:
72
+ return F.avg_pool3d(x, 2)
73
+
74
+
75
+ class UpsampleBlock3d(nn.Module):
76
+ def __init__(
77
+ self,
78
+ in_channels: int,
79
+ out_channels: int,
80
+ mode: Literal["conv", "nearest"] = "conv",
81
+ ):
82
+ assert mode in ["conv", "nearest"], f"Invalid mode {mode}"
83
+
84
+ super().__init__()
85
+ self.in_channels = in_channels
86
+ self.out_channels = out_channels
87
+
88
+ if mode == "conv":
89
+ self.conv = nn.Conv3d(in_channels, out_channels*8, 3, padding=1)
90
+ elif mode == "nearest":
91
+ assert in_channels == out_channels, "Nearest mode requires in_channels to be equal to out_channels"
92
+
93
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
94
+ if hasattr(self, "conv"):
95
+ x = self.conv(x)
96
+ return pixel_shuffle_3d(x, 2)
97
+ else:
98
+ return F.interpolate(x, scale_factor=2, mode="nearest")
99
+
100
+
101
+ class SparseStructureEncoder(nn.Module):
102
+ """
103
+ Encoder for Sparse Structure (\mathcal{E}_S in the paper Sec. 3.3).
104
+
105
+ Args:
106
+ in_channels (int): Channels of the input.
107
+ latent_channels (int): Channels of the latent representation.
108
+ num_res_blocks (int): Number of residual blocks at each resolution.
109
+ channels (List[int]): Channels of the encoder blocks.
110
+ num_res_blocks_middle (int): Number of residual blocks in the middle.
111
+ norm_type (Literal["group", "layer"]): Type of normalization layer.
112
+ use_fp16 (bool): Whether to use FP16.
113
+ """
114
+ def __init__(
115
+ self,
116
+ in_channels: int,
117
+ latent_channels: int,
118
+ num_res_blocks: int,
119
+ channels: List[int],
120
+ num_res_blocks_middle: int = 2,
121
+ norm_type: Literal["group", "layer"] = "layer",
122
+ use_fp16: bool = False,
123
+ ):
124
+ super().__init__()
125
+ self.in_channels = in_channels
126
+ self.latent_channels = latent_channels
127
+ self.num_res_blocks = num_res_blocks
128
+ self.channels = channels
129
+ self.num_res_blocks_middle = num_res_blocks_middle
130
+ self.norm_type = norm_type
131
+ self.use_fp16 = use_fp16
132
+ self.dtype = torch.float16 if use_fp16 else torch.float32
133
+
134
+ self.input_layer = nn.Conv3d(in_channels, channels[0], 3, padding=1)
135
+
136
+ self.blocks = nn.ModuleList([])
137
+ for i, ch in enumerate(channels):
138
+ self.blocks.extend([
139
+ ResBlock3d(ch, ch)
140
+ for _ in range(num_res_blocks)
141
+ ])
142
+ if i < len(channels) - 1:
143
+ self.blocks.append(
144
+ DownsampleBlock3d(ch, channels[i+1])
145
+ )
146
+
147
+ self.middle_block = nn.Sequential(*[
148
+ ResBlock3d(channels[-1], channels[-1])
149
+ for _ in range(num_res_blocks_middle)
150
+ ])
151
+
152
+ self.out_layer = nn.Sequential(
153
+ norm_layer(norm_type, channels[-1]),
154
+ nn.SiLU(),
155
+ nn.Conv3d(channels[-1], latent_channels*2, 3, padding=1)
156
+ )
157
+
158
+ if use_fp16:
159
+ self.convert_to_fp16()
160
+
161
+ @property
162
+ def device(self) -> torch.device:
163
+ """
164
+ Return the device of the model.
165
+ """
166
+ return next(self.parameters()).device
167
+
168
+ def convert_to_fp16(self) -> None:
169
+ """
170
+ Convert the torso of the model to float16.
171
+ """
172
+ self.use_fp16 = True
173
+ self.dtype = torch.float16
174
+ self.blocks.apply(convert_module_to_f16)
175
+ self.middle_block.apply(convert_module_to_f16)
176
+
177
+ def convert_to_fp32(self) -> None:
178
+ """
179
+ Convert the torso of the model to float32.
180
+ """
181
+ self.use_fp16 = False
182
+ self.dtype = torch.float32
183
+ self.blocks.apply(convert_module_to_f32)
184
+ self.middle_block.apply(convert_module_to_f32)
185
+
186
+ def forward(self, x: torch.Tensor, sample_posterior: bool = False, return_raw: bool = False) -> torch.Tensor:
187
+ h = self.input_layer(x)
188
+ h = h.type(self.dtype)
189
+
190
+ for block in self.blocks:
191
+ h = block(h)
192
+ h = self.middle_block(h)
193
+
194
+ h = h.type(x.dtype)
195
+ h = self.out_layer(h)
196
+
197
+ mean, logvar = h.chunk(2, dim=1)
198
+
199
+ if sample_posterior:
200
+ std = torch.exp(0.5 * logvar)
201
+ z = mean + std * torch.randn_like(std)
202
+ else:
203
+ z = mean
204
+
205
+ if return_raw:
206
+ return z, mean, logvar
207
+ return z
208
+
209
+
210
+ class SparseStructureDecoder(nn.Module):
211
+ """
212
+ Decoder for Sparse Structure (\mathcal{D}_S in the paper Sec. 3.3).
213
+
214
+ Args:
215
+ out_channels (int): Channels of the output.
216
+ latent_channels (int): Channels of the latent representation.
217
+ num_res_blocks (int): Number of residual blocks at each resolution.
218
+ channels (List[int]): Channels of the decoder blocks.
219
+ num_res_blocks_middle (int): Number of residual blocks in the middle.
220
+ norm_type (Literal["group", "layer"]): Type of normalization layer.
221
+ use_fp16 (bool): Whether to use FP16.
