Spaces:
Runtime error
Runtime error
File size: 9,839 Bytes
e276be2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 |
import math
from contextlib import nullcontext
from functools import partial
from typing import Dict, List, Optional, Tuple, Union
import kornia
import numpy as np
import torch
import torch.nn as nn
from einops import rearrange, repeat
from omegaconf import ListConfig
from torch.utils.checkpoint import checkpoint
from transformers import (
T5EncoderModel,
T5Tokenizer,
)
from ...util import (
append_dims,
autocast,
count_params,
default,
disabled_train,
expand_dims_like,
instantiate_from_config,
)
class AbstractEmbModel(nn.Module):
def __init__(self):
super().__init__()
self._is_trainable = None
self._ucg_rate = None
self._input_key = None
@property
def is_trainable(self) -> bool:
return self._is_trainable
@property
def ucg_rate(self) -> Union[float, torch.Tensor]:
return self._ucg_rate
@property
def input_key(self) -> str:
return self._input_key
@is_trainable.setter
def is_trainable(self, value: bool):
self._is_trainable = value
@ucg_rate.setter
def ucg_rate(self, value: Union[float, torch.Tensor]):
self._ucg_rate = value
@input_key.setter
def input_key(self, value: str):
self._input_key = value
@is_trainable.deleter
def is_trainable(self):
del self._is_trainable
@ucg_rate.deleter
def ucg_rate(self):
del self._ucg_rate
@input_key.deleter
def input_key(self):
del self._input_key
class GeneralConditioner(nn.Module):
OUTPUT_DIM2KEYS = {2: "vector", 3: "crossattn", 4: "concat", 5: "concat"}
KEY2CATDIM = {"vector": 1, "crossattn": 2, "concat": 1}
def __init__(self, emb_models: Union[List, ListConfig], cor_embs=[], cor_p=[]):
super().__init__()
embedders = []
for n, embconfig in enumerate(emb_models):
embedder = instantiate_from_config(embconfig)
assert isinstance(
embedder, AbstractEmbModel
), f"embedder model {embedder.__class__.__name__} has to inherit from AbstractEmbModel"
embedder.is_trainable = embconfig.get("is_trainable", False)
embedder.ucg_rate = embconfig.get("ucg_rate", 0.0)
if not embedder.is_trainable:
embedder.train = disabled_train
for param in embedder.parameters():
param.requires_grad = False
embedder.eval()
print(
f"Initialized embedder #{n}: {embedder.__class__.__name__} "
f"with {count_params(embedder, False)} params. Trainable: {embedder.is_trainable}"
)
if "input_key" in embconfig:
embedder.input_key = embconfig["input_key"]
elif "input_keys" in embconfig:
embedder.input_keys = embconfig["input_keys"]
else:
raise KeyError(f"need either 'input_key' or 'input_keys' for embedder {embedder.__class__.__name__}")
embedder.legacy_ucg_val = embconfig.get("legacy_ucg_value", None)
if embedder.legacy_ucg_val is not None:
embedder.ucg_prng = np.random.RandomState()
embedders.append(embedder)
self.embedders = nn.ModuleList(embedders)
if len(cor_embs) > 0:
assert len(cor_p) == 2 ** len(cor_embs)
self.cor_embs = cor_embs
self.cor_p = cor_p
def possibly_get_ucg_val(self, embedder: AbstractEmbModel, batch: Dict) -> Dict:
assert embedder.legacy_ucg_val is not None
p = embedder.ucg_rate
val = embedder.legacy_ucg_val
for i in range(len(batch[embedder.input_key])):
if embedder.ucg_prng.choice(2, p=[1 - p, p]):
batch[embedder.input_key][i] = val
return batch
def surely_get_ucg_val(self, embedder: AbstractEmbModel, batch: Dict, cond_or_not) -> Dict:
assert embedder.legacy_ucg_val is not None
val = embedder.legacy_ucg_val
for i in range(len(batch[embedder.input_key])):
if cond_or_not[i]:
batch[embedder.input_key][i] = val
return batch
def get_single_embedding(
self,
embedder,
batch,
output,
cond_or_not: Optional[np.ndarray] = None,
force_zero_embeddings: Optional[List] = None,
):
embedding_context = nullcontext if embedder.is_trainable else torch.no_grad
with embedding_context():
