-
Notifications
You must be signed in to change notification settings - Fork 3
/
tinyimagenet_eval.py
129 lines (106 loc) · 5.83 KB
/
tinyimagenet_eval.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
import argparse
import torch
from transformers import BertGenerationTokenizer, BertGenerationDecoder, BertGenerationConfig
import os
from dataloaders.ZO_Clip_loaders import tinyimage_single_isolated_class_loader
from clip.simple_tokenizer import SimpleTokenizer as clip_tokenizer
from tqdm import tqdm
import numpy as np
from sklearn.metrics import roc_auc_score
def tokenize_for_clip(batch_sentences, tokenizer):
default_length = 77 # CLIP default
sot_token = tokenizer.encoder['<|startoftext|>']
eot_token = tokenizer.encoder['<|endoftext|>']
tokenized_list = []
for sentence in batch_sentences:
text_tokens = [sot_token] + tokenizer.encode(sentence) + [eot_token]
tokenized = torch.zeros((default_length), dtype=torch.long)
tokenized[:len(text_tokens)] = torch.tensor(text_tokens)
tokenized_list.append(tokenized)
tokenized_list = torch.stack(tokenized_list)
return tokenized_list
def greedysearch_generation_topk(clip_embed):
max_len=77
N = 1 # batch has single sample
target_list = [torch.tensor(berttokenizer.bos_token_id)]
top_k_list = []
bert_model.eval()
for i in range(max_len):
target = torch.LongTensor(target_list).unsqueeze(0)
position_ids = torch.arange(0, len(target)).expand(N, len(target)).to(device)
with torch.no_grad():
out = bert_model(input_ids=target.to(device),
position_ids=position_ids,
attention_mask=torch.ones(len(target)).unsqueeze(0).to(device),
encoder_hidden_states=clip_embed.unsqueeze(1).to(device),
)
pred_idx = out.logits.argmax(2)[:, -1]
_, top_k = torch.topk(out.logits, dim=2, k=35)
top_k_list.append(top_k[:, -1].flatten())
target_list.append(pred_idx)
#if pred_idx == berttokenizer.eos_token_id or len(target_list)==10: #the entitiy word is in at most first 10 words
if len(target_list) == 10: # the entitiy word is in at most first 10 words
break
top_k_list = torch.cat(top_k_list)
return target_list, top_k_list
def image_decoder(clip_model, berttokenizer, device, split, image_loaders=None):
seen_labels = split[:20]
seen_descriptions = [f"This is a photo of a {label}" for label in seen_labels]
targets = torch.tensor(1000*[0] + 9000*[1])
#targets = torch.tensor(100*[0] + 900*[1])
ood_probs_sum = []
max_num_entities = 0
for i, semantic_label in enumerate(split):
loader = image_loaders[semantic_label]
for idx, image in enumerate(tqdm(loader)):
#if idx==5:break
with torch.no_grad():
clip_out = clip_model.encode_image(image.to(device)).float()
clip_extended_embed = clip_out.repeat(1, 2).type(torch.FloatTensor)
#greedy generation
target_list, topk_list = greedysearch_generation_topk(clip_extended_embed)
target_tokens = [berttokenizer.decode(int(pred_idx.cpu().numpy())) for pred_idx in target_list]
topk_tokens = [berttokenizer.decode(int(pred_idx.cpu().numpy())) for pred_idx in topk_list]
unique_entities = list(set(topk_tokens))
if len(unique_entities) > max_num_entities:
max_num_entities = len(unique_entities)
all_desc = seen_descriptions + [f"This is a photo of a {label}" for label in unique_entities]
all_desc_ids = tokenize_for_clip(all_desc, cliptokenizer)
image_feature = clip_model.encode_image(image.cuda()).float()
image_feature /= image_feature.norm(dim=-1, keepdim=True)
text_features = clip_model.encode_text(all_desc_ids.cuda()).float()
text_features /= text_features.norm(dim=-1, keepdim=True)
zeroshot_probs = (100.0 * image_feature @ text_features.T).softmax(dim=-1).squeeze()
#detection score is accumulative sum of probs of generated entities
ood_prob_sum = np.sum(zeroshot_probs[20:].detach().cpu().numpy())
ood_probs_sum.append(ood_prob_sum)
auc_sum = roc_auc_score(np.array(targets), np.squeeze(ood_probs_sum))
print(' sum_ood AUROC={}'.format(auc_sum))
return auc_sum
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--trained_path', type=str, default='./trained_models/COCO/')
args = parser.parse_args()
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
args.saved_model_path = args.trained_path + '/ViT-B32/'
if not os.path.exists(args.saved_model_path):
os.makedirs(args.saved_model_path)
# initialize tokenizers for clip and bert, these two use different tokenizers
berttokenizer = BertGenerationTokenizer.from_pretrained('google/bert_for_seq_generation_L-24_bbc_encoder')
clip_model = torch.jit.load(os.path.join('./trained_models', "{}.pt".format('ViT-B-32'))).to(device).eval()
cliptokenizer = clip_tokenizer()
bert_config = BertGenerationConfig.from_pretrained("google/bert_for_seq_generation_L-24_bbc_encoder")
bert_config.is_decoder=True
bert_config.add_cross_attention=True
bert_model = BertGenerationDecoder.from_pretrained('google/bert_for_seq_generation_L-24_bbc_encoder',
config=bert_config).to(device).train()
bert_model.load_state_dict(torch.load(args.saved_model_path + 'model.pt')['net'])
splits, tinyimg_loaders = tinyimage_single_isolated_class_loader()
max_scores = []
sum_scores = []
mean_scores = []
for split in splits:
sum_score = image_decoder(clip_model, berttokenizer, device, split=split, image_loaders=tinyimg_loaders)
sum_scores.append(sum_score)
print('sum auc on 5 splits:', sum_scores)
print('average on 5 splits with sum score:', np.mean(sum_scores), 'stdv:', np.std(sum_scores))