File size: 14,048 Bytes
a43ef32 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 |
"""
Copyright (c) 2023, salesforce.com, inc.
All rights reserved.
SPDX-License-Identifier: BSD-3-Clause
For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause
"""
import logging
import torch
import torch.nn as nn
from torch.cuda.amp import autocast as autocast
from transformers import T5TokenizerFast
from lavis.common.registry import registry
from lavis.models.blip2_models.blip2 import Blip2Base, disabled_train
from lavis.models.blip2_models.modeling_t5 import T5Config, T5ForConditionalGeneration
@registry.register_model("blip2_t5")
class Blip2T5(Blip2Base):
"""
BLIP2 T5 model.
Supported model types:
- pretrain_flant5xl: pretrained model with FlanT5-XL
- pretrain_flant5xl_vitL: pretrained model with FlanT5-XL
- pretrain_flant5xxl: pretrained model with FlanT5-XXL
- caption_coco_flant5xl: fintuned image captioning model with FlanT5-XL
Usage:
>>> from lavis.models import load_model
>>> model = load_model("blip2_t5", "pretrain_flant5xl")
"""
PRETRAINED_MODEL_CONFIG_DICT = {
"pretrain_flant5xl": "configs/models/blip2/blip2_pretrain_flant5xl.yaml",
"pretrain_flant5xl_vitL": "configs/models/blip2/blip2_pretrain_flant5xl_vitL.yaml",
"pretrain_flant5xxl": "configs/models/blip2/blip2_pretrain_flant5xxl.yaml",
"caption_coco_flant5xl": "configs/models/blip2/blip2_caption_flant5xl.yaml",
}
def __init__(
self,
vit_model="eva_clip_g",
img_size=224,
drop_path_rate=0,
use_grad_checkpoint=False,
vit_precision="fp16",
freeze_vit=True,
num_query_token=32,
t5_model="google/flan-t5-xl",
prompt="",
max_txt_len=32,
apply_lemmatizer=False,
):
"""
apply_lemmatizer: when set to True, postprocess predict_answers() result with lemmas.
"""
super().__init__()
self.tokenizer = self.init_tokenizer()
self.visual_encoder, self.ln_vision = self.init_vision_encoder(
vit_model, img_size, drop_path_rate, use_grad_checkpoint, vit_precision
)
if freeze_vit:
for name, param in self.visual_encoder.named_parameters():
param.requires_grad = False
self.visual_encoder = self.visual_encoder.eval()
self.visual_encoder.train = disabled_train
logging.info("freeze vision encoder")
self.Qformer, self.query_tokens = self.init_Qformer(
num_query_token, self.visual_encoder.num_features
)
self.Qformer.cls = None
self.Qformer.bert.embeddings.word_embeddings = None
self.Qformer.bert.embeddings.position_embeddings = None
for layer in self.Qformer.bert.encoder.layer:
layer.output = None
layer.intermediate = None
self.t5_tokenizer = T5TokenizerFast.from_pretrained(t5_model)
t5_config = T5Config.from_pretrained(t5_model)
t5_config.dense_act_fn = "gelu"
self.t5_model = T5ForConditionalGeneration.from_pretrained(
t5_model, config=t5_config
)
for name, param in self.t5_model.named_parameters():
param.requires_grad = False
param.data = param.data.bfloat16()
self.t5_proj = nn.Linear(
self.Qformer.config.hidden_size, self.t5_model.config.hidden_size
)
self.max_txt_len = max_txt_len
self.prompt = prompt
self._apply_lemmatizer = apply_lemmatizer
self._lemmatizer = None
def forward(self, samples):
image = samples["image"]
with self.maybe_autocast():
image_embeds = self.ln_vision(self.visual_encoder(image))
image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(
image.device
)
query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
query_output = self.Qformer.bert(
query_embeds=query_tokens,
encoder_hidden_states=image_embeds,
encoder_attention_mask=image_atts,
return_dict=True,
)
inputs_t5 = self.t5_proj(query_output.last_hidden_state)
atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(image.device)
with self.maybe_autocast(dtype=torch.bfloat16):
input_tokens = self.t5_tokenizer(
samples["text_input"],
padding="longest",
truncation=True,
max_length=self.max_txt_len,
return_tensors="pt",
).to(image.device)
output_tokens = self.t5_tokenizer(
samples["text_output"],
padding="longest",
truncation=True,
max_length=self.max_txt_len,
return_tensors="pt",
).to(image.device)
encoder_atts = torch.cat([atts_t5, input_tokens.attention_mask], dim=1)
targets = output_tokens.input_ids.masked_fill(
output_tokens.input_ids == self.t5_tokenizer.pad_token_id, -100
)
inputs_embeds = self.t5_model.encoder.embed_tokens(input_tokens.input_ids)
inputs_embeds = torch.cat([inputs_t5, inputs_embeds], dim=1)
outputs = self.t5_model(
inputs_embeds=inputs_embeds,
attention_mask=encoder_atts,
decoder_attention_mask=output_tokens.attention_mask,
return_dict=True,
labels=targets,
)
loss = outputs.loss
return {"loss": loss}
@torch.no_grad()
def generate(
self,
samples,
use_nucleus_sampling=False,
num_beams=5,
max_length=30,
min_length=1,
top_p=0.9,
repetition_penalty=1.0,
length_penalty=1.0,
num_captions=1,
temperature=1,
):
"""
Args:
samples (dict): A dictionary containing the following keys:
- image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)
use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.
num_beams (int): Number of beams for beam search. 1 means no beam search.
max_length (int): The maximum length of the sequence to be generated.
min_length (int): The minimum length of the sequence to be generated.
top_p (float): The cumulative probability for nucleus sampling.
repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.
num_captions (int): Number of captions to be generated for each image.
