-
Notifications
You must be signed in to change notification settings - Fork 15
/
xvlm_models.py
267 lines (220 loc) · 13.8 KB
/
xvlm_models.py
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
import os
import torch
import yaml
import subprocess
import numpy as np
from tqdm import tqdm
import torch.nn.functional as F
from .xvlm_utils.xvlm import XVLM
from .xvlm_utils.tokenization_bert import BertTokenizer
from .xvlm_utils.tokenization_roberta import RobertaTokenizer
from .blip_utils.utils import MetricLogger
# All of the below URLs are taken from, and most of the implementation are heavily inspired from the wonderful https://github.com/salesforce/BLIP repo.
download_urls = {
"xvlm-flickr" : {
"model_url": "1vhdtH3iFaoZuMqOGm-8YM-diPWVfRJzv",
"vision_config_url": "https://github.com/zengyan-97/X-VLM/raw/e7b960256d194952321b5adad39770c03e6ce9c2/configs/config_swinB_384.json",
"config_url": "13-GCckeAh7QUeFVGwye7qLJamwl_hXdf",
"bert_config_url": "https://github.com/zengyan-97/X-VLM/raw/e7b960256d194952321b5adad39770c03e6ce9c2/configs/config_bert.json"
},
"xvlm-coco" : {
"model_url": "1bv6_pZOsXW53EhlwU0ZgSk03uzFI61pN",
"vision_config_url": "https://github.com/zengyan-97/X-VLM/raw/e7b960256d194952321b5adad39770c03e6ce9c2/configs/config_swinB_384.json",
"config_url": "11pdOukGXZzmPubvjLhJ2Sr1BIBRTEM-P",
"bert_config_url": "https://github.com/zengyan-97/X-VLM/raw/e7b960256d194952321b5adad39770c03e6ce9c2/configs/config_bert.json"
},
}
class XVLMWrapper:
def __init__(self, root_dir, device, variant="xvlm-flickr"):
self.variant = variant
self.root_dir = root_dir
self.config_path = os.path.join(root_dir, f"{self.variant}-config")
self.model_path = os.path.join(root_dir, f"{self.variant}.pth")
self.bert_config_path = os.path.join(root_dir, "configs", download_urls[variant]['bert_config_url'].split("/")[-1])
self.vision_config_path = os.path.join(root_dir, "configs", download_urls[variant]['vision_config_url'].split("/")[-1])
if not (os.path.exists(self.config_path) and os.path.exists(self.model_path) and os.path.exists(self.bert_config_path) and os.path.exists(self.vision_config_path)):
self.download()
self.config = yaml.load(open(self.config_path, 'r'), Loader=yaml.Loader)
self.config["vision_config"] = self.vision_config_path
self.config["text_config"] = self.bert_config_path
self.model = XVLM(self.config)
if "xvlm-pretrained" in variant:
is_pretrained=True
else:
is_pretrained=False
self.model.load_pretrained(self.model_path, self.config, is_eval=True, is_pretrained=is_pretrained)
self.model = self.model.to(device)
self.model = self.model.eval()
self.device = device
if self.config['use_roberta']:
self.tokenizer = RobertaTokenizer.from_pretrained(self.config['text_encoder'])
else:
# TODO: Hack. We should use the tokenizer from the config
self.tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
def download(self):
import gdown
print(f"Downloading XVLM model to {self.root_dir}...")
