File size: 15,432 Bytes
9cabbed |
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 |
import torch
from transformers import (
AutoTokenizer, AutoModel,
BertTokenizer, BertModel,
CLIPTokenizer, CLIPTextModel
)
import torch.nn as nn
import pytorch_lightning as pl
from typing import List
from abc import ABC, abstractmethod
import random
import os
os.environ["TOKENIZERS_PARALLELISM"] = "false"
def taglist_index_to_sentence(taglist_vocab, tag_vocab, taglist_indices, subsample: bool = True):
"""
Convert a tensor or list of taglist indices to a list of tag sentences.
Optionally, randomly shuffle and sample a subset of tags for each sentence.
Args:
taglist_vocab: List of tuples of tag IDs.
tag_vocab: Dictionary mapping tag ID to tag string.
taglist_indices: Tensor or list of indices into taglist_vocab.
seed: Random seed for reproducibility.
subsample: If True, randomly subsample tags in each sentence.
Returns:
tag_sentences: List of strings (tag sentences).
"""
if isinstance(taglist_indices, torch.Tensor):
taglist_indices = taglist_indices.view(-1).tolist()
tag_sentences = []
for idx in taglist_indices:
tag_ids = taglist_vocab[idx]
tags = [tag_vocab[tid].lower().replace('=', ' ') for tid in tag_ids]
if subsample and len(tags) > 1:
n_sample = random.randint(1, len(tags)) # Choose how many tags to keep
tags = random.sample(tags, n_sample) # Sample without replacement
random.shuffle(tags) # Randomize order
sentence = ' '.join(tags)
tag_sentences.append(sentence)
return tag_sentences
def average_pool(last_hidden_states, attention_mask):
masked_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)
return masked_hidden.sum(dim=1) / attention_mask.sum(dim=1, keepdim=True)
class BaseTextEncoder(nn.Module, ABC):
def __init__(self, model_name: str):
super().__init__()
self.model_name = model_name
self.tokenizer = None
self.model = None
self.embedding_dim = None
@abstractmethod
def encode(self, sentences: List[str], device: str = 'cpu') -> torch.Tensor:
"""
Encode a list of sentences into a tensor of embeddings.
Must be implemented by subclasses.
"""
pass
class BertTextEncoder(BaseTextEncoder):
def __init__(self, model_name='bert-base-uncased'):
super().__init__(model_name)
self.tokenizer = BertTokenizer.from_pretrained(model_name)
self.model = BertModel.from_pretrained(model_name)
self.embedding_dim = self.model.config.hidden_size
def encode(self, sentences, device='cpu'):
self.model.to(device)
inputs = self.tokenizer(sentences, return_tensors='pt', padding=True, truncation=True).to(device)
return self.model(**inputs).pooler_output
class CLIPTextEncoder(BaseTextEncoder):
def __init__(self, model_name='openai/clip-vit-large-patch14', local_tokenizer_path=None):
super().__init__(model_name)
local_tokenizer_path = "/u/cherd/.cache/huggingface/hub/models--openai--clip-vit-large-patch14/snapshots/32bd64288804d66eefd0ccbe215aa642df71cc41"
if local_tokenizer_path is not None:
self.tokenizer = CLIPTokenizer.from_pretrained(local_tokenizer_path)
self.model = CLIPTextModel.from_pretrained(local_tokenizer_path)
else:
self.tokenizer = CLIPTokenizer.from_pretrained(model_name)
self.model = CLIPTextModel.from_pretrained(model_name, from_flax=True)
self.embedding_dim = self.model.config.hidden_size
def encode(self, sentences, device='cpu'):
self.model.to(device)
inputs = self.tokenizer(sentences, return_tensors='pt', padding=True, truncation=True).to(device)
input_ids = inputs['input_ids']
