-
Notifications
You must be signed in to change notification settings - Fork 0
/
metrics.py
210 lines (179 loc) · 8.34 KB
/
metrics.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
import torch
import numpy as np
import pickle
import time
import json
from pycocoevalcap.bleu.bleu import Bleu
from pycocoevalcap.cider.cider import Cider
from pycocoevalcap.meteor.meteor import Meteor
from pycocoevalcap.rouge.rouge import Rouge
# Device configuration
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
def _imp_generator(predicted_imps, batch_idx, vocab):
""""generate the impression from predicted id"""
impression_ids = predicted_imps[batch_idx]
impression_words = []
for word_id in impression_ids:
word = vocab.id2word[word_id]
if word == '<start>' or word == '<pad>' or word == "<unk>":
continue
# <end> will stop the generation of one sentence
elif word == '<end>':
break
else:
impression_words.append(word)
impression = ' '.join(impression_words)
# Take punctuation as a single word
impression = impression.lower()
return impression
def _fin_generator(predicted_fins, batch_idx, vocab):
""""generate the finding from predicted id"""
finding_ids = predicted_fins[batch_idx]
# print(finding_ids)
finding_sentences = []
for num_sen in range(finding_ids.shape[0]):
single_sentence_ids = finding_ids[num_sen]
single_sentence_words = []
for word_id in single_sentence_ids:
word = vocab.id2word[word_id]
if word == '<start>' or word == '<pad>' or word == "<unk>":
continue
# <end> will stop the generation of one sentence
elif word == '<end>':
break
else:
single_sentence_words.append(word)
single_sentence = ' '.join(single_sentence_words)
# empty sentence will stop the generation
if not single_sentence:
break
else:
# Take punctuation as a single word
single_sentence = single_sentence.lower()
finding_sentences.append(single_sentence)
finding = ' '.join(finding_sentences)
return finding
def _gt_imp_generator(para):
""" Modify the ground truth impression to be consistent with our training.
Every impression sentence must end with '.',
all '.' in the middle of impression will be changed to ',' """
para = para.split('.')
gt_imp = ''
for i, sentence in enumerate(para):
# remove the leading or trailing spaces
sentence = sentence.strip()
if len(sentence) > 0:
sentence = sentence + ' , '
gt_imp = gt_imp + sentence
gt_imp_lst = gt_imp.split()
# change the last ',' to '.'
gt_imp_lst[-1] = '.'
gt_imp = ' '.join(gt_imp_lst)
# leave space between the original , and the word before it
gt_imp = gt_imp.lower().replace(', ', ' , ')
# fix the two-space issue caused by last operation
gt_imp = gt_imp.replace(' , ', ' , ')
return gt_imp
def _gt_fin_generator(para):
"""Modify the ground truth finding to be consistent with our training
Every finding sentence must be ended with '.' """
gt_fin = para.lower().replace(', ', ' , ').replace('. ', ' . ')
return gt_fin
def _generate_imp_fin_dict(predicted_imps_lst, predicted_fins_lst, image_ids_lst, args):
pre_imp_dict = {}
pre_fin_dict = {}
pre_imp_fin_dict = {}
gt_imp_dict = {}
gt_fin_dict = {}
gt_imp_fin_dict = {}
with open(args.vocab_path, 'rb') as f:
vocab = pickle.load(f)
with open(args.eval_json_dir) as f:
data = json.load(f)
# The length of predicted impression lst is the number of batches
assert len(predicted_imps_lst) == len(predicted_fins_lst) == len(image_ids_lst)
for idx in range(len(predicted_imps_lst)):
if torch.cuda.is_available():
predicted_imps = predicted_imps_lst[idx].cpu().data.numpy()
predicted_fins = predicted_fins_lst[idx].cpu().data.numpy()
else:
predicted_imps = predicted_imps_lst[idx].data.numpy()
predicted_fins = predicted_fins_lst[idx].data.numpy()
image_ids = np.asarray(image_ids_lst[idx])
