File size: 13,167 Bytes
0f90f73 |
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 |
import logging
import warnings
import torch
import torch.nn as nn
from dataclasses import dataclass, field
from typing import Optional, Dict, Sequence, Union, List, Tuple, Any
from transformers import (
LlamaForCausalLM,
Blip2PreTrainedModel,
Blip2VisionModel,
Blip2Config,
Blip2QFormerModel,
GenerationConfig,
)
from transformers.utils import ModelOutput
warnings.filterwarnings('ignore')
logger = logging.getLogger(__name__)
@dataclass
class Blip2ForConditionalGenerationModelOutput(ModelOutput):
"""
Class defining the outputs of [`Blip2ForConditionalGeneration`].
Args:
loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided, `torch.FloatTensor` of shape `(1,)`):
Language modeling loss from the language model.
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
Prediction scores of the language modeling head of the language model.
vision_outputs (`BaseModelOutputWithPooling`):
Outputs of the vision encoder.
qformer_outputs (`BaseModelOutputWithPoolingAndCrossAttentions`):
Outputs of the Q-Former (Querying Transformer).
language_model_outputs (`CausalLMOutputWithPast` or `Seq2SeqLMOutput`):
Outputs of the language model.
"""
loss: Optional[Tuple[torch.FloatTensor]] = None
logits: Optional[Tuple[torch.FloatTensor]] = None
vision_outputs: Optional[torch.FloatTensor] = None
qformer_outputs: Optional[Tuple[torch.FloatTensor]] = None
language_model_outputs: Optional[Tuple[torch.FloatTensor]] = None
def to_tuple(self) -> Tuple[Any]:
return tuple(
self[k]
if k not in ["vision_outputs", "qformer_outputs", "language_model_outputs"]
else getattr(self, k).to_tuple()
for k in self.keys()
)
class Blip2LlaMAForConditionalGeneration(Blip2PreTrainedModel):
config_class = Blip2Config
main_input_name = "pixel_values"
def __init__(self, config: Blip2Config):
super().__init__(config)
self.vision_model = Blip2VisionModel(config.vision_config)
self.query_tokens = nn.Parameter(torch.zeros(1, config.num_query_tokens, config.qformer_config.hidden_size))
self.qformer = Blip2QFormerModel(config.qformer_config)
language_model = LlamaForCausalLM(config.text_config)
self.language_model = language_model
self.language_projection = nn.Linear(config.qformer_config.hidden_size, language_model.config.hidden_size)
self.config.hidden_size = config.text_config.hidden_size
self.num_queries = config.num_query_tokens
self.offset = 5
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self):
return self.language_model.get_input_embeddings()
def set_input_embeddings(self, value):
self.language_model.set_input_embeddings(value)
def set_output_embeddings(self, new_embeddings):
self.language_model.set_output_embeddings(new_embeddings)
def get_output_embeddings(self) -> nn.Module:
return self.language_model.get_output_embeddings()
def get_encoder(self):
return self.language_model.get_encoder()
def get_decoder(self):
return self.language_model.get_decoder()
def extract_feature(
self,
pixel_values: torch.FloatTensor,
):
image_embeds = self.vision_model(pixel_values, return_dict=True).last_hidden_state
image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long, device=image_embeds.device)
query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
query_outputs = self.qformer(
query_embeds=query_tokens,
encoder_hidden_states=image_embeds,
encoder_attention_mask=image_attention_mask,
return_dict=True,
)
query_output = query_outputs.last_hidden_state
language_model_inputs = self.language_projection(query_output)
return language_model_inputs
def _tie_weights(self):
if not self.config.use_decoder_only_language_model:
self.language_model.encoder.embed_tokens = self.language_model.shared
self.language_model.decoder.embed_tokens = self.language_model.shared
def _preprocess_accelerate(self):
r"""
Some pre-processing hacks to make the model `accelerate` compatible. Check
https://github.com/huggingface/transformers/pull/21707 for more details.
"""
hf_device_map = self.hf_device_map
if len(hf_device_map) > 1 and "language_model" not in hf_device_map and torch.cuda.device_count() > 1:
# warn users about unexpected behavior when using multi-GPU + BLIP-2 + `accelerate`.
logger.warning(
"The `language_model` is not in the `hf_device_map` dictionary and you are running your script"
" in a multi-GPU environment. this may lead to unexpected behavior when using `accelerate`."
" Please pass a `device_map` that contains `language_model` to remove this warning."