222
+ """
223
+ def __init__(
224
+ self,
225
+ out_channels: int,
226
+ latent_channels: int,
227
+ num_res_blocks: int,
228
+ channels: List[int],
229
+ num_res_blocks_middle: int = 2,
230
+ norm_type: Literal["group", "layer"] = "layer",
231
+ use_fp16: bool = False,
232
+ ):
233
+ super().__init__()
234
+ self.out_channels = out_channels
235
+ self.latent_channels = latent_channels
236
+ self.num_res_blocks = num_res_blocks
237
+ self.channels = channels
238
+ self.num_res_blocks_middle = num_res_blocks_middle
239
+ self.norm_type = norm_type
240
+ self.use_fp16 = use_fp16
241
+ self.dtype = torch.float16 if use_fp16 else torch.float32
242
+
243
+ self.input_layer = nn.Conv3d(latent_channels, channels[0], 3, padding=1)
244
+
245
+ self.middle_block = nn.Sequential(*[
246
+ ResBlock3d(channels[0], channels[0])
247
+ for _ in range(num_res_blocks_middle)
248
+ ])
249
+
250
+ self.blocks = nn.ModuleList([])
251
+ for i, ch in enumerate(channels):
252
+ self.blocks.extend([
253
+ ResBlock3d(ch, ch)
254
+ for _ in range(num_res_blocks)
255
+ ])
256
+ if i < len(channels) - 1:
257
+ self.blocks.append(
258
+ UpsampleBlock3d(ch, channels[i+1])
259
+ )
260
+
261
+ self.out_layer = nn.Sequential(
262
+ norm_layer(norm_type, channels[-1]),
263
+ nn.SiLU(),
264
+ nn.Conv3d(channels[-1], out_channels, 3, padding=1)
265
+ )
266
+
267
+ if use_fp16:
268
+ self.convert_to_fp16()
269
+
270
+ @property
271
+ def device(self) -> torch.device:
272
+ """
273
+ Return the device of the model.
274
+ """
275
+ return next(self.parameters()).device
276
+
277
+ def convert_to_fp16(self) -> None:
278
+ """
279
+ Convert the torso of the model to float16.
280
+ """
281
+ self.use_fp16 = True
282
+ self.dtype = torch.float16
283
+ self.blocks.apply(convert_module_to_f16)
284
+ self.middle_block.apply(convert_module_to_f16)
285
+
286
+ def convert_to_fp32(self) -> None:
287
+ """
288
+ Convert the torso of the model to float32.
289
+ """
290
+ self.use_fp16 = False
291
+ self.dtype = torch.float32
292
+ self.blocks.apply(convert_module_to_f32)
293
+ self.middle_block.apply(convert_module_to_f32)
294
+
295
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
296
+ h = self.input_layer(x)
297
+
298
+ h = h.type(self.dtype)
299
+
300
+ h = self.middle_block(h)
301
+ for block in self.blocks:
302
+ h = block(h)
303
+
304
+ h = h.type(x.dtype)
305
+ h = self.out_layer(h)
306
+ return h
trellis/models/structured_latent_flow.py ADDED
@@ -0,0 +1,276 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import *
2
+ import torch
3
+ import torch.nn as nn
4
+ import torch.nn.functional as F
5
+ import numpy as np
6
+ from ..modules.utils import zero_module, convert_module_to_f16, convert_module_to_f32
7
+ from ..modules.transformer import AbsolutePositionEmbedder
8
+ from ..modules.norm import LayerNorm32
9
+ from ..modules import sparse as sp
10
+ from ..modules.sparse.transformer import ModulatedSparseTransformerCrossBlock
11
+ from .sparse_structure_flow import TimestepEmbedder
12
+ from .sparse_elastic_mixin import SparseTransformerElasticMixin
13
+
14
+
15
+ class SparseResBlock3d(nn.Module):
16
+ def __init__(
17
+ self,
18
+ channels: int,
19
+ emb_channels: int,
20
+ out_channels: Optional[int] = None,
21
+ downsample: bool = False,
22
+ upsample: bool = False,
23
+ ):
24
+ super().__init__()
25
+ self.channels = channels
26
+ self.emb_channels = emb_channels
27
+ self.out_channels = out_channels or channels
28
+ self.downsample = downsample
29
+ self.upsample = upsample
30
+
31
+ assert not (downsample and upsample), "Cannot downsample and upsample at the same time"
32
+
33
+ self.norm1 = LayerNorm32(channels, elementwise_affine=True, eps=1e-6)
34
+ self.norm2 = LayerNorm32(self.out_channels, elementwise_affine=False, eps=1e-6)
35
+ self.conv1 = sp.SparseConv3d(channels, self.out_channels, 3)
36
+ self.conv2 = zero_module(sp.SparseConv3d(self.out_channels, self.out_channels, 3))
37
+ self.emb_layers = nn.Sequential(
38
+ nn.SiLU(),
39
+ nn.Linear(emb_channels, 2 * self.out_channels, bias=True),
40
+ )
41
+ self.skip_connection = sp.SparseLinear(channels, self.out_channels) if channels != self.out_channels else nn.Identity()
42
+ self.updown = None
43
+ if self.downsample:
44
+ self.updown = sp.SparseDownsample(2)
45
+ elif self.upsample:
46
+ self.updown = sp.SparseUpsample(2)
47
+
48
+ def _updown(self, x: sp.SparseTensor) -> sp.SparseTensor:
49
+ if self.updown is not None:
50
+ x = self.updown(x)
51
+ return x
52
+
53
+ def forward(self, x: sp.SparseTensor, emb: torch.Tensor) -> sp.SparseTensor:
54
+ emb_out = self.emb_layers(emb).type(x.dtype)
55
+ scale, shift = torch.chunk(emb_out, 2, dim=1)
56
+
57
+ x = self._updown(x)
58
+ h = x.replace(self.norm1(x.feats))
59
+ h = h.replace(F.silu(h.feats))
