File size: 11,822 Bytes
5110eb7 |
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 |
from med import BertConfig, BertModel, BertLMHeadModel
from vit import VisionTransformer, interpolate_pos_embed
import torch
from torch import nn
import torch.nn.functional as F
from transformers import BertTokenizer
import numpy as np
from pathlib import Path
from urllib.parse import urlparse
from timm.models.hub import download_cached_file
import os
# General helpers
def is_url(url_or_filename):
parsed = urlparse(url_or_filename)
return parsed.scheme in ("http", "https")
def load_checkpoint(model,url_or_filename):
if is_url(url_or_filename):
cached_file = download_cached_file(url_or_filename, check_hash=False, progress=True)
checkpoint = torch.load(cached_file, map_location='cpu')
elif os.path.isfile(url_or_filename):
checkpoint = torch.load(url_or_filename, map_location='cpu')
else:
raise RuntimeError('checkpoint url or path is invalid')
state_dict = checkpoint['model']
state_dict['visual_encoder.pos_embed'] = interpolate_pos_embed(state_dict['visual_encoder.pos_embed'],model.visual_encoder)
if 'visual_encoder_m.pos_embed' in model.state_dict().keys():
state_dict['visual_encoder_m.pos_embed'] = interpolate_pos_embed(state_dict['visual_encoder_m.pos_embed'],
model.visual_encoder_m)
for key in model.state_dict().keys():
if key in state_dict.keys():
if state_dict[key].shape!=model.state_dict()[key].shape:
del state_dict[key]
msg = model.load_state_dict(state_dict,strict=False)
print('load checkpoint from %s'%url_or_filename)
return model,msg
def init_tokenizer():
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
tokenizer.add_special_tokens({'bos_token':'[DEC]'})
tokenizer.add_special_tokens({'additional_special_tokens':['[ENC]']})
tokenizer.enc_token_id = tokenizer.additional_special_tokens_ids[0]
return tokenizer
def create_vit(vit, image_size, use_grad_checkpointing=False, ckpt_layer=0, drop_path_rate=0):
assert vit in ['base', 'large'], "vit parameter must be base or large"
if vit=='base':
vision_width = 768
visual_encoder = VisionTransformer(img_size=image_size, patch_size=16, embed_dim=vision_width, depth=12,
num_heads=12, use_grad_checkpointing=use_grad_checkpointing, ckpt_layer=ckpt_layer,
drop_path_rate=0 or drop_path_rate
)
elif vit=='large':
vision_width = 1024
visual_encoder = VisionTransformer(img_size=image_size, patch_size=16, embed_dim=vision_width, depth=24,
num_heads=16, use_grad_checkpointing=use_grad_checkpointing, ckpt_layer=ckpt_layer,
drop_path_rate=0.1 or drop_path_rate
)
return visual_encoder, vision_width
class BLIP_VQA(nn.Module):
def __init__(self,
med_config = str(Path(__file__).parent / 'configs/med_config.json'),
image_size = 480,
vit = 'base',
vit_grad_ckpt = False,
vit_ckpt_layer = 0,
):
"""
Args:
med_config (str): path for the mixture of encoder-decoder model's configuration file
image_size (int): input image size
vit (str): model size of vision transformer
"""
super().__init__()
self.visual_encoder, vision_width = create_vit(vit, image_size, vit_grad_ckpt, vit_ckpt_layer, drop_path_rate=0.1)
self.tokenizer = init_tokenizer()
encoder_config = BertConfig.from_json_file(med_config)
encoder_config.encoder_width = vision_width
self.text_encoder = BertModel(config=encoder_config, add_pooling_layer=False)
decoder_config = BertConfig.from_json_file(med_config)
self.text_decoder = BertLMHeadModel(config=decoder_config)
def forward(self, image, question, answer=None, n=None, weights=None, train=True, inference='rank', k_test=128):
image_embeds = self.visual_encoder(image)
image_atts = torch.ones(image_embeds.size()[:-1],dtype=torch.long).to(image.device)
question = self.tokenizer(question, padding='longest', truncation=True, max_length=35,
return_tensors="pt").to(image.device)
question.input_ids[:,0] = self.tokenizer.enc_token_id
if train:
'''
n: number of answers for each question
weights: weight for each answer
'''
answer = self.tokenizer(answer, padding='longest', return_tensors="pt").to(image.device)
answer.input_ids[:,0] = self.tokenizer.bos_token_id
