/
eval_sentence_predictions.py
220 lines (178 loc) · 8.07 KB
/
eval_sentence_predictions.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
import argparse
import json
import time
import datetime
import numpy as np
import code
import socket
import os
import cPickle as pickle
import math
from imagernn.data_provider import getDataProvider
from imagernn.solver import Solver
from imagernn.imagernn_utils import decodeGenerator, eval_split
def main(params):
# load the checkpoint
checkpoint_path = params['checkpoint_path']
max_images = params['max_images']
print 'loading checkpoint %s' % (checkpoint_path, )
checkpoint = pickle.load(open(checkpoint_path, 'rb'))
checkpoint_params = checkpoint['params']
dataset = checkpoint_params['dataset']
model = checkpoint['model']
dump_folder = params['dump_folder']
if dump_folder:
print 'creating dump folder ' + dump_folder
os.system('mkdir -p ' + dump_folder)
# fetch the data provider
dp = getDataProvider(dataset)
misc = {}
misc['wordtoix'] = checkpoint['wordtoix']
ixtoword = checkpoint['ixtoword']
blob = {} # output blob which we will dump to JSON for visualizing the results
blob['params'] = params
blob['checkpoint_params'] = checkpoint_params
blob['imgblobs'] = []
# iterate over all images in test set and predict sentences
BatchGenerator = decodeGenerator(checkpoint_params)
n = 0
all_references = []
all_candidates = []
captions_res = []
for img in dp.iterImages(split = 'test', max_images = max_images):
n+=1
print 'image %d/%d:' % (n, max_images)
references = [' '.join(x['tokens']) for x in img['sentences']] # as list of lists of tokens
kwparams = { 'beam_size' : params['beam_size'] }
Ys = BatchGenerator.predict([{'image':img}], model, checkpoint_params, **kwparams)
img_blob = {} # we will build this up
img_blob['img_path'] = img['local_file_path']
img_blob['imgid'] = img['imgid']
img_blob['id'] = img['id']
if dump_folder:
# copy source file to some folder. This makes it easier to distribute results
# into a webpage, because all images that were predicted on are in a single folder
source_file = img['local_file_path']
target_file = os.path.join(dump_folder, os.path.basename(img['local_file_path']))
os.system('cp %s %s' % (source_file, target_file))
# encode the human-provided references
img_blob['references'] = []
flag = True
for gtsent in references:
if flag:
print 'GT: ' + gtsent
flag = False
img_blob['references'].append({'text': gtsent})
# now evaluate and encode the top prediction
top_predictions = Ys[0] # take predictions for the first (and only) image we passed in
top_prediction = top_predictions[0] # these are sorted with highest on top
candidate = ' '.join([ixtoword[ix] for ix in top_prediction[1] if ix > 0]) # ix 0 is the END token, skip that
print 'PRED: (%f) %s' % (top_prediction[0], candidate)
# save for later eval
all_references.append(references)
all_candidates.append(candidate)
captions_res.append({'image_id':img_blob['id'],'caption':candidate})
img_blob['candidate'] = {'text': candidate, 'logprob': top_prediction[0]}
blob['imgblobs'].append(img_blob)
# use perl script to eval BLEU score for fair comparison to other research work
# first write intermediate files
print 'writing intermediate files into eval/'
open('eval/output', 'w').write('\n'.join(all_candidates))
for q in xrange(5):
open('eval/reference'+`q`, 'w').write('\n'.join([x[q] for x in all_references]))
# invoke the perl script to get BLEU scores
print 'invoking eval/multi-bleu.perl script...'