60
+ h = self.conv1(h)
61
+ h = h.replace(self.norm2(h.feats)) * (1 + scale) + shift
62
+ h = h.replace(F.silu(h.feats))
63
+ h = self.conv2(h)
64
+ h = h + self.skip_connection(x)
65
+
66
+ return h
67
+
68
+
69
+ class SLatFlowModel(nn.Module):
70
+ def __init__(
71
+ self,
72
+ resolution: int,
73
+ in_channels: int,
74
+ model_channels: int,
75
+ cond_channels: int,
76
+ out_channels: int,
77
+ num_blocks: int,
78
+ num_heads: Optional[int] = None,
79
+ num_head_channels: Optional[int] = 64,
80
+ mlp_ratio: float = 4,
81
+ patch_size: int = 2,
82
+ num_io_res_blocks: int = 2,
83
+ io_block_channels: List[int] = None,
84
+ pe_mode: Literal["ape", "rope"] = "ape",
85
+ use_fp16: bool = False,
86
+ use_checkpoint: bool = False,
87
+ use_skip_connection: bool = True,
88
+ share_mod: bool = False,
89
+ qk_rms_norm: bool = False,
90
+ qk_rms_norm_cross: bool = False,
91
+ ):
92
+ super().__init__()
93
+ self.resolution = resolution
94
+ self.in_channels = in_channels
95
+ self.model_channels = model_channels
96
+ self.cond_channels = cond_channels
97
+ self.out_channels = out_channels
98
+ self.num_blocks = num_blocks
99
+ self.num_heads = num_heads or model_channels // num_head_channels
100
+ self.mlp_ratio = mlp_ratio
101
+ self.patch_size = patch_size
102
+ self.num_io_res_blocks = num_io_res_blocks
103
+ self.io_block_channels = io_block_channels
104
+ self.pe_mode = pe_mode
105
+ self.use_fp16 = use_fp16
106
+ self.use_checkpoint = use_checkpoint
107
+ self.use_skip_connection = use_skip_connection
108
+ self.share_mod = share_mod
109
+ self.qk_rms_norm = qk_rms_norm
110
+ self.qk_rms_norm_cross = qk_rms_norm_cross
111
+ self.dtype = torch.float16 if use_fp16 else torch.float32
112
+
113
+ if self.io_block_channels is not None:
114
+ assert int(np.log2(patch_size)) == np.log2(patch_size), "Patch size must be a power of 2"
115
+ assert np.log2(patch_size) == len(io_block_channels), "Number of IO ResBlocks must match the number of stages"
116
+
117
+ self.t_embedder = TimestepEmbedder(model_channels)
118
+ if share_mod:
119
+ self.adaLN_modulation = nn.Sequential(
120
+ nn.SiLU(),
121
+ nn.Linear(model_channels, 6 * model_channels, bias=True)
122
+ )
123
+
124
+ if pe_mode == "ape":
125
+ self.pos_embedder = AbsolutePositionEmbedder(model_channels)
126
+
127
+ self.input_layer = sp.SparseLinear(in_channels, model_channels if io_block_channels is None else io_block_channels[0])
128
+
129
+ self.input_blocks = nn.ModuleList([])
130
+ if io_block_channels is not None:
131
+ for chs, next_chs in zip(io_block_channels, io_block_channels[1:] + [model_channels]):
132
+ self.input_blocks.extend([
133
+ SparseResBlock3d(
134
+ chs,
135
+ model_channels,
136
+ out_channels=chs,
137
+ )
138
+ for _ in range(num_io_res_blocks-1)
139
+ ])
140
+ self.input_blocks.append(
141
+ SparseResBlock3d(
142
+ chs,
143
+ model_channels,
144
+ out_channels=next_chs,
145
+ downsample=True,
146
+ )
147
+ )
148
+
149
+ self.blocks = nn.ModuleList([
150
+ ModulatedSparseTransformerCrossBlock(
151
+ model_channels,
152
+ cond_channels,
153
+ num_heads=self.num_heads,
154
+ mlp_ratio=self.mlp_ratio,
155
+ attn_mode='full',
156
+ use_checkpoint=self.use_checkpoint,
157
+ use_rope=(pe_mode == "rope"),
158
+ share_mod=self.share_mod,
159
+ qk_rms_norm=self.qk_rms_norm,
160
+ qk_rms_norm_cross=self.qk_rms_norm_cross,
161
+ )
162
+ for _ in range(num_blocks)
163
+ ])
164
+
165
+ self.out_blocks = nn.ModuleList([])
166
+ if io_block_channels is not None:
167
+ for chs, prev_chs in zip(reversed(io_block_channels), [model_channels] + list(reversed(io_block_channels[1:]))):
168
+ self.out_blocks.append(
169
+ SparseResBlock3d(
170
+ prev_chs * 2 if self.use_skip_connection else prev_chs,
171
+ model_channels,
172
+ out_channels=chs,
173
+ upsample=True,
174
+ )
175
+ )
176
+ self.out_blocks.extend([
177
+ SparseResBlock3d(
178
+ chs * 2 if self.use_skip_connection else chs,
179
+ model_channels,
180
+ out_channels=chs,
181
+ )
182
+ for _ in range(num_io_res_blocks-1)
183
+ ])
184
+
185
+ self.out_layer = sp.SparseLinear(model_channels if io_block_channels is None else io_block_channels[0], out_channels)
186
+
187
+ self.initialize_weights()
188
+ if use_fp16:
189
+ self.convert_to_fp16()
190
+
191
+ @property
192
+ def device(self) -> torch.device:
193
+ """
194
+ Return the device of the model.
195
+ """
196
+ return next(self.parameters()).device
197
+
198
+ def convert_to_fp16(self) -> None:
199
+ """
200
+ Convert the torso of the model to float16.
201
+ """
202
+ self.input_blocks.apply(convert_module_to_f16)
203
+ self.blocks.apply(convert_module_to_f16)
204
+ self.out_blocks.apply(convert_module_to_f16)
205
+
206
+ def convert_to_fp32(self) -> None:
207
+ """
208
+ Convert the torso of the model to float32.
