-
Notifications
You must be signed in to change notification settings - Fork 0
/
predictor_ITV_batch_querysets.py
402 lines (341 loc) · 20.1 KB
/
predictor_ITV_batch_querysets.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
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
from __future__ import print_function
import pickle
import time
import os
import sys
# os.environ["CUDA_VISIBLE_DEVICES"] = "0"
sys.path.append('./util')
from util.vocab import Concept
import torch
torch.backends.cudnn.enabled = False
from torch.autograd import Variable
import evaluation
from model import ITV
import util.data_provider as data
from util.vocab import Vocabulary
from util.text2vec import get_text_encoder
import h5py
import logging
import json
import numpy as np
import argparse
from util.util import read_dict,Progbar,makedirsforfile, checkToSkip
from util.bigfile import BigFile
from util.constant import ROOT_PATH
def parse_args():
# Hyper Parameters
parser = argparse.ArgumentParser()
parser.add_argument('testCollection', type=str, help='test collection')
parser.add_argument('--rootpath', type=str, default=ROOT_PATH, help='path to datasets.')
parser.add_argument('--overwrite', type=int, default=0, choices=[0,1], help='overwrite existed file. (default: 0)')
parser.add_argument('--batch_size', default=128, type=int, help='Size of a training mini-batch.')
parser.add_argument('--workers', default=10, type=int, help='Number of data loader workers.')
parser.add_argument('--logger_name', default='runs', help='Path to save the model and Tensorboard log.')
parser.add_argument('--checkpoint_name', default='model_best.pth.match.tar', type=str, help='name of checkpoint (default: model_best.pth.tar)')
parser.add_argument('--query_sets', type=str, default='tv16.avs.txt,tv17.avs.txt,tv18.avs.txt',
help='test query sets, tv16.avs.txt,tv17.avs.txt,tv18.avs.txt for TRECVID 16/17/18.')
parser.add_argument('--query_num_all', type=int, default=90,
help='number of querys for test.')
parser.add_argument('--query_sigmoid_threshold', type=float, default=0.99,
help='threshold for concept selection.')
parser.add_argument('--concept_selection', type=str, default=None,help='way for concept selection')
args = parser.parse_args()
return args
def encode_data(model, data_loader, return_ids=True,sigmoid=True,dim=11147):
"""Encode all videos and captions loadable by `data_loader`
"""
# numpy array to keep all the embeddings
embeddings = None
sigmoid_outs = None
ids = ['']*len(data_loader.dataset)
pbar = Progbar(len(data_loader.dataset))
for i, (datas, idxs, data_ids) in enumerate(data_loader):
# compute the embeddings
if sigmoid:
emb,sigmoid_out = model(datas,sigmoid_output=sigmoid)
prob_mask = sigmoid_out<=0.5
sigmoid_out[prob_mask]=0
else:
emb = model(datas,sigmoid_output=sigmoid)
# initialize the numpy arrays given the size of the embeddings
if embeddings is None:
embeddings = np.zeros((len(data_loader.dataset), emb.size(1)))
sigmoid_outs = np.zeros((len(data_loader.dataset),dim))
# preserve the embeddings by copying from gpu and converting to numpy
embeddings[idxs] = emb.data.cpu().numpy().copy()
if sigmoid:
sigmoid_outs[idxs] = sigmoid_out.data.cpu().numpy().copy()
for j, idx in enumerate(idxs):
ids[idx] = data_ids[j]
del datas
pbar.add(len(idxs))
if sigmoid:
return embeddings, sigmoid_outs,ids
else:
if return_ids == True:
return embeddings, ids,
else:
return embeddings
def compute_distances(model, data_loader,query_embs,bert_sim_concept_vectors,iw2v_concept_vectors,w2v_concept_vectors,nonUL_concept_vectors_ori_all,UL_concept_vectors_combined_all,return_ids=True,sigmoid=True,dim=11147):
"""Encode all videos and captions loadable by `data_loader`
"""
# numpy array to keep all the embeddings
embedding_matrix = None
bert_sim_concept_matrix_all = None
iw2v_concept_matrix_all = None
w2v_sim_concept_matrix_all = None
nonUL_concept_decoded_matrix_all = None
UL_concept_decoded_combined_matrix_all = None
ids = ['']*len(data_loader.dataset)
pbar = Progbar(len(data_loader.dataset))
for i, (datas, idxs, data_ids) in enumerate(data_loader):
# compute the embeddings
if sigmoid:
emb,sigmoid_out = model(datas)
prob_mask = sigmoid_out<=0.5
sigmoid_out[prob_mask]=0
else:
emb = model(datas)
# initialize the numpy arrays given the size of the embeddings
if embedding_matrix is None:
embedding_matrix = np.zeros([query_embs.shape[0],len(data_loader.dataset)])