" Please refer to https://github.com/huggingface/blog/blob/main/accelerate-large-models.md for",
" more details on creating a `device_map` for large models.",
)
if hasattr(self.language_model, "_hf_hook"):
self.language_model._hf_hook.io_same_device = True # For `generate` compatibility
def forward(
self,
pixel_values: torch.FloatTensor,
input_ids: torch.FloatTensor,
attention_mask: Optional[torch.LongTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
labels: Optional[torch.LongTensor] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, Blip2ForConditionalGenerationModelOutput]:
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# step 1: forward the images through the vision encoder,
# to get image embeddings of shape (batch_size, seq_len, hidden_size)
vision_outputs = self.vision_model(
pixel_values=pixel_values,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
image_embeds = vision_outputs[0]
# step 2: forward the query tokens through the QFormer, using the image embeddings for cross-attention
image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long, device=image_embeds.device)
query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
query_outputs = self.qformer(
query_embeds=query_tokens,
encoder_hidden_states=image_embeds,
encoder_attention_mask=image_attention_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
query_output = query_outputs[0]
# step 3: use the language model, conditioned on the query outputs and the prompt
language_model_inputs = self.language_projection(query_output)
assert language_model_inputs.shape[1] == self.num_queries
inputs_embeds = self.language_model.get_input_embeddings()(input_ids)
# Human: <img><IMAGE></img>. Give the describe Assistant:
# position of <image>: [offset: offset+num_queries]
inputs_embeds[:, self.offset:self.offset + self.num_queries, :] = language_model_inputs
if attention_mask is None:
attention_mask = torch.ones_like(input_ids)
outputs = self.language_model(
inputs_embeds=inputs_embeds,
attention_mask=attention_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
logits = outputs.logits if return_dict else outputs[0]
loss = None
# we compute the loss here since we need to take into account the sequence length of the query embeds
if labels is not None:
logits = logits[:, -labels.size(1):, :]
# Shift so that tokens < n predict n
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous().to(logits.device).to(torch.long)
# Flatten the tokens
loss_fct = nn.CrossEntropyLoss(reduction="mean")
loss = loss_fct(shift_logits.view(-1, self.config.text_config.vocab_size), shift_labels.view(-1))
if not return_dict:
output = (logits, vision_outputs, query_outputs, outputs)
return ((loss,) + output) if loss is not None else output
return Blip2ForConditionalGenerationModelOutput(
loss=loss,
logits=logits,
vision_outputs=vision_outputs,
qformer_outputs=query_outputs,
language_model_outputs=outputs,
)
@torch.no_grad()
def generate(
self,
pixel_values: torch.FloatTensor,
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
language_model_inputs: Optional[torch.FloatTensor] = None,
generation_config: Optional[GenerationConfig] = None,
**generate_kwargs,
) -> torch.LongTensor:
"""
Overrides `generate` function to be able to use the model as a conditional generator.
Args:
pixel_values (`torch.FloatTensor` of shape (batch_size, num_channels, height, width)):
Input images to be processed.
input_ids (`torch.LongTensor` of shape (batch_size, sequence_length), *optional*):
The sequence used as a prompt for the generation.
attention_mask (`torch.LongTensor` of shape (batch_size, sequence_length), *optional*):
Mask to avoid performing attention on padding token indices
generation_config (`~generation.GenerationConfig`, *optional*):
The generation configuration to be used as base parametrization for the generation call. `**kwargs`
passed to generate matching the attributes of `generation_config` will override them. If
`generation_config` is not provided, the default will be used, which had the following loading
priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
default values, whose documentation should be checked to parameterize generation.
Returns:
captions (list): A list of strings of length batch_size * num_captions.
"""
if hasattr(self, "hf_device_map"):
# preprocess for `accelerate`
self._preprocess_accelerate()
if language_model_inputs is None:
batch_size = pixel_values.shape[0]
image_embeds = self.vision_model(pixel_values, return_dict=True).last_hidden_state
image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long, device=image_embeds.device)
query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
query_outputs = self.qformer(
query_embeds=query_tokens,
encoder_hidden_states=image_embeds,
encoder_attention_mask=image_attention_mask,
return_dict=True,
)
query_output = query_outputs.last_hidden_state
language_model_inputs = self.language_projection(query_output)
assert language_model_inputs.shape[1] == self.num_queries
if input_ids is None:
input_ids = (
torch.LongTensor([[self.config.text_config.bos_token_id]])
.repeat(batch_size, 1)
.to(image_embeds.device)
)
if attention_mask is None:
attention_mask = torch.ones_like(input_ids)
inputs_embeds = self.language_model.get_input_embeddings()(input_ids)
# position of <image>: [offset: offset+num_queries]
inputs_embeds[:, self.offset:self.offset + self.num_queries, :] = language_model_inputs
outputs = self.language_model.generate(
inputs_embeds=inputs_embeds,
attention_mask=attention_mask,
generation_config=generation_config,
**generate_kwargs,
)
return outputs
|