209
+ """
210
+ self.input_blocks.apply(convert_module_to_f32)
211
+ self.blocks.apply(convert_module_to_f32)
212
+ self.out_blocks.apply(convert_module_to_f32)
213
+
214
+ def initialize_weights(self) -> None:
215
+ # Initialize transformer layers:
216
+ def _basic_init(module):
217
+ if isinstance(module, nn.Linear):
218
+ torch.nn.init.xavier_uniform_(module.weight)
219
+ if module.bias is not None:
220
+ nn.init.constant_(module.bias, 0)
221
+ self.apply(_basic_init)
222
+
223
+ # Initialize timestep embedding MLP:
224
+ nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
225
+ nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
226
+
227
+ # Zero-out adaLN modulation layers in DiT blocks:
228
+ if self.share_mod:
229
+ nn.init.constant_(self.adaLN_modulation[-1].weight, 0)
230
+ nn.init.constant_(self.adaLN_modulation[-1].bias, 0)
231
+ else:
232
+ for block in self.blocks:
233
+ nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
234
+ nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
235
+
236
+ # Zero-out output layers:
237
+ nn.init.constant_(self.out_layer.weight, 0)
238
+ nn.init.constant_(self.out_layer.bias, 0)
239
+
240
+ def forward(self, x: sp.SparseTensor, t: torch.Tensor, cond: torch.Tensor) -> sp.SparseTensor:
241
+ h = self.input_layer(x).type(self.dtype)
242
+ t_emb = self.t_embedder(t)
243
+ if self.share_mod:
244
+ t_emb = self.adaLN_modulation(t_emb)
245
+ t_emb = t_emb.type(self.dtype)
246
+ cond = cond.type(self.dtype)
247
+
248
+ skips = []
249
+ # pack with input blocks
250
+ for block in self.input_blocks:
251
+ h = block(h, t_emb)
252
+ skips.append(h.feats)
253
+
254
+ if self.pe_mode == "ape":
255
+ h = h + self.pos_embedder(h.coords[:, 1:]).type(self.dtype)
256
+ for block in self.blocks:
257
+ h = block(h, t_emb, cond)
258
+
259
+ # unpack with output blocks
260
+ for block, skip in zip(self.out_blocks, reversed(skips)):
261
+ if self.use_skip_connection:
262
+ h = block(h.replace(torch.cat([h.feats, skip], dim=1)), t_emb)
263
+ else:
264
+ h = block(h, t_emb)
265
+
266
+ h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
267
+ h = self.out_layer(h.type(x.dtype))
268
+ return h
269
+
270
+
271
+ class ElasticSLatFlowModel(SparseTransformerElasticMixin, SLatFlowModel):
272
+ """
273
+ SLat Flow Model with elastic memory management.
274
+ Used for training with low VRAM.
275
+ """
276
+ pass
trellis/models/structured_latent_vae/__init__.py ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ from .encoder import SLatEncoder, ElasticSLatEncoder
2
+ from .decoder_gs import SLatGaussianDecoder, ElasticSLatGaussianDecoder
3
+ from .decoder_rf import SLatRadianceFieldDecoder, ElasticSLatRadianceFieldDecoder
4
+ from .decoder_mesh import SLatMeshDecoder, ElasticSLatMeshDecoder
trellis/models/structured_latent_vae/base.py ADDED
@@ -0,0 +1,117 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import *
2
+ import torch
3
+ import torch.nn as nn
4
+ from ...modules.utils import convert_module_to_f16, convert_module_to_f32
5
+ from ...modules import sparse as sp
6
+ from ...modules.transformer import AbsolutePositionEmbedder
7
+ from ...modules.sparse.transformer import SparseTransformerBlock
8
+
9
+
10
+ def block_attn_config(self):
11
+ """
12
+ Return the attention configuration of the model.
13
+ """
14
+ for i in range(self.num_blocks):
15
+ if self.attn_mode == "shift_window":
16
+ yield "serialized", self.window_size, 0, (16 * (i % 2),) * 3, sp.SerializeMode.Z_ORDER
17
+ elif self.attn_mode == "shift_sequence":
18
+ yield "serialized", self.window_size, self.window_size // 2 * (i % 2), (0, 0, 0), sp.SerializeMode.Z_ORDER
19
+ elif self.attn_mode == "shift_order":
20
+ yield "serialized", self.window_size, 0, (0, 0, 0), sp.SerializeModes[i % 4]
21
+ elif self.attn_mode == "full":
22
+ yield "full", None, None, None, None
23
+ elif self.attn_mode == "swin":
24
+ yield "windowed", self.window_size, None, self.window_size // 2 * (i % 2), None
25
+
26
+
27
+ class SparseTransformerBase(nn.Module):
28
+ """
29
+ Sparse Transformer without output layers.
30
+ Serve as the base class for encoder and decoder.