# shape 0 is the number of samples in a batch
assert predicted_imps.shape[0] == predicted_fins.shape[0] == image_ids.shape[0]
for batch_idx in range(predicted_imps.shape[0]):
img_id = image_ids[batch_idx]
# Impressions:
pre_imp = _imp_generator(predicted_imps, batch_idx, vocab)
gt_imp = data[img_id][0]
gt_imp = _gt_imp_generator(gt_imp)
pre_imp_dict[img_id] = [pre_imp]
gt_imp_dict[img_id] = [gt_imp]
# Findings:
pre_fin = _fin_generator(predicted_fins, batch_idx, vocab)
gt_fin = data[img_id][1]
gt_fin = _gt_fin_generator(gt_fin)
pre_fin_dict[img_id] = [pre_fin]
gt_fin_dict[img_id] = [gt_fin]
# Impression+Finding
pre_imp_fin = pre_imp + ' ' + pre_fin
gt_imp_fin = gt_imp + ' ' + gt_fin
pre_imp_fin_dict[img_id] = [pre_imp_fin]
gt_imp_fin_dict[img_id] = [gt_imp_fin]
return gt_imp_dict, pre_imp_dict, gt_fin_dict, pre_fin_dict, gt_imp_fin_dict, pre_imp_fin_dict
def _define_metrics(gts, res):
bleu_scorer = Bleu(n=4)
bleu, _ = bleu_scorer.compute_score(gts=gts, res=res)
rouge_scorer = Rouge()
rouge, _ = rouge_scorer.compute_score(gts=gts, res=res)
cider_scorer = Cider()
cider, _ = cider_scorer.compute_score(gts=gts, res=res)
meteor_scorer = Meteor()
meteor, _ = meteor_scorer.compute_score(gts=gts, res=res)
for i in range(4):
bleu[i] = round(bleu[i], 4)
return bleu, round(meteor, 4), round(rouge, 4), round(cider, 4)
def compute_metrics(predicted_imps_lst, predicted_fins_lst, image_ids_lst, args):
gt_imp_dic, pre_imp_dic, gt_fin_dic, pre_fin_dic, gt_imp_fin_dic, pre_imp_fin_dic = \
_generate_imp_fin_dict(predicted_imps_lst, predicted_fins_lst, image_ids_lst, args)
if args.imp_fin_only:
imp_fin_bleu, imp_fin_meteor, imp_fin_rouge, imp_fin_cider = _define_metrics(gt_imp_fin_dic, pre_imp_fin_dic)
print('Impression + Finding: bleu = %s, meteor = %s, rouge = %s, cider = %s' % (
imp_fin_bleu, imp_fin_meteor, imp_fin_rouge, imp_fin_cider))
else:
imp_bleu, imp_meteor, imp_rouge, imp_cider = _define_metrics(gt_imp_dic, pre_imp_dic)
fin_bleu, fin_meteor, fin_rouge, fin_cider = _define_metrics(gt_fin_dic, pre_fin_dic)
imp_fin_bleu, imp_fin_meteor, imp_fin_rouge, imp_fin_cider = _define_metrics(gt_imp_fin_dic, pre_imp_fin_dic)
print(
'Impression: bleu = %s, meteor = %s, rouge = %s, cider = %s' % (imp_bleu, imp_meteor, imp_rouge, imp_cider))
print('Finding: bleu = %s, meteor = %s, rouge = %s, cider = %s' % (fin_bleu, fin_meteor, fin_rouge, fin_cider))
print('Impression + Finding: bleu = %s, meteor = %s, rouge = %s, cider = %s' % (
imp_fin_bleu, imp_fin_meteor, imp_fin_rouge, imp_fin_cider))
return imp_fin_bleu, imp_fin_meteor, imp_fin_rouge, imp_fin_cider
def _writer(txt, imp_dict, fin_dict):
assert imp_dict.keys() == fin_dict.keys()
for key in imp_dict.keys():
txt.write(key)
txt.write("\n")
txt.write("Impression: ")
txt.write(imp_dict[key][0])
txt.write("\n")
txt.write("Findings: ")
txt.write(fin_dict[key][0])
txt.write("\n")
def generate_text_file(predicted_imps_lst, predicted_fins_lst, image_ids_lst, num_run, args):
gt_imp_dic, pre_imp_dic, gt_fin_dic, pre_fin_dic, gt_imp_fin_dic, pre_imp_fin_dic = \
_generate_imp_fin_dict(predicted_imps_lst, predicted_fins_lst, image_ids_lst, args)
# num_run为字符串“test”, 生成测试txt
if isinstance(num_run, str):
gt_txt = open("results/{}_gt_results_{}.txt".format(num_run + "run", time.strftime('%Y-%m-%d-%H-%M')), "+w")
pre_txt = open("results/{}_pre_results_{}.txt".format(num_run + "run", time.strftime('%Y-%m-%d-%H-%M')), "+w")
# num_run为int 3 , 训练过程中生成
else:
gt_txt = open("results/{}_gt_results_{}.txt".format(str(num_run + 1) + "run", time.strftime('%Y-%m-%d-%H-%M')),
"+w")
pre_txt = open(
"results/{}_pre_results_{}.txt".format(str(num_run + 1) + "run", time.strftime('%Y-%m-%d-%H-%M')), "+w")
_writer(gt_txt, gt_imp_dic, gt_fin_dic)
_writer(pre_txt, pre_imp_dic, pre_fin_dic)