eos_token_id = self.tokenizer.eos_token_id
pad_token_id = self.tokenizer.pad_token_id
outputs = self.model(**inputs)
last_hidden = outputs.last_hidden_state # [B, T, D]
batch_size = input_ids.size(0)
embeddings = []
for i in range(batch_size):
input_seq = input_ids[i]
eos_positions = (input_seq == eos_token_id).nonzero(as_tuple=True)[0]
if len(eos_positions) > 0:
eos_idx = eos_positions[-1] # take last EOS (safe for duplicates)
else:
eos_idx = (input_seq != pad_token_id).sum() - 1 # fallback to last non-padding token
embeddings.append(last_hidden[i, eos_idx, :])
return torch.stack(embeddings)
class E5TextEncoder(BaseTextEncoder):
def __init__(self, model_name='intfloat/e5-base'):
super().__init__(model_name)
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
self.model = AutoModel.from_pretrained(model_name)
self.model.pooler = None
self.embedding_dim = self.model.config.hidden_size
def encode(self, sentences, device='cpu'):
self.model.to(device)
sentences = [f"query: {s}" for s in sentences] # official prompt for e5 (for features as per documentation)
inputs = self.tokenizer(sentences, return_tensors='pt', padding=True, truncation=True).to(device)
outputs = self.model(**inputs)
return average_pool(outputs.last_hidden_state, inputs['attention_mask'])
class GritLMTextEncoder(BaseTextEncoder):
def __init__(self, model_name='nomic-ai/nomic-bert-base-punc'):
super().__init__(model_name)
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
self.model = AutoModel.from_pretrained(model_name)
self.embedding_dim = self.model.config.hidden_size
self.proj_head = nn.Linear(self.embedding_dim, 768) # to match other encoders
def encode(self, sentences, device='cpu'):
self.model.to(device)
inputs = self.tokenizer(sentences, return_tensors='pt', padding=True, truncation=True).to(device)
outputs = self.model(**inputs)
pooled = average_pool(outputs.last_hidden_state, inputs['attention_mask'])
return self.proj_head(pooled)
class TextEncoder(pl.LightningModule):
def __init__(self, taglist_vocab: List[tuple], tag_vocab: dict, model_name='bert'):
super().__init__()
self.taglist_vocab = taglist_vocab
self.tag_vocab = tag_vocab
model_name = model_name.lower()
encoder_map = {
'bert': lambda: BertTextEncoder('bert-base-uncased'),
'clip': lambda: CLIPTextEncoder('openai/clip-vit-large-patch14'),
'e5': lambda: E5TextEncoder('intfloat/e5-base'),
'gritlm': lambda: GritLMTextEncoder('nomic-ai/nomic-bert-base-punc')
}
if model_name not in encoder_map:
raise ValueError(f"Unsupported model_name: {model_name}. Choose from {list(encoder_map.keys())}")
print(f"Text backbone: {model_name}")
self.encoder = encoder_map[model_name]() # Instantiate the selected encoder
# self.embedding_dim = 768
def forward(self, taglist_tensor: torch.Tensor) -> torch.Tensor:
tag_indices = taglist_tensor.tolist()
tag_sentences = taglist_index_to_sentence(self.taglist_vocab, self.tag_vocab, tag_indices, subsample=True) # randomize subsampling tags
embeddings = self.encoder.encode(tag_sentences, device=self.device)
return embeddings
def encode_raw_text(self, raw_text: str) -> torch.Tensor:
"""
Encode a single raw string into an embedding for queries
"""
return self.encoder.encode([raw_text], device=self.device)[0]
def encode_batch(self, raw_texts: List[str]) -> torch.Tensor:
"""
Encode a batch of raw strings into embeddings for queries
"""
return self.encoder.encode(raw_texts, device=self.device)