31
+ """
32
+ def __init__(
33
+ self,
34
+ in_channels: int,
35
+ model_channels: int,
36
+ num_blocks: int,
37
+ num_heads: Optional[int] = None,
38
+ num_head_channels: Optional[int] = 64,
39
+ mlp_ratio: float = 4.0,
40
+ attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
41
+ window_size: Optional[int] = None,
42
+ pe_mode: Literal["ape", "rope"] = "ape",
43
+ use_fp16: bool = False,
44
+ use_checkpoint: bool = False,
45
+ qk_rms_norm: bool = False,
46
+ ):
47
+ super().__init__()
48
+ self.in_channels = in_channels
49
+ self.model_channels = model_channels
50
+ self.num_blocks = num_blocks
51
+ self.window_size = window_size
52
+ self.num_heads = num_heads or model_channels // num_head_channels
53
+ self.mlp_ratio = mlp_ratio
54
+ self.attn_mode = attn_mode
55
+ self.pe_mode = pe_mode
56
+ self.use_fp16 = use_fp16
57
+ self.use_checkpoint = use_checkpoint
58
+ self.qk_rms_norm = qk_rms_norm
59
+ self.dtype = torch.float16 if use_fp16 else torch.float32
60
+
61
+ if pe_mode == "ape":
62
+ self.pos_embedder = AbsolutePositionEmbedder(model_channels)
63
+
64
+ self.input_layer = sp.SparseLinear(in_channels, model_channels)
65
+ self.blocks = nn.ModuleList([
66
+ SparseTransformerBlock(
67
+ model_channels,
68
+ num_heads=self.num_heads,
69
+ mlp_ratio=self.mlp_ratio,
70
+ attn_mode=attn_mode,
71
+ window_size=window_size,
72
+ shift_sequence=shift_sequence,
73
+ shift_window=shift_window,
74
+ serialize_mode=serialize_mode,
75
+ use_checkpoint=self.use_checkpoint,
76
+ use_rope=(pe_mode == "rope"),
77
+ qk_rms_norm=self.qk_rms_norm,
78
+ )
79
+ for attn_mode, window_size, shift_sequence, shift_window, serialize_mode in block_attn_config(self)
80
+ ])
81
+
82
+ @property
83
+ def device(self) -> torch.device:
84
+ """
85
+ Return the device of the model.
86
+ """
87
+ return next(self.parameters()).device
88
+
89
+ def convert_to_fp16(self) -> None:
90
+ """
91
+ Convert the torso of the model to float16.
92
+ """
93
+ self.blocks.apply(convert_module_to_f16)
94
+
95
+ def convert_to_fp32(self) -> None:
96
+ """
97
+ Convert the torso of the model to float32.
98
+ """
99
+ self.blocks.apply(convert_module_to_f32)
100
+
101
+ def initialize_weights(self) -> None:
102
+ # Initialize transformer layers:
103
+ def _basic_init(module):
104
+ if isinstance(module, nn.Linear):
105
+ torch.nn.init.xavier_uniform_(module.weight)
106
+ if module.bias is not None:
107
+ nn.init.constant_(module.bias, 0)
108
+ self.apply(_basic_init)
109
+
110
+ def forward(self, x: sp.SparseTensor) -> sp.SparseTensor:
111
+ h = self.input_layer(x)
112
+ if self.pe_mode == "ape":
113
+ h = h + self.pos_embedder(x.coords[:, 1:])
114
+ h = h.type(self.dtype)
115
+ for block in self.blocks:
116
+ h = block(h)
117
+ return h
trellis/models/structured_latent_vae/decoder_gs.py ADDED
@@ -0,0 +1,131 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import *
2
+ import torch
3
+ import torch.nn as nn
4
+ import torch.nn.functional as F
5
+ from ...modules import sparse as sp
6
+ from ...utils.random_utils import hammersley_sequence
7
+ from .base import SparseTransformerBase
8
+ from ...representations import Gaussian
9
+ from ..sparse_elastic_mixin import SparseTransformerElasticMixin
10
+
11
+
12
+ class SLatGaussianDecoder(SparseTransformerBase):
13
+ def __init__(
14
+ self,
15
+ resolution: int,
16
+ model_channels: int,
17
+ latent_channels: int,
18
+ num_blocks: int,
19
+ num_heads: Optional[int] = None,
20
+ num_head_channels: Optional[int] = 64,
21
+ mlp_ratio: float = 4,
22
+ attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
23
+ window_size: int = 8,
24
+ pe_mode: Literal["ape", "rope"] = "ape",
25
+ use_fp16: bool = False,
26
+ use_checkpoint: bool = False,
27
+ qk_rms_norm: bool = False,
28
+ representation_config: dict = None,
29
+ ):
30
+ super().__init__(
31
+ in_channels=latent_channels,
32
+ model_channels=model_channels,
33
+ num_blocks=num_blocks,
34
+ num_heads=num_heads,
35
+ num_head_channels=num_head_channels,
36
+ mlp_ratio=mlp_ratio,
37
+ attn_mode=attn_mode,
38
+ window_size=window_size,
39
+ pe_mode=pe_mode,
40
+ use_fp16=use_fp16,
41
+ use_checkpoint=use_checkpoint,
42
+ qk_rms_norm=qk_rms_norm,
43
+ )
44
+ self.resolution = resolution
45
+ self.rep_config = representation_config
46
+ self._calc_layout()
47
+ self.out_layer = sp.SparseLinear(model_channels, self.out_channels)
48
+ self._build_perturbation()
49
+
50
+ self.initialize_weights()
51
+ if use_fp16:
52
+ self.convert_to_fp16()
53
+
54
+ def initialize_weights(self) -> None:
55
+ super().initialize_weights()
56
+ # Zero-out output layers:
57
+ nn.init.constant_(self.out_layer.weight, 0)
58
+ nn.init.constant_(self.out_layer.bias, 0)
59
+
60
+ def _build_perturbation(self) -> None:
61
+ perturbation = [hammersley_sequence(3, i, self.rep_config['num_gaussians']) for i in range(self.rep_config['num_gaussians'])]
62
+ perturbation = torch.tensor(perturbation).float() * 2 - 1
63
+ perturbation = perturbation / self.rep_config['voxel_size']
64
+ perturbation = torch.atanh(perturbation).to(self.device)
65
+ self.register_buffer('offset_perturbation', perturbation)
66
+
67
+ def _calc_layout(self) -> None:
68
+ self.layout = {
69
+ '_xyz' : {'shape': (self.rep_config['num_gaussians'], 3), 'size': self.rep_config['num_gaussians'] * 3},
70
+ '_features_dc' : {'shape': (self.rep_config['num_gaussians'], 1, 3), 'size': self.rep_config['num_gaussians'] * 3},
71
+ '_scaling' : {'shape': (self.rep_config['num_gaussians'], 3), 'size': self.rep_config['num_gaussians'] * 3},
72
+ '_rotation' : {'shape': (self.rep_config['num_gaussians'], 4), 'size': self.rep_config['num_gaussians'] * 4},
73
+ '_opacity' : {'shape': (self.rep_config['num_gaussians'], 1), 'size': self.rep_config['num_gaussians']},
74
+ }
75
+ start = 0
76
+ for k, v in self.layout.items():
77
+ v['range'] = (start, start + v['size'])
78
+ start += v['size']
79
+ self.out_channels = start
80
+
81
+ def to_representation(self, x: sp.SparseTensor) -> List[Gaussian]:
82
+ """
83
+ Convert a batch of network outputs to 3D representations.