Returns:
captions (list): A list of strings of length batch_size * num_captions.
"""
image = samples["image"]
with self.maybe_autocast():
image_embeds = self.ln_vision(self.visual_encoder(image))
image_embeds = image_embeds.float()
image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(
image.device
)
query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
query_output = self.Qformer.bert(
query_embeds=query_tokens,
encoder_hidden_states=image_embeds,
encoder_attention_mask=image_atts,
return_dict=True,
)
inputs_t5 = self.t5_proj(query_output.last_hidden_state)
atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(image.device)
if "prompt" in samples.keys():
prompt = samples["prompt"]
else:
prompt = self.prompt
if isinstance(prompt, str):
prompt = [prompt] * image.size(0)
else:
assert len(prompt) == image.size(
0
), "The number of prompts must be equal to the batch size."
input_tokens = self.t5_tokenizer(
prompt, padding="longest", return_tensors="pt"
).to(image.device)
encoder_atts = torch.cat([atts_t5, input_tokens.attention_mask], dim=1)
with self.maybe_autocast(dtype=torch.bfloat16):
inputs_embeds = self.t5_model.encoder.embed_tokens(input_tokens.input_ids)
inputs_embeds = torch.cat([inputs_t5, inputs_embeds], dim=1)
outputs = self.t5_model.generate(
inputs_embeds=inputs_embeds,
attention_mask=encoder_atts,
do_sample=use_nucleus_sampling,
top_p=top_p,
temperature=temperature,
num_beams=num_beams,
max_new_tokens=max_length,
min_length=min_length,
repetition_penalty=repetition_penalty,
length_penalty=length_penalty,
num_return_sequences=num_captions,
)
output_text = self.t5_tokenizer.batch_decode(
outputs, skip_special_tokens=True
)
return output_text
def predict_answers(
self,
samples,
num_beams=5,
inference_method="generate",
max_len=10,
min_len=1,
num_ans_candidates=128,
answer_list=None,
prompt="",
length_penalty=-1,
**kwargs
):
image = samples["image"]
with self.maybe_autocast():
image_embeds = self.ln_vision(self.visual_encoder(image))
image_embeds = image_embeds.float()
image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(
image.device
)
query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
query_output = self.Qformer.bert(
query_embeds=query_tokens,
encoder_hidden_states=image_embeds,
encoder_attention_mask=image_atts,
return_dict=True,
)
inputs_t5 = self.t5_proj(query_output.last_hidden_state)
atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(image.device)
if isinstance(samples["text_input"], str):
samples["text_input"] = [samples["text_input"]]
if prompt:
text_input = [prompt.format(question) for question in samples["text_input"]]
else:
text_input = samples["text_input"]
input_tokens = self.t5_tokenizer(
text_input, padding="longest", return_tensors="pt"
).to(image.device)
encoder_atts = torch.cat([atts_t5, input_tokens.attention_mask], dim=1)
with self.maybe_autocast(dtype=torch.bfloat16):
inputs_embeds = self.t5_model.encoder.embed_tokens(input_tokens.input_ids)
inputs_embeds = torch.cat([inputs_t5, inputs_embeds], dim=1)
outputs = self.t5_model.generate(
inputs_embeds=inputs_embeds,
attention_mask=encoder_atts,
do_sample=False,
num_beams=num_beams,
max_new_tokens=max_len,
min_length=min_len,
length_penalty=length_penalty,
)
output_text = self.t5_tokenizer.batch_decode(
outputs, skip_special_tokens=True
)
if self._apply_lemmatizer:
output_text = self._lemmatize(output_text)
return output_text
def _lemmatize(self, answers):
def apply(answer):
doc = self.lemmatizer(answer)
words = []
for token in doc:
if token.pos_ in ["NOUN", "VERB"]:
words.append(token.lemma_)
else:
words.append(token.text)
answer = " ".join(words)
return answer
return [apply(answer) for answer in answers]
@property
def lemmatizer(self):
if self._lemmatizer is None:
try:
import spacy
self._lemmatizer = spacy.load("en_core_web_sm")
except ImportError:
logging.error(
"""
Please install spacy and en_core_web_sm model to apply lemmatization.
python -m spacy download en_core_web_sm
OR
import spacy.cli
spacy.cli.download("en_core_web_sm")
"""
)
exit(1)
return self._lemmatizer
@classmethod
def from_config(cls, cfg):
vit_model = cfg.get("vit_model", "eva_clip_g")
img_size = cfg.get("image_size")
num_query_token = cfg.get("num_query_token")
t5_model = cfg.get("t5_model")
drop_path_rate = cfg.get("drop_path_rate", 0)
use_grad_checkpoint = cfg.get("use_grad_checkpoint", False)
vit_precision = cfg.get("vit_precision", "fp16")
freeze_vit = cfg.get("freeze_vit", True)
prompt = cfg.get("prompt", "")
max_txt_len = cfg.get("max_txt_len", 32)
apply_lemmatizer = cfg.get("apply_lemmatizer", False)
model = cls(
vit_model=vit_model,
img_size=img_size,
drop_path_rate=drop_path_rate,
use_grad_checkpoint=use_grad_checkpoint,
vit_precision=vit_precision,
freeze_vit=freeze_vit,
num_query_token=num_query_token,
t5_model=t5_model,
prompt=prompt,
max_txt_len=max_txt_len,
apply_lemmatizer=apply_lemmatizer,
)
model.load_checkpoint_from_config(cfg)
return model
|