-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy patheval_utils.py
176 lines (145 loc) · 6.47 KB
/
eval_utils.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
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import torch
import torch.nn as nn
from torch.autograd import Variable
import numpy as np
import json
from json import encoder
import random
import string
import time
import os
import sys
import misc.utils as utils
def language_eval_chinese(dataset, preds, model_id, split):
from caption_eval.coco_caption.pycxtools.coco import COCO
from caption_eval.coco_caption.pycxevalcap.eval import COCOEvalCap
m1_score = {}
m1_score['error'] = 0
encoder.FLOAT_REPR = lambda o: format(o, '.3f')
if not os.path.isdir('eval_results'):
os.mkdir('eval_results')
cache_path = os.path.join('eval_results/', model_id + '_' + split + '.json')
reference_file = '/home/hc/image_root/ai_challenger_caption_validation_20170910/coco_val_caption_validation_annotations_20170910.json'
coco = COCO(reference_file)
valids = coco.getImgIds()
# filter results to only those in MSCOCO validation set (will be about a third)
preds_filt = [p for p in preds if p['image_id'] in valids]
print('using %d/%d predictions' % (len(preds_filt), len(preds)))
json.dump(preds_filt, open(cache_path, 'w')) # serialize to temporary json file. Sigh, COCO API...
coco_res = coco.loadRes(cache_path)
coco_eval = COCOEvalCap(coco, coco_res)
coco_eval.params['image_id'] = coco_res.getImgIds()
# evaluate results
coco_eval.evaluate()
# print output evaluation scores
for metric, score in coco_eval.eval.items():
print('%s: %.3f' % (metric, score))
m1_score[metric] = score
return m1_score
def language_eval(dataset, preds, model_id, split):
import sys
if 'coco' in dataset:
sys.path.append("coco-caption")
annFile = 'coco-caption/annotations/captions_val2014.json'
else:
sys.path.append("f30k-caption")
annFile = 'f30k-caption/annotations/dataset_flickr30k.json'
from pycocotools.coco import COCO
from pycocoevalcap.eval import COCOEvalCap
encoder.FLOAT_REPR = lambda o: format(o, '.3f')
if not os.path.isdir('eval_results'):
os.mkdir('eval_results')
cache_path = os.path.join('eval_results/', model_id + '_' + split + '.json')
coco = COCO(annFile)
valids = coco.getImgIds()
# filter results to only those in MSCOCO validation set (will be about a third)
preds_filt = [p for p in preds if p['image_id'] in valids]
print('using %d/%d predictions' % (len(preds_filt), len(preds)))
json.dump(preds_filt, open(cache_path, 'w')) # serialize to temporary json file. Sigh, COCO API...
cocoRes = coco.loadRes(cache_path)
cocoEval = COCOEvalCap(coco, cocoRes)
cocoEval.params['image_id'] = cocoRes.getImgIds()
cocoEval.evaluate()
# create output dictionary
out = {}
for metric, score in cocoEval.eval.items():
out[metric] = score
imgToEval = cocoEval.imgToEval
for p in preds_filt:
image_id, caption = p['image_id'], p['caption']
imgToEval[image_id]['caption'] = caption
with open(cache_path, 'w') as outfile:
json.dump({'overall': out, 'imgToEval': imgToEval}, outfile)
return out
def eval_split(model, crit, loader, eval_kwargs={}):
verbose = eval_kwargs.get('verbose', True)
num_images = eval_kwargs.get('num_images', eval_kwargs.get('val_images_use', -1))
split = eval_kwargs.get('split', 'val')
lang_eval = eval_kwargs.get('language_eval', 0)
dataset = eval_kwargs.get('dataset', 'coco')
beam_size = eval_kwargs.get('beam_size', 1)
# Make sure in the evaluation mode
model.eval()
loader.reset_iterator(split)
n = 0
loss = 0
loss_sum = 0
loss_evals = 1e-8
predictions = []
while True:
data = loader.get_batch(split)
n = n + loader.batch_size
if data.get('labels', None) is not None:
# forward the model to get loss
tmp = [data['fc_feats'], data['att_feats'], data['labels'], data['masks']]
tmp = [Variable(torch.from_numpy(_), volatile=True).cuda() for _ in tmp]
fc_feats, att_feats, labels, masks = tmp
loss = crit(model(fc_feats, att_feats, labels), labels[:,1:], masks[:,1:]).data[0]
loss_sum = loss_sum + loss
loss_evals = loss_evals + 1
# forward the model to also get generated samples for each image
# Only leave one feature for each image, in case duplicate sample
tmp = [data['fc_feats'][np.arange(loader.batch_size) * loader.seq_per_img],
data['att_feats'][np.arange(loader.batch_size) * loader.seq_per_img]]
tmp = [Variable(torch.from_numpy(_), volatile=True).cuda() for _ in tmp]
fc_feats, att_feats = tmp
# forward the model to also get generated samples for each image
seq, _ = model.sample_beam(fc_feats, att_feats, eval_kwargs)
#set_trace()
sents = utils.decode_sequence(loader.get_vocab(), seq)
for k, sent in enumerate(sents):
sent = sent.replace(" ","")
entry = {'image_id': data['infos'][k]['id'], 'caption': sent}
if eval_kwargs.get('dump_path', 0) == 1:
entry['file_name'] = data['infos'][k]['file_path']
predictions.append(entry)
if eval_kwargs.get('dump_images', 0) == 1:
# dump the raw image to vis/ folder
cmd = 'cp "' + os.path.join(eval_kwargs['image_root'], data['infos'][k]['file_path']) + '" vis/imgs/img' + entry['image_id'] + '.jpg' # bit gross str(len(predictions))
print(cmd)
os.system(cmd)
if verbose:
print('image %s: %s' %(entry['image_id'], entry['caption']))
# if we wrapped around the split or used up val imgs budget then bail
ix0 = data['bounds']['it_pos_now']
ix1 = data['bounds']['it_max']
if num_images != -1:
ix1 = min(ix1, num_images)
for i in range(n - ix1):
predictions.pop()
if verbose:
print('evaluating validation preformance... %d/%d (%f)' %(ix0 - 1, ix1, loss))
if data['bounds']['wrapped']:
break
if num_images >= 0 and n >= num_images:
break
#break # use to debug whole network performance
lang_stats = None
if lang_eval == 1:
lang_stats = language_eval_chinese(dataset, predictions, eval_kwargs['id'], split)
# Switch back to training mode
model.train()
return loss_sum/loss_evals, predictions, lang_stats