84
+
85
+ Args:
86
+ x: The [N x * x C] sparse tensor output by the network.
87
+
88
+ Returns:
89
+ list of representations
90
+ """
91
+ ret = []
92
+ for i in range(x.shape[0]):
93
+ representation = Gaussian(
94
+ sh_degree=0,
95
+ aabb=[-0.5, -0.5, -0.5, 1.0, 1.0, 1.0],
96
+ mininum_kernel_size = self.rep_config['3d_filter_kernel_size'],
97
+ scaling_bias = self.rep_config['scaling_bias'],
98
+ opacity_bias = self.rep_config['opacity_bias'],
99
+ scaling_activation = self.rep_config['scaling_activation']
100
+ )
101
+ xyz = (x.coords[x.layout[i]][:, 1:].float() + 0.5) / self.resolution
102
+ for k, v in self.layout.items():
103
+ if k == '_xyz':
104
+ offset = x.feats[x.layout[i]][:, v['range'][0]:v['range'][1]].reshape(-1, *v['shape'])
105
+ offset = offset * self.rep_config['lr'][k]
106
+ if self.rep_config['perturb_offset']:
107
+ offset = offset + self.offset_perturbation
108
+ offset = torch.tanh(offset) / self.resolution * 0.5 * self.rep_config['voxel_size']
109
+ _xyz = xyz.unsqueeze(1) + offset
110
+ setattr(representation, k, _xyz.flatten(0, 1))
111
+ else:
112
+ feats = x.feats[x.layout[i]][:, v['range'][0]:v['range'][1]].reshape(-1, *v['shape']).flatten(0, 1)
113
+ feats = feats * self.rep_config['lr'][k]
114
+ setattr(representation, k, feats)
115
+ ret.append(representation)
116
+ return ret
117
+
118
+ def forward(self, x: sp.SparseTensor) -> List[Gaussian]:
119
+ h = super().forward(x)
120
+ h = h.type(x.dtype)
121
+ h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
122
+ h = self.out_layer(h)
123
+ return self.to_representation(h)
124
+
125
+
126
+ class ElasticSLatGaussianDecoder(SparseTransformerElasticMixin, SLatGaussianDecoder):
127
+ """
128
+ Slat VAE Gaussian decoder with elastic memory management.
129
+ Used for training with low VRAM.
130
+ """
131
+ pass
trellis/models/structured_latent_vae/decoder_mesh.py ADDED
@@ -0,0 +1,176 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import *
2
+ import torch
3
+ import torch.nn as nn
4
+ import torch.nn.functional as F
5
+ import numpy as np
6
+ from ...modules.utils import zero_module, convert_module_to_f16, convert_module_to_f32
7
+ from ...modules import sparse as sp
8
+ from .base import SparseTransformerBase
9
+ from ...representations import MeshExtractResult
10
+ from ...representations.mesh import SparseFeatures2Mesh
11
+ from ..sparse_elastic_mixin import SparseTransformerElasticMixin
12
+
13
+
14
+ class SparseSubdivideBlock3d(nn.Module):
15
+ """
16
+ A 3D subdivide block that can subdivide the sparse tensor.
17
+
18
+ Args:
19
+ channels: channels in the inputs and outputs.
20
+ out_channels: if specified, the number of output channels.
21
+ num_groups: the number of groups for the group norm.
22
+ """
23
+ def __init__(
24
+ self,
25
+ channels: int,
26
+ resolution: int,
27
+ out_channels: Optional[int] = None,
28
+ num_groups: int = 32
29
+ ):
30
+ super().__init__()
31
+ self.channels = channels
32
+ self.resolution = resolution
33
+ self.out_resolution = resolution * 2
34
+ self.out_channels = out_channels or channels
35
+
36
+ self.act_layers = nn.Sequential(
37
+ sp.SparseGroupNorm32(num_groups, channels),
38
+ sp.SparseSiLU()
39
+ )
40
+
41
+ self.sub = sp.SparseSubdivide()
42
+
43
+ self.out_layers = nn.Sequential(
44
+ sp.SparseConv3d(channels, self.out_channels, 3, indice_key=f"res_{self.out_resolution}"),
45
+ sp.SparseGroupNorm32(num_groups, self.out_channels),
46
+ sp.SparseSiLU(),
47
+ zero_module(sp.SparseConv3d(self.out_channels, self.out_channels, 3, indice_key=f"res_{self.out_resolution}")),
48
+ )
49
+
50
+ if self.out_channels == channels:
51
+ self.skip_connection = nn.Identity()
52
+ else:
53
+ self.skip_connection = sp.SparseConv3d(channels, self.out_channels, 1, indice_key=f"res_{self.out_resolution}")
54
+
55
+ def forward(self, x: sp.SparseTensor) -> sp.SparseTensor:
56
+ """
57
+ Apply the block to a Tensor, conditioned on a timestep embedding.
58
+
59
+ Args:
60
+ x: an [N x C x ...] Tensor of features.
61
+ Returns:
62
+ an [N x C x ...] Tensor of outputs.