model_url = download_urls[self.variant]["model_url"]
config_url = download_urls[self.variant]["config_url"]
bert_config_url = download_urls[self.variant]["bert_config_url"]
vision_config_url = download_urls[self.variant]["vision_config_url"]
os.makedirs(os.path.join(self.root_dir, "configs"), exist_ok=True)
gdown.download(id=model_url, output=self.model_path, quiet=False)
gdown.download(id=config_url, output=self.config_path, quiet=False)
subprocess.call(["wget", "-c", bert_config_url, "-O", self.bert_config_path])
subprocess.call(["wget", "-c", vision_config_url, "-O", self.vision_config_path])
@torch.no_grad()
def get_text_embeddings(self, texts, text_batch_size=256):
num_text = len(texts)
text_bs = 256
text_ids = []
text_embeds = []
text_atts = []
for i in range(0, num_text, text_bs):
text = texts[i: min(num_text, i+text_bs)]
text_input = self.tokenizer(text, padding='max_length', truncation=True, max_length=self.config["max_tokens"], return_tensors="pt").to(self.device)
text_output = self.model.text_encoder(text_input.input_ids, attention_mask = text_input.attention_mask, mode='text')
text_embed = F.normalize(self.model.text_proj(text_output.last_hidden_state[:,0,:]))
text_embeds.append(text_embed)
text_ids.append(text_input.input_ids)
text_atts.append(text_input.attention_mask)
text_embeds = torch.cat(text_embeds,dim=0)
text_ids = torch.cat(text_ids,dim=0)
text_atts = torch.cat(text_atts,dim=0)
#text_ids[:,0] = self.tokenizer.enc_token_id
return text_embeds, text_ids, text_atts
@torch.no_grad()
def get_image_embeddings(self, image_loader):
image_feats = []
image_embeds = []
for batch in tqdm(image_loader):
image = batch["image"]
image = image.to(self.device)
image_feat = self.model.vision_encoder(image)
image_embed = self.model.vision_proj(image_feat[:,0,:])
image_embed = F.normalize(image_embed,dim=-1)
image_feats.append(image_feat.cpu())
image_embeds.append(image_embed)
image_feats = torch.cat(image_feats,dim=0)
image_embeds = torch.cat(image_embeds,dim=0)
return image_feats, image_embeds
@torch.no_grad()
def get_retrieval_scores_dataset(self, loader):
texts = loader.dataset.text
metric_logger = MetricLogger(delimiter=" ")
text_embeds, text_ids, text_atts = self.get_text_embeddings(texts)
image_feats, image_embeds = self.get_image_embeddings(loader)
sims_matrix = image_embeds @ text_embeds.t()
score_matrix_i2t = torch.full((image_embeds.shape[0],len(texts)),-100.0).to(self.device)
num_tasks = 1
rank = 0
step = sims_matrix.size(0)//num_tasks + 1
start = rank*step
end = min(sims_matrix.size(0),start+step)
for i,sims in enumerate(metric_logger.log_every(sims_matrix[start:end], 50, "Evaluation i2T")):
topk_sim, topk_idx = sims.topk(k=self.config['k_test'], dim=0)
encoder_output = image_feats[start+i].repeat(self.config['k_test'],1,1).to(self.device)
encoder_att = torch.ones(encoder_output.size()[:-1],dtype=torch.long).to(self.device)
output = self.model.text_encoder(text_ids[topk_idx],
attention_mask = text_atts[topk_idx],
encoder_hidden_states = encoder_output,
encoder_attention_mask = encoder_att,
return_dict = True,
)
score = self.model.itm_head(output.last_hidden_state[:,0,:])[:,1]
score_matrix_i2t[start+i,topk_idx] = score + topk_sim
sims_matrix = sims_matrix.t()
score_matrix_t2i = torch.full((len(texts),image_feats.shape[0]),-100.0).to(self.device)
step = sims_matrix.size(0)//num_tasks + 1
start = rank*step
end = min(sims_matrix.size(0),start+step)
for i,sims in enumerate(metric_logger.log_every(sims_matrix[start:end], 50, "Evaluation T2i")):
topk_sim, topk_idx = sims.topk(k=self.config['k_test'], dim=0)
encoder_output = image_feats[topk_idx].to(self.device)
encoder_att = torch.ones(encoder_output.size()[:-1],dtype=torch.long).to(self.device)
output = self.model.text_encoder(text_ids[start+i].repeat(self.config['k_test'],1),
attention_mask = text_atts[start+i].repeat(self.config['k_test'],1),
encoder_hidden_states = encoder_output,
encoder_attention_mask = encoder_att,
return_dict = True,
)
score = self.model.itm_head(output.last_hidden_state[:,0,:])[:,1]
score_matrix_t2i[start+i,topk_idx] = score + topk_sim
return score_matrix_i2t.cpu().numpy(), score_matrix_t2i.cpu().numpy()