if hasattr(embedder, "input_key") and (embedder.input_key is not None):
if embedder.legacy_ucg_val is not None:
if cond_or_not is None:
batch = self.possibly_get_ucg_val(embedder, batch)
else:
batch = self.surely_get_ucg_val(embedder, batch, cond_or_not)
emb_out = embedder(batch[embedder.input_key])
elif hasattr(embedder, "input_keys"):
emb_out = embedder(*[batch[k] for k in embedder.input_keys])
assert isinstance(
emb_out, (torch.Tensor, list, tuple)
), f"encoder outputs must be tensors or a sequence, but got {type(emb_out)}"
if not isinstance(emb_out, (list, tuple)):
emb_out = [emb_out]
for emb in emb_out:
out_key = self.OUTPUT_DIM2KEYS[emb.dim()]
if embedder.ucg_rate > 0.0 and embedder.legacy_ucg_val is None:
if cond_or_not is None:
emb = (
expand_dims_like(
torch.bernoulli((1.0 - embedder.ucg_rate) * torch.ones(emb.shape[0], device=emb.device)),
emb,
)
* emb
)
else:
emb = (
expand_dims_like(
torch.tensor(1 - cond_or_not, dtype=emb.dtype, device=emb.device),
emb,
)
* emb
)
if hasattr(embedder, "input_key") and embedder.input_key in force_zero_embeddings:
emb = torch.zeros_like(emb)
if out_key in output:
output[out_key] = torch.cat((output[out_key], emb), self.KEY2CATDIM[out_key])
else:
output[out_key] = emb
return output
def forward(self, batch: Dict, force_zero_embeddings: Optional[List] = None) -> Dict:
output = dict()
if force_zero_embeddings is None:
force_zero_embeddings = []
if len(self.cor_embs) > 0:
batch_size = len(batch[list(batch.keys())[0]])
rand_idx = np.random.choice(len(self.cor_p), size=(batch_size,), p=self.cor_p)
for emb_idx in self.cor_embs:
cond_or_not = rand_idx % 2
rand_idx //= 2
output = self.get_single_embedding(
self.embedders[emb_idx],
batch,
output=output,
cond_or_not=cond_or_not,
force_zero_embeddings=force_zero_embeddings,
)
for i, embedder in enumerate(self.embedders):
if i in self.cor_embs:
continue
output = self.get_single_embedding(
embedder, batch, output=output, force_zero_embeddings=force_zero_embeddings
)
return output
def get_unconditional_conditioning(self, batch_c, batch_uc=None, force_uc_zero_embeddings=None):
if force_uc_zero_embeddings is None:
force_uc_zero_embeddings = []
ucg_rates = list()
for embedder in self.embedders:
ucg_rates.append(embedder.ucg_rate)
embedder.ucg_rate = 0.0
cor_embs = self.cor_embs
cor_p = self.cor_p
self.cor_embs = []
self.cor_p = []
c = self(batch_c)
uc = self(batch_c if batch_uc is None else batch_uc, force_uc_zero_embeddings)
for embedder, rate in zip(self.embedders, ucg_rates):
embedder.ucg_rate = rate
self.cor_embs = cor_embs
self.cor_p = cor_p
return c, uc
class FrozenT5Embedder(AbstractEmbModel):
"""Uses the T5 transformer encoder for text"""
def __init__(
self,
model_dir="google/t5-v1_1-xxl",
device="cuda",
max_length=77,
freeze=True,
cache_dir=None,
):
super().__init__()
if model_dir is not "google/t5-v1_1-xxl":
self.tokenizer = T5Tokenizer.from_pretrained(model_dir)
self.transformer = T5EncoderModel.from_pretrained(model_dir)
else:
self.tokenizer = T5Tokenizer.from_pretrained(model_dir, cache_dir=cache_dir)
self.transformer = T5EncoderModel.from_pretrained(model_dir, cache_dir=cache_dir)
self.device = device
self.max_length = max_length
if freeze:
self.freeze()
def freeze(self):
self.transformer = self.transformer.eval()
for param in self.parameters():
param.requires_grad = False
# @autocast
def forward(self, text):
batch_encoding = self.tokenizer(
text,
truncation=True,
max_length=self.max_length,
return_length=True,
return_overflowing_tokens=False,
padding="max_length",
return_tensors="pt",
)
tokens = batch_encoding["input_ids"].to(self.device)
with torch.autocast("cuda", enabled=False):
outputs = self.transformer(input_ids=tokens)
z = outputs.last_hidden_state
return z
def encode(self, text):
return self(text)
|