-
Notifications
You must be signed in to change notification settings - Fork 0
/
saliency_eval.py
135 lines (98 loc) · 5.34 KB
/
saliency_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
130
131
132
133
134
135
import torch.backends.cudnn as cudnn
import torch.optim
import torch.nn.functional as F
import torch.utils.data
import torchvision.transforms as transforms
from nltk.translate.bleu_score import corpus_bleu
from scorer import compute_metrics
from datasets import *
from utils import *
from tqdm import tqdm
data_root = 'final_dataset'
data_name = 'coco_5_cap_per_img_5_min_word_freq'
checkpoint_file = 'results/BEST_24Saliency_checkpoint_coco_5_cap_per_img_5_min_word_freq.pth.tar'
mapping_file = 'WORDMAP_coco_5_cap_per_img_5_min_word_freq.json'
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
cudnn.benchmark = True
mapping_file = os.path.join(data_root, 'WORDMAP_' + data_name + '.json')
with open(mapping_file, 'r') as j:
mapping = json.load(j)
reverse_mapping = {value: key for key, value in mapping.items()}
vocabSize = len(reverse_mapping)
# Load model
torch.nn.Module.dump_patches = True
checkpoint = torch.load(checkpoint_file, map_location = device)
decoder = checkpoint['decoder']
decoder = decoder.to(device)
decoder.eval() # turn on evaluation mode
def evaluate(beam_size):
groundtruths = list() # true captions, list of lists as each image may have several truth captions
predictions = list() # predictions , list of predictions
# DataLoader
test_loader = torch.utils.data.DataLoader(SaliencyCustomDataset(data_root, data_name, 'TEST'), batch_size=1, shuffle=True, num_workers=1, pin_memory=torch.cuda.is_available())
for i, (imagefeatures, salfeatures, sequence, sequencelength, sequences_generated) in enumerate(tqdm(test_loader)):
k = beam_size
imagefeatures = imagefeatures.to(device)
sequence = sequence.to(device)
sequencelength = sequencelength.to(device)
salfeatures = salfeatures.to(device)
sequences_generated = sequences_generated.to(device)
imagefeatures_mean = torch.mean(imagefeatures, dim=1).expand(k, 2048)
best_k_scores = torch.zeros(k, 1).to(device)
prev_k_sequences = torch.LongTensor([[mapping['<start>']]] * k).to(device)
k_sequences = prev_k_sequences
hidden1, cell1 = decoder.init_hidden_state(k)
hidden2, cell2 = decoder.init_hidden_state(k)
complete_seqs = list()
complete_seqs_scores = list()
los = 1
while True:
embeddings = decoder.embedding(prev_k_sequences).squeeze(1)
hidden1,cell1 = decoder.TD(torch.cat([hidden2,imagefeatures_mean,embeddings], dim=1),(hidden1,cell1))
attention_weighted_encoding = decoder.attModule(imagefeatures,salfeatures,hidden1)
hidden2,cell2 = decoder.lang_layer(torch.cat([attention_weighted_encoding,hidden1], dim=1),(hidden2,cell2))
scores = decoder.linear(hidden2)
scores = F.log_softmax(scores, dim=1)
scores = torch.add(scores, best_k_scores.expand(scores.size()))
if los == 1:
best_k_scores, best_k_sequences = scores[0].topk(k, 0, True, True)
else:
best_k_scores, best_k_sequences = scores.view(-1).topk(k, dim=0, largest=True, sorted=True)
prev_word_inds = torch.div(best_k_sequences, vocabSize, rounding_mode="trunc")
word_inds = best_k_sequences % vocabSize
k_sequences = torch.cat([k_sequences[prev_word_inds], word_inds.unsqueeze(1)], dim=1)
continue_indices = [index for index, word in enumerate(word_inds) if word != mapping['<end>']]
end_indices = list(set(range(len(word_inds))).difference(set(continue_indices)))
if len(end_indices) > 0:
complete_seqs.extend(k_sequences[end_indices].tolist())
complete_seqs_scores.extend(best_k_scores[end_indices])
k -= len(end_indices)
if k == 0:
break
hidden1 = hidden1[prev_word_inds[continue_indices]]
cell1 = cell1[prev_word_inds[continue_indices]]
hidden2 = hidden2[prev_word_inds[continue_indices]]
cell2 = cell2[prev_word_inds[continue_indices]]
imagefeatures_mean = imagefeatures_mean[prev_word_inds[continue_indices]]
k_sequences = k_sequences[continue_indices]
best_k_scores = best_k_scores[continue_indices].unsqueeze(1)
prev_k_sequences = word_inds[continue_indices].unsqueeze(1)
if los > 50:
break
los += 1
x = complete_seqs_scores.index(max(complete_seqs_scores))
best_sequence = complete_seqs[x]
cap = sequences_generated[0].tolist()
caps = list(
map(lambda c: [reverse_mapping[w] for w in c if w not in {mapping['<start>'], mapping['<end>'], mapping['<pad>']}], cap)) # remove <start> and pads
str_caps = [' '.join(c) for c in caps]
groundtruths.append(str_caps)
prediction = ([reverse_mapping[w] for w in best_sequence if w not in {mapping['<start>'], mapping['<end>'], mapping['<pad>']}])
prediction = ' '.join(prediction)
predictions.append(prediction)
metrics_dict = compute_metrics(groundtruths, predictions)
return metrics_dict
if __name__ == '__main__':
beam_size = 5
metrics_dict = evaluate(beam_size)
print(metrics_dict)