bert_sim_concept_matrix_all = np.zeros([query_embs.shape[0],len(data_loader.dataset)])
iw2v_concept_matrix_all = np.zeros([query_embs.shape[0],len(data_loader.dataset)])
w2v_sim_concept_matrix_all = np.zeros([query_embs.shape[0],len(data_loader.dataset)])
nonUL_concept_decoded_matrix_all = np.zeros([query_embs.shape[0],len(data_loader.dataset)])
UL_concept_decoded_combined_matrix_all = np.zeros([query_embs.shape[0],len(data_loader.dataset)])
# preserve the embeddings by copying from gpu and converting to numpy
embedding_matrix[:,idxs] = query_embs.dot(emb.data.cpu().numpy().copy().T)
if sigmoid:
bert_sim_concept_matrix_all[:, idxs] = bert_sim_concept_vectors.dot(sigmoid_out.data.cpu().numpy().copy().T)
iw2v_concept_matrix_all[:, idxs] = iw2v_concept_vectors.dot(sigmoid_out.data.cpu().numpy().copy().T)
w2v_sim_concept_matrix_all[:, idxs] = w2v_concept_vectors.dot(sigmoid_out.data.cpu().numpy().copy().T)
nonUL_concept_decoded_matrix_all[:, idxs] = nonUL_concept_vectors_ori_all.dot(sigmoid_out.data.cpu().numpy().copy().T)
UL_concept_decoded_combined_matrix_all[:, idxs] = UL_concept_vectors_combined_all.dot(sigmoid_out.data.cpu().numpy().copy().T)
for j, idx in enumerate(idxs):
ids[idx] = data_ids[j]
del datas
pbar.add(len(idxs))
if sigmoid:
return embedding_matrix, bert_sim_concept_matrix_all,iw2v_concept_matrix_all,w2v_sim_concept_matrix_all,nonUL_concept_decoded_matrix_all,UL_concept_decoded_combined_matrix_all,ids
else:
if return_ids == True:
return embedding_matrix, ids,
else:
return embedding_matrix
def main():
opt = parse_args()
print(json.dumps(vars(opt), indent=2))
rootpath = opt.rootpath
testCollection = opt.testCollection
resume = os.path.join(opt.logger_name, opt.checkpoint_name)
concept_selection = opt.concept_selection
# encoder_resume_name = os.path.join(opt.encoder_resume_name, opt.checkpoint_name)
if not os.path.exists(resume):
logging.info(resume + ' not exists.')
sys.exit(0)
checkpoint = torch.load(resume)
start_epoch = checkpoint['epoch']
matching_best_rsum = checkpoint['matching_best_rsum']
classification_best_rsum = checkpoint['classification_best_rsum']
print("=> loaded checkpoint '{}' (epoch {}, matching_best_rsum {},classification_best_rsum {})"
.format(resume, start_epoch, matching_best_rsum, classification_best_rsum))
options = checkpoint['opt']
model = ITV(options)
model.load_state_dict(checkpoint['model'])
model.vid_encoder.eval()
model.text_encoder.eval()
model.unify_decoder.eval()
trainCollection = options.trainCollection
valCollection = options.valCollection
visual_feat_file = BigFile(os.path.join(rootpath, testCollection, 'FeatureData', options.visual_feature))
assert options.visual_feat_dim == visual_feat_file.ndims
if 'motion_feature' in options:
motion_feat_file = BigFile(os.path.join(rootpath, testCollection, 'FeatureData', options.motion_feature))
assert options.motion_feat_dim == motion_feat_file.ndims
video2frames = read_dict(os.path.join(rootpath, testCollection, 'FeatureData', options.visual_feature,'video2frames.txt'))
## set bow vocabulary and encoding
bow_vocab_file = os.path.join(rootpath, options.trainCollection, 'TextData', 'vocabulary', 'bow',
options.vocab + '.pkl')
bow_vocab = pickle.load(open(bow_vocab_file, 'rb'))
bow2vec = get_text_encoder('bow')(bow_vocab)
options.bow_vocab_size = len(bow_vocab)
## set rnn vocabulary
rnn_vocab_file = os.path.join(rootpath, options.trainCollection, 'TextData', 'vocabulary', 'rnn',
options.vocab + '.pkl')
rnn_vocab = pickle.load(open(rnn_vocab_file, 'rb'))
options.vocab_size = len(rnn_vocab)
# set concept concept list for multi-label classification
concept=model.concept
concept2vec = get_text_encoder('bow')(concept, istimes=0)
options.concept_list_size = len(concept)
concept_file = os.path.join(rootpath, options.trainCollection, 'TextData', 'concept', 'concept_frequency_count_gt' + str(
options.concept_fre_threshold)+'.'+options.concept_bank+'.txt')
contradiction_file = concept_file+'.contradict.contradict_pairs'
with open(contradiction_file, 'r') as reader:
lines = reader.readlines()
for line in lines:
if line.find('//') < 0:
concept.add_contradict(line)
visual_loader = data.get_vis_data_loader(visual_feat_file,motion_feat_file, opt.batch_size, opt.workers, video2frames)
modelname = opt.logger_name[opt.logger_name.index('run'):]
query_num = opt.query_num_all