63
+ """
64
+ h = self.act_layers(x)
65
+ h = self.sub(h)
66
+ x = self.sub(x)
67
+ h = self.out_layers(h)
68
+ h = h + self.skip_connection(x)
69
+ return h
70
+
71
+
72
+ class SLatMeshDecoder(SparseTransformerBase):
73
+ def __init__(
74
+ self,
75
+ resolution: int,
76
+ model_channels: int,
77
+ latent_channels: int,
78
+ num_blocks: int,
79
+ num_heads: Optional[int] = None,
80
+ num_head_channels: Optional[int] = 64,
81
+ mlp_ratio: float = 4,
82
+ attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
83
+ window_size: int = 8,
84
+ pe_mode: Literal["ape", "rope"] = "ape",
85
+ use_fp16: bool = False,
86
+ use_checkpoint: bool = False,
87
+ qk_rms_norm: bool = False,
88
+ representation_config: dict = None,
89
+ ):
90
+ super().__init__(
91
+ in_channels=latent_channels,
92
+ model_channels=model_channels,
93
+ num_blocks=num_blocks,
94
+ num_heads=num_heads,
95
+ num_head_channels=num_head_channels,
96
+ mlp_ratio=mlp_ratio,
97
+ attn_mode=attn_mode,
98
+ window_size=window_size,
99
+ pe_mode=pe_mode,
100
+ use_fp16=use_fp16,
101
+ use_checkpoint=use_checkpoint,
102
+ qk_rms_norm=qk_rms_norm,
103
+ )
104
+ self.resolution = resolution
105
+ self.rep_config = representation_config
106
+ self.mesh_extractor = SparseFeatures2Mesh(res=self.resolution*4, use_color=self.rep_config.get('use_color', False))
107
+ self.out_channels = self.mesh_extractor.feats_channels
108
+ self.upsample = nn.ModuleList([
109
+ SparseSubdivideBlock3d(
110
+ channels=model_channels,
111
+ resolution=resolution,
112
+ out_channels=model_channels // 4
113
+ ),
114
+ SparseSubdivideBlock3d(
115
+ channels=model_channels // 4,
116
+ resolution=resolution * 2,
117
+ out_channels=model_channels // 8
118
+ )
119
+ ])
120
+ self.out_layer = sp.SparseLinear(model_channels // 8, self.out_channels)
121
+
122
+ self.initialize_weights()
123
+ if use_fp16:
124
+ self.convert_to_fp16()
125
+
126
+ def initialize_weights(self) -> None:
127
+ super().initialize_weights()
128
+ # Zero-out output layers:
129
+ nn.init.constant_(self.out_layer.weight, 0)
130
+ nn.init.constant_(self.out_layer.bias, 0)
131
+
132
+ def convert_to_fp16(self) -> None:
133
+ """
134
+ Convert the torso of the model to float16.
135
+ """
136
+ super().convert_to_fp16()
137
+ self.upsample.apply(convert_module_to_f16)
138
+
139
+ def convert_to_fp32(self) -> None:
140
+ """
141
+ Convert the torso of the model to float32.
142
+ """
143
+ super().convert_to_fp32()
144
+ self.upsample.apply(convert_module_to_f32)
145
+
146
+ def to_representation(self, x: sp.SparseTensor) -> List[MeshExtractResult]:
147
+ """
148
+ Convert a batch of network outputs to 3D representations.
149
+
150
+ Args:
151
+ x: The [N x * x C] sparse tensor output by the network.
152
+
153
+ Returns:
154
+ list of representations
155
+ """
156
+ ret = []
157
+ for i in range(x.shape[0]):
158
+ mesh = self.mesh_extractor(x[i], training=self.training)
159
+ ret.append(mesh)
160
+ return ret
161
+
162
+ def forward(self, x: sp.SparseTensor) -> List[MeshExtractResult]:
163
+ h = super().forward(x)
164
+ for block in self.upsample:
165
+ h = block(h)
166
+ h = h.type(x.dtype)
167
+ h = self.out_layer(h)
168
+ return self.to_representation(h)
169
+
170
+
171
+ class ElasticSLatMeshDecoder(SparseTransformerElasticMixin, SLatMeshDecoder):
172
+ """
173
+ Slat VAE Mesh decoder with elastic memory management.
174
+ Used for training with low VRAM.
175
+ """
176
+ pass
trellis/models/structured_latent_vae/decoder_rf.py ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import *
2
+ import torch
3
+ import torch.nn as nn
4
+ import torch.nn.functional as F
5
+ import numpy as np
6
+ from ...modules import sparse as sp
7
+ from .base import SparseTransformerBase
8
+ from ...representations import Strivec
9
+ from ..sparse_elastic_mixin import SparseTransformerElasticMixin
10
+
11
+
12
+ class SLatRadianceFieldDecoder(SparseTransformerBase):
13
+ def __init__(
14
+ self,
15
+ resolution: int,
16
+ model_channels: int,
17
+ latent_channels: int,
18
+ num_blocks: int,
19
+ num_heads: Optional[int] = None,
20
+ num_head_channels: Optional[int] = 64,
21
+ mlp_ratio: float = 4,
22
+ attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
23
+ window_size: int = 8,
24
+ pe_mode: Literal["ape", "rope"] = "ape",
25
+ use_fp16: bool = False,
26
+ use_checkpoint: bool = False,
27
+ qk_rms_norm: bool = False,
28
+ representation_config: dict = None,
29
+ ):
30
+ super().__init__(
31
+ in_channels=latent_channels,
32
+ model_channels=model_channels,
33
+ num_blocks=num_blocks,
34
+ num_heads=num_heads,
35
+ num_head_channels=num_head_channels,
36
+ mlp_ratio=mlp_ratio,
37
+ attn_mode=attn_mode,
38
+ window_size=window_size,
39
+ pe_mode=pe_mode,
40
+ use_fp16=use_fp16,
41
+ use_checkpoint=use_checkpoint,
42
+ qk_rms_norm=qk_rms_norm,
43
+ )
44
+ self.resolution = resolution
45
+ self.rep_config = representation_config
46
+ self._calc_layout()
47
+ self.out_layer = sp.SparseLinear(model_channels, self.out_channels)
48
+
49
+ self.initialize_weights()
50
+ if use_fp16:
51
+ self.convert_to_fp16()
52
+
53
+ def initialize_weights(self) -> None:
54
+ super().initialize_weights()
55
+ # Zero-out output layers:
56
+ nn.init.constant_(self.out_layer.weight, 0)
57
+ nn.init.constant_(self.out_layer.bias, 0)
58
+
59
+ def _calc_layout(self) -> None:
60
+ self.layout = {
61
+ 'trivec': {'shape': (self.rep_config['rank'], 3, self.rep_config['dim']), 'size': self.rep_config['rank'] * 3 * self.rep_config['dim']},
62
+ 'density': {'shape': (self.rep_config['rank'],), 'size': self.rep_config['rank']},
63
+ 'features_dc': {'shape': (self.rep_config['rank'], 1, 3), 'size': self.rep_config['rank'] * 3},
64
+ }
65
+ start = 0
66
+ for k, v in self.layout.items():
67
+ v['range'] = (start, start + v['size'])
68
+ start += v['size']
69
+ self.out_channels = start
70
+
71
+ def to_representation(self, x: sp.SparseTensor) -> List[Strivec]:
72
+ """
73
+ Convert a batch of network outputs to 3D representations.