owd = os.getcwd()
os.chdir('eval')
os.system('./multi-bleu.perl reference < output')
os.chdir(owd)
# # now also evaluate test split perplexity
# gtppl = eval_split('test', dp, model, checkpoint_params, misc, eval_max_images = max_images)
# print 'perplexity of ground truth words based on dictionary of %d words: %f' % (len(ixtoword), gtppl)
# blob['gtppl'] = gtppl
# dump result struct to file
print 'saving result struct to %s' % (params['result_struct_filename'], )
json.dump(blob, open(params['result_struct_filename'], 'w'))
alg_name = params['checkpoint_path'].split('_')[1]
res_file_name = params['out_dir']+'/captions_val_'+alg_name+'_results.json'
json.dump(captions_res, open(res_file_name, 'w'))
from eval_tools import metrics
scores = metrics.run(dataset,alg_name,params['out_dir'])
return scores
def run(checkpoint):
max_images = -1
dump_folder = ''
checkpoint_params = checkpoint['params']
dataset = checkpoint_params['dataset']
model = checkpoint['model']
beam_size = 1
# dump_folder = params['dump_folder']
# fetch the data provider
dp = getDataProvider(dataset)
misc = {}
misc['wordtoix'] = checkpoint['wordtoix']
ixtoword = checkpoint['ixtoword']
blob = {} # output blob which we will dump to JSON for visualizing the results
# blob['params'] = params
blob['checkpoint_params'] = checkpoint_params
blob['imgblobs'] = []
# iterate over all images in test set and predict sentences
BatchGenerator = decodeGenerator(checkpoint_params)
n = 0
all_references = []
all_candidates = []
captions_res = []
for img in dp.iterImages(split = 'test', max_images = max_images):
n+=1
print 'image %d/%d:' % (n, max_images)
references = [' '.join(x['tokens']) for x in img['sentences']] # as list of lists of tokens
kwparams = { 'beam_size' : beam_size }
Ys = BatchGenerator.predict([{'image':img}], model, checkpoint_params, **kwparams)
img_blob = {} # we will build this up
img_blob['img_path'] = img['local_file_path']
img_blob['imgid'] = img['imgid']
img_blob['id'] = img['id']
if dump_folder:
# copy source file to some folder. This makes it easier to distribute results
# into a webpage, because all images that were predicted on are in a single folder
source_file = img['local_file_path']
target_file = os.path.join(dump_folder, os.path.basename(img['local_file_path']))
os.system('cp %s %s' % (source_file, target_file))
# encode the human-provided references
img_blob['references'] = []
flag = True
for gtsent in references:
if flag:
print 'GT: ' + gtsent
flag = False
img_blob['references'].append({'text': gtsent})
# now evaluate and encode the top prediction
top_predictions = Ys[0] # take predictions for the first (and only) image we passed in
top_prediction = top_predictions[0] # these are sorted with highest on top
candidate = ' '.join([ixtoword[ix] for ix in top_prediction[1] if ix > 0]) # ix 0 is the END token, skip that
print 'PRED: (%f) %s' % (top_prediction[0], candidate)
# save for later eval
all_references.append(references)
all_candidates.append(candidate)
captions_res.append({'image_id':img_blob['id'],'caption':candidate})
img_blob['candidate'] = {'text': candidate, 'logprob': top_prediction[0]}
blob['imgblobs'].append(img_blob)
alg_name = checkpoint['algorithm']
res_file_name = checkpoint['outdir']+'/captions_val_'+alg_name+'_results.json'
json.dump(captions_res, open(res_file_name, 'w'))
from eval_tools import metrics
scores = metrics.run(dataset,alg_name,checkpoint['outdir'])
return scores
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('checkpoint_path', type=str, help='the input checkpoint')
parser.add_argument('-b', '--beam_size', type=int, default=1, help='beam size in inference. 1 indicates greedy per-word max procedure. Good value is approx 20 or so, and more = better.')
parser.add_argument('--result_struct_filename', type=str, default='result_struct.json', help='filename of the result struct to save')
parser.add_argument('-m', '--max_images', type=int, default=-1, help='max images to use')
parser.add_argument('-d', '--dump_folder', type=str, default="", help='dump the relevant images to a separate folder with this name?')
parser.add_argument('-o', '--out_dir', type=str, default="results", help='where predictions will be saved')
args = parser.parse_args()
params = vars(args) # convert to ordinary dict
print 'parsed parameters:'
print json.dumps(params, indent = 2)
main(params)