def run_scores_batched(self, image_embeds, image_feats, text_embeds, text_ids, text_atts):
# Should return something with shape (n_tests, n_image_options, n_text_options)
# Image embeds and all: (n_tests, n_image_options, embed_dim)
# Text embeds and all: (n_tests, n_text_options, embed_dim)
# Score matrix should be of the size: (n_tests, n_image_options, n_text_options)
sims_matrix = torch.einsum('ijk,ilk->ijl', image_embeds, text_embeds) # (n_tests, n_image_options, n_text_options)
score_matrix_i2t = torch.full((sims_matrix.shape[0], sims_matrix.shape[1], sims_matrix.shape[2]),-100.0).to(self.device)
for i, sims in enumerate(sims_matrix):
for j in range(sims.shape[0]):
encoder_output = image_feats[i, j].repeat(sims_matrix.shape[2],1,1).to(self.device)
encoder_att = torch.ones(encoder_output.size()[:-1],dtype=torch.long).to(self.device)
output = self.model.text_encoder(text_ids[i],
attention_mask = text_atts[i],
encoder_hidden_states = encoder_output,
encoder_attention_mask = encoder_att,
return_dict = True)
score = self.model.itm_head(output.last_hidden_state[:,0,:])[:,1]
score_matrix_i2t[i,j] = score + sims[j]
sims_matrix = sims_matrix.permute(0,2,1) # (n_tests, n_text_options, n_image_options)
score_matrix_t2i = torch.full((sims_matrix.shape[0], sims_matrix.shape[1], sims_matrix.shape[2]),-100.0).to(self.device)
for i, sims in enumerate(sims_matrix):
for j in range(sims.shape[0]):
encoder_output = image_feats[i].to(self.device)
encoder_att = torch.ones(encoder_output.size()[:-1],dtype=torch.long).to(self.device)
output = self.model.text_encoder(text_ids[i, j].repeat(sims_matrix.shape[2],1),
attention_mask = text_atts[i, j].repeat(sims_matrix.shape[2],1),
encoder_hidden_states = encoder_output,
encoder_attention_mask = encoder_att,
return_dict = True)
score = self.model.itm_head(output.last_hidden_state[:,0,:])[:,1]
score_matrix_t2i[i,j] = score + sims[j]
return score_matrix_i2t.cpu().numpy(), score_matrix_t2i.cpu().numpy()
@torch.no_grad()
def get_retrieval_scores_batched(self, joint_loader):
"""Computes the scores for each image_option / caption_option pair in the joint loader.
Args:
joint_loader (DataLoader): batches have "image_options" and "caption_options" fields.
"image_options" is a list of images, and "caption_options" is a list of captions.
Returns:
all_scores: A numpy array containing the scores of the shape NxKxL,
where N is the number of test cases, K is the number of image options per the test case,
and L is the number of caption options per the test case.
"""
t2i_scores, i2t_scores = [], []
for batch in tqdm(joint_loader):
image_feats = []
image_embeds = []
for i_option in batch["image_options"]:
image_feat = self.model.vision_encoder(i_option.to(self.device))
image_embed = self.model.vision_proj(image_feat[:,0,:]) # B x D
image_embed = F.normalize(image_embed,dim=-1)
image_feats.append(image_feat.unsqueeze(1))
image_embeds.append(image_embed.unsqueeze(1))
image_feats = torch.cat(image_feats,dim=1)
image_embeds = torch.cat(image_embeds,dim=1)
text_ids = []
text_embeds = []
text_atts = []
for c_option in batch["caption_options"]:
c_option = list(c_option)
text_input = self.tokenizer(c_option, padding='max_length', truncation=True, max_length=self.config["max_tokens"], return_tensors="pt").to(self.device)
text_output = self.model.text_encoder(text_input.input_ids, attention_mask = text_input.attention_mask, mode='text')
text_embed = F.normalize(self.model.text_proj(text_output.last_hidden_state[:,0,:]))
text_embeds.append(text_embed.unsqueeze(1))
text_ids.append(text_input.input_ids.unsqueeze(1))
text_atts.append(text_input.attention_mask.unsqueeze(1))
text_embeds = torch.cat(text_embeds,dim=1)
text_ids = torch.cat(text_ids,dim=1)
text_atts = torch.cat(text_atts,dim=1)
s_i2t, s_t2i = self.run_scores_batched(image_embeds, image_feats, text_embeds, text_ids, text_atts)
t2i_scores.append(s_t2i)
i2t_scores.append(s_i2t)
t2i_scores = np.concatenate(t2i_scores, axis=0) # N x N_t x N_i
t2i_scores = np.transpose(t2i_scores, (0, 2, 1)) # N x N_i x N_t
i2t_scores = np.concatenate(i2t_scores, axis=0) # N x N_i x N_t
print(t2i_scores.shape, i2t_scores.shape)
return t2i_scores, i2t_scores