# import torch
# from transformers import (
# AutoTokenizer, AutoModel,
# BertTokenizer, BertModel,
# CLIPTokenizer, CLIPTextModel
# )
# import torch.nn as nn
# import pytorch_lightning as pl
# from typing import List
# from abc import ABC, abstractmethod
# import random
# import os
# os.environ["TOKENIZERS_PARALLELISM"] = "false"
# def taglist_index_to_sentence(taglist_vocab, tag_vocab, taglist_indices, subsample: bool = True):
# """
# Convert a tensor or list of taglist indices to a list of tag sentences.
# Optionally, randomly shuffle and sample a subset of tags for each sentence.
# Args:
# taglist_vocab: List of tuples of tag IDs.
# tag_vocab: Dictionary mapping tag ID to tag string.
# taglist_indices: Tensor or list of indices into taglist_vocab.
# seed: Random seed for reproducibility.
# subsample: If True, randomly subsample tags in each sentence.
# Returns:
# tag_sentences: List of strings (tag sentences).
# """
# if isinstance(taglist_indices, torch.Tensor):
# taglist_indices = taglist_indices.view(-1).tolist()
# tag_sentences = []
# for idx in taglist_indices:
# tag_ids = taglist_vocab[idx]
# tags = [tag_vocab[tid].lower().replace('=', ' ') for tid in tag_ids]
# if subsample and len(tags) > 1:
# n_sample = random.randint(1, len(tags)) # Choose how many tags to keep
# tags = random.sample(tags, n_sample) # Sample without replacement
# random.shuffle(tags) # Randomize order
# sentence = ' '.join(tags)
# tag_sentences.append(sentence)
# return tag_sentences
# def average_pool(last_hidden_states, attention_mask):
# masked_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)
# return masked_hidden.sum(dim=1) / attention_mask.sum(dim=1, keepdim=True)
# class BaseTextEncoder(nn.Module, ABC):
# def __init__(self, model_name: str):
# super().__init__()
# self.model_name = model_name
# self.tokenizer = None
# self.model = None
# self.embedding_dim = None
# @abstractmethod
# def encode(self, sentences: List[str], device: str = 'cpu') -> torch.Tensor:
# """
# Encode a list of sentences into a tensor of embeddings.
# Must be implemented by subclasses.
# """
# pass
# class BertTextEncoder(BaseTextEncoder):
# def __init__(self, model_name='bert-base-uncased'):
# super().__init__(model_name)
# self.tokenizer = BertTokenizer.from_pretrained(model_name)
# self.model = BertModel.from_pretrained(model_name)
# self.embedding_dim = self.model.config.hidden_size
# def encode(self, sentences, device='cpu'):
# self.model.to(device)
# inputs = self.tokenizer(sentences, return_tensors='pt', padding=True, truncation=True).to(device)
# return self.model(**inputs).pooler_output
# class CLIPTextEncoder(BaseTextEncoder):
# def __init__(self, model_name='openai/clip-vit-large-patch14'):
# super().__init__(model_name)
# self.tokenizer = CLIPTokenizer.from_pretrained(model_name)
# self.model = CLIPTextModel.from_pretrained(model_name)
# self.embedding_dim = self.model.config.hidden_size
# def encode(self, sentences, device='cpu'):
# self.model.to(device)
# inputs = self.tokenizer(sentences, return_tensors='pt', padding=True, truncation=True).to(device)
# input_ids = inputs['input_ids']
# eos_token_id = self.tokenizer.eos_token_id
# pad_token_id = self.tokenizer.pad_token_id
# outputs = self.model(**inputs)
# last_hidden = outputs.last_hidden_state # [B, T, D]
# batch_size = input_ids.size(0)
# embeddings = []
# for i in range(batch_size):
# input_seq = input_ids[i]
# eos_positions = (input_seq == eos_token_id).nonzero(as_tuple=True)[0]
# if len(eos_positions) > 0:
# eos_idx = eos_positions[-1] # take last EOS (safe for duplicates)
# else:
# eos_idx = (input_seq != pad_token_id).sum() - 1 # fallback to last non-padding token
# embeddings.append(last_hidden[i, eos_idx, :])
# return torch.stack(embeddings)
# class E5TextEncoder(BaseTextEncoder):
# def __init__(self, model_name='intfloat/e5-base'):
# super().__init__(model_name)
# self.tokenizer = AutoTokenizer.from_pretrained(model_name)
# self.model = AutoModel.from_pretrained(model_name)
# self.model.pooler = None
# self.embedding_dim = self.model.config.hidden_size
# def encode(self, sentences, device='cpu'):
# self.model.to(device)
# sentences = [f"query: {s}" for s in sentences] # official prompt for e5 (for features as per documentation)
# inputs = self.tokenizer(sentences, return_tensors='pt', padding=True, truncation=True).to(device)
# outputs = self.model(**inputs)
# return average_pool(outputs.last_hidden_state, inputs['attention_mask'])
# class GritLMTextEncoder(BaseTextEncoder):
# def __init__(self, model_name='nomic-ai/nomic-bert-base-punc'):
# super().__init__(model_name)
# self.tokenizer = AutoTokenizer.from_pretrained(model_name)
# self.model = AutoModel.from_pretrained(model_name)
# self.embedding_dim = self.model.config.hidden_size
# self.proj_head = nn.Linear(self.embedding_dim, 768) # to match other encoders
# def encode(self, sentences, device='cpu'):
# self.model.to(device)
# inputs = self.tokenizer(sentences, return_tensors='pt', padding=True, truncation=True).to(device)
# outputs = self.model(**inputs)
# pooled = average_pool(outputs.last_hidden_state, inputs['attention_mask'])
# return self.proj_head(pooled)
# class TextEncoder(pl.LightningModule):
# def __init__(self, taglist_vocab: List[tuple], tag_vocab: dict, model_name='bert'):
# super().__init__()
# self.taglist_vocab = taglist_vocab
# self.tag_vocab = tag_vocab
# model_name = model_name.lower()
# encoder_map = {
# 'bert': lambda: BertTextEncoder('bert-base-uncased'),
# 'clip': lambda: CLIPTextEncoder('openai/clip-vit-large-patch14'),
# 'e5': lambda: E5TextEncoder('intfloat/e5-base'),
# 'gritlm': lambda: GritLMTextEncoder('nomic-ai/nomic-bert-base-punc')
# }
# if model_name not in encoder_map:
# raise ValueError(f"Unsupported model_name: {model_name}. Choose from {list(encoder_map.keys())}")
# print(f"Text backbone: {model_name}")
# self.encoder = encoder_map[model_name]() # Instantiate the selected encoder
# # self.embedding_dim = 768
# def forward(self, taglist_tensor: torch.Tensor) -> torch.Tensor:
# tag_indices = taglist_tensor.tolist()
# tag_sentences = taglist_index_to_sentence(self.taglist_vocab, self.tag_vocab, tag_indices, subsample=True) # randomize subsampling tags
# embeddings = self.encoder.encode(tag_sentences, device=self.device)
# return embeddings
# def encode_raw_text(self, raw_text: str) -> torch.Tensor:
# """
# Encode a single raw string into an embedding for queries
# """
# return self.encoder.encode([raw_text], device=self.device)[0] |