74
+
75
+ Args:
76
+ x: The [N x * x C] sparse tensor output by the network.
77
+
78
+ Returns:
79
+ list of representations
80
+ """
81
+ ret = []
82
+ for i in range(x.shape[0]):
83
+ representation = Strivec(
84
+ sh_degree=0,
85
+ resolution=self.resolution,
86
+ aabb=[-0.5, -0.5, -0.5, 1, 1, 1],
87
+ rank=self.rep_config['rank'],
88
+ dim=self.rep_config['dim'],
89
+ device='cuda',
90
+ )
91
+ representation.density_shift = 0.0
92
+ representation.position = (x.coords[x.layout[i]][:, 1:].float() + 0.5) / self.resolution
93
+ representation.depth = torch.full((representation.position.shape[0], 1), int(np.log2(self.resolution)), dtype=torch.uint8, device='cuda')
94
+ for k, v in self.layout.items():
95
+ setattr(representation, k, x.feats[x.layout[i]][:, v['range'][0]:v['range'][1]].reshape(-1, *v['shape']))
96
+ representation.trivec = representation.trivec + 1
97
+ ret.append(representation)
98
+ return ret
99
+
100
+ def forward(self, x: sp.SparseTensor) -> List[Strivec]:
101
+ h = super().forward(x)
102
+ h = h.type(x.dtype)
103
+ h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
104
+ h = self.out_layer(h)
105
+ return self.to_representation(h)
106
+
107
+
108
+ class ElasticSLatRadianceFieldDecoder(SparseTransformerElasticMixin, SLatRadianceFieldDecoder):
109
+ """
110
+ Slat VAE Radiance Field Decoder with elastic memory management.
111
+ Used for training with low VRAM.
112
+ """
113
+ pass
trellis/models/structured_latent_vae/encoder.py ADDED
@@ -0,0 +1,80 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import *
2
+ import torch
3
+ import torch.nn as nn
4
+ import torch.nn.functional as F
5
+ from ...modules import sparse as sp
6
+ from .base import SparseTransformerBase
7
+ from ..sparse_elastic_mixin import SparseTransformerElasticMixin
8
+
9
+
10
+ class SLatEncoder(SparseTransformerBase):
11
+ def __init__(
12
+ self,
13
+ resolution: int,
14
+ in_channels: int,
15
+ model_channels: int,
16
+ latent_channels: int,
17
+ num_blocks: int,
18
+ num_heads: Optional[int] = None,
19
+ num_head_channels: Optional[int] = 64,
20
+ mlp_ratio: float = 4,
21
+ attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
22
+ window_size: int = 8,
23
+ pe_mode: Literal["ape", "rope"] = "ape",
24
+ use_fp16: bool = False,
25
+ use_checkpoint: bool = False,
26
+ qk_rms_norm: bool = False,
27
+ ):
28
+ super().__init__(
29
+ in_channels=in_channels,
30
+ model_channels=model_channels,
31
+ num_blocks=num_blocks,
32
+ num_heads=num_heads,
33
+ num_head_channels=num_head_channels,
34
+ mlp_ratio=mlp_ratio,
35
+ attn_mode=attn_mode,
36
+ window_size=window_size,
37
+ pe_mode=pe_mode,
38
+ use_fp16=use_fp16,
39
+ use_checkpoint=use_checkpoint,
40
+ qk_rms_norm=qk_rms_norm,
41
+ )
42
+ self.resolution = resolution
43
+ self.out_layer = sp.SparseLinear(model_channels, 2 * latent_channels)
44
+
45
+ self.initialize_weights()
46
+ if use_fp16:
47
+ self.convert_to_fp16()
48
+
49
+ def initialize_weights(self) -> None:
50
+ super().initialize_weights()
51
+ # Zero-out output layers:
52
+ nn.init.constant_(self.out_layer.weight, 0)
53
+ nn.init.constant_(self.out_layer.bias, 0)
54
+
55
+ def forward(self, x: sp.SparseTensor, sample_posterior=True, return_raw=False):
56
+ h = super().forward(x)
57
+ h = h.type(x.dtype)
58
+ h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
59
+ h = self.out_layer(h)
60
+
61
+ # Sample from the posterior distribution
62
+ mean, logvar = h.feats.chunk(2, dim=-1)
63
+ if sample_posterior:
64
+ std = torch.exp(0.5 * logvar)
65
+ z = mean + std * torch.randn_like(std)
66
+ else:
67
+ z = mean
68
+ z = h.replace(z)
69
+
70
+ if return_raw:
71
+ return z, mean, logvar
72
+ else:
73
+ return z
74
+
75
+
76
+ class ElasticSLatEncoder(SparseTransformerElasticMixin, SLatEncoder):
77
+ """
78
+ SLat VAE encoder with elastic memory management.
79
+ Used for training with low VRAM.
80
+ """