concept_dim = len(concept)
thetas = [0.0,0.3,0.5,1.0]
output_dir = resume.replace(trainCollection, testCollection)
query_sets = []
queryset2queryidxs = {}
queryidxstart = 0
query_embs_all = np.zeros([query_num,options.visual_mapping_layers[1]])
bert_sim_concept_vectors_all= np.zeros([query_num,options.concept_list_size])
iw2v_concept_vectors_all = np.zeros([query_num, options.concept_list_size])
w2vsim_query_concept_vectors_all = np.zeros([query_num, options.concept_list_size])
nonUL_query_decoded_combined_concept_vectors_all= np.zeros([query_num, options.concept_list_size])
UL_query_decoded_combined_concept_vectors_all= np.zeros([query_num, options.concept_list_size])
query_ids_all = []
query_sigmoid_threshold = opt.query_sigmoid_threshold
for query_set in opt.query_sets.strip().split(','):
query_sets.append(query_set)
output_dir_tmp = output_dir.replace(valCollection, '%s/%s/%s' % (query_set, trainCollection, valCollection))
output_dir_tmp = output_dir_tmp.replace('/%s/' % options.cv_name, '/results/')
pred_result_file = os.path.join(output_dir_tmp, 'id.sent')
print(pred_result_file)
if checkToSkip(pred_result_file, opt.overwrite):
continue
try:
makedirsforfile(pred_result_file)
except Exception as e:
print(e)
# data loader prepare
query_file = os.path.join(rootpath, testCollection, 'TextData', query_set)
# set data loader
query_loader = data.get_txt_data_loader(query_file, rnn_vocab, bow2vec, opt.batch_size, opt.workers)
query_selections=None
print("matched concepts are from file:"+query_file+'\n')
##way1 direct match for concept selection
# concept_vectors, query_ids2,query_selections = evaluation.get_concept_vector(query_file, concept2vec)
##way2 similarity match for concept selection
query_file = os.path.join(rootpath, testCollection, 'TextData', query_set)
bert_sim_concept_vectors, query_ids2_bert_sim,query_selections_bert_sim = evaluation.get_concept_vector_BySim(query_file, concept2vec=concept2vec)
savename = query_file + '.query_selection.bert_sim.lemma'
lines=[]
for i, iquery_id in enumerate(query_ids2_bert_sim):
lines.append("#%s:%s\n" % (iquery_id, query_selections_bert_sim[i]))
print("#%s:%s" % (iquery_id, query_selections_bert_sim[i]))
with open(savename, 'w') as writer:
writer.writelines(lines)
print('save in %s\n' % savename)
query_file = os.path.join(rootpath, testCollection, 'TextData', query_set)
# #way3 extraction from the saved similarity concept selection
if concept_selection is not None:
nonUL_concept_file = os.path.join(rootpath, testCollection, 'TextData',query_set+'.decoded_egt%fAndContraryConcept.lemma.enhanced.%s'%(query_sigmoid_threshold,concept_selection))
nonUL_concept_vectors, nonUL_query_ids2,nonUL_query_selections = evaluation.get_concept_vector(nonUL_concept_file, concept2vec,spliter=',')
##print concept selection:
lines = []
for i,iquery_id in enumerate(nonUL_query_ids2):
lines.append("#%s:%s\n"%(iquery_id,nonUL_query_selections[i]))
print("#%s:%s"%(iquery_id,nonUL_query_selections[i]))
start = time.time()
query_concept_sigmoid=None
# query_embs, query_ids = encode_data(model.embed_txt, query_loader,sigmoid=False,dim=concept_dim)
query_embs, query_concept_sigmoid,query_ids = encode_data(model.embed_txt, query_loader,sigmoid=True,dim=concept_dim)
print("encode text time: %.3f s" % (time.time() - start))
query_concept_sigmoid_new = np.zeros([len(query_ids),query_concept_sigmoid.shape[1]])
query_concept_sigmoid_combined_new = np.zeros([len(query_ids),query_concept_sigmoid.shape[1]])
# query_concept_sigmoid_minusContrary = np.zeros([len(query_ids),query_concept_sigmoid.shape[1]])
if query_concept_sigmoid is not None:
lines = []
for i,iquery_id in enumerate(query_ids):
contrary_words=[]
query_sigmoid = query_concept_sigmoid[i,:]
query_concept_decoding =[concept2vec.vocab.idx2concept[idx] for idx in np.where(query_sigmoid >= query_sigmoid_threshold)[0]]
concept_mapping =[]
for idx in np.where(query_sigmoid >= query_sigmoid_threshold)[0]:
word = concept.idx2concept[idx]
# if word=="video" or word=="clip"or word=="show"or word=="something"or word=="someone":
if word=="video" or word=="clip":
continue
concept_mapping.append(word+':%.3f'%query_sigmoid[idx])
query_concept_sigmoid_new[i,idx] =query_sigmoid[idx]
query_concept_sigmoid_combined_new[i,idx] =query_sigmoid[idx]
# query_concept_sigmoid_minusContrary[i,idx] =query_sigmoid[