answer_targets = answer.input_ids.masked_fill(answer.input_ids == self.tokenizer.pad_token_id, -100)
question_output = self.text_encoder(question.input_ids,
attention_mask = question.attention_mask,
encoder_hidden_states = image_embeds,
encoder_attention_mask = image_atts,
return_dict = True)
question_states = []
question_atts = []
for b, n in enumerate(n):
question_states += [question_output.last_hidden_state[b]]*n
question_atts += [question.attention_mask[b]]*n
question_states = torch.stack(question_states,0)
question_atts = torch.stack(question_atts,0)
answer_output = self.text_decoder(answer.input_ids,
attention_mask = answer.attention_mask,
encoder_hidden_states = question_states,
encoder_attention_mask = question_atts,
labels = answer_targets,
return_dict = True,
reduction = 'none',
)
loss = weights * answer_output.loss
loss = loss.sum()/image.size(0)
return loss
else:
question_output = self.text_encoder(question.input_ids,
attention_mask = question.attention_mask,
encoder_hidden_states = image_embeds,
encoder_attention_mask = image_atts,
return_dict = True)
if inference=='generate':
num_beams = 3
question_states = question_output.last_hidden_state.repeat_interleave(num_beams,dim=0)
question_atts = torch.ones(question_states.size()[:-1],dtype=torch.long).to(question_states.device)
model_kwargs = {"encoder_hidden_states": question_states, "encoder_attention_mask":question_atts}
bos_ids = torch.full((image.size(0),1),fill_value=self.tokenizer.bos_token_id,device=image.device)
outputs = self.text_decoder.generate(input_ids=bos_ids,
max_length=10,
min_length=1,
num_beams=num_beams,
eos_token_id=self.tokenizer.sep_token_id,
pad_token_id=self.tokenizer.pad_token_id,
**model_kwargs)
answers = []
for output in outputs:
answer = self.tokenizer.decode(output, skip_special_tokens=True)
answers.append(answer)
return answers
elif inference=='rank':
max_ids = self.rank_answer(question_output.last_hidden_state, question.attention_mask,
answer.input_ids, answer.attention_mask, k_test)
return max_ids
def rank_answer(self, question_states, question_atts, answer_ids, answer_atts, k):
num_ques = question_states.size(0)
start_ids = answer_ids[0,0].repeat(num_ques,1) # bos token
start_output = self.text_decoder(start_ids,
encoder_hidden_states = question_states,
encoder_attention_mask = question_atts,
return_dict = True,
reduction = 'none')
logits = start_output.logits[:,0,:] # first token's logit
# topk_probs: top-k probability
# topk_ids: [num_question, k]
answer_first_token = answer_ids[:,1]
prob_first_token = F.softmax(logits,dim=1).index_select(dim=1, index=answer_first_token)
topk_probs, topk_ids = prob_first_token.topk(k,dim=1)
# answer input: [num_question*k, answer_len]
input_ids = []
input_atts = []
for b, topk_id in enumerate(topk_ids):
input_ids.append(answer_ids.index_select(dim=0, index=topk_id))
input_atts.append(answer_atts.index_select(dim=0, index=topk_id))
input_ids = torch.cat(input_ids,dim=0)
input_atts = torch.cat(input_atts,dim=0)
targets_ids = input_ids.masked_fill(input_ids == self.tokenizer.pad_token_id, -100)
# repeat encoder's output for top-k answers
question_states = tile(question_states, 0, k)
question_atts = tile(question_atts, 0, k)
output = self.text_decoder(input_ids,
attention_mask = input_atts,
encoder_hidden_states = question_states,
encoder_attention_mask = question_atts,
labels = targets_ids,
return_dict = True,
reduction = 'none')
log_probs_sum = -output.loss
log_probs_sum = log_probs_sum.view(num_ques,k)
max_topk_ids = log_probs_sum.argmax(dim=1)
max_ids = topk_ids[max_topk_ids>=0,max_topk_ids]
return max_ids
def blip_vqa(pretrained='',**kwargs):
model = BLIP_VQA(**kwargs)
if pretrained:
model,msg = load_checkpoint(model,pretrained)
# assert(len(msg.missing_keys)==0)
return model
def tile(x, dim, n_tile):
init_dim = x.size(dim)
repeat_idx = [1] * x.dim()
repeat_idx[dim] = n_tile
x = x.repeat(*(repeat_idx))
order_index = torch.LongTensor(np.concatenate([init_dim * np.arange(n_tile) + i for i in range(init_dim)]))
return torch.index_select(x, dim, order_index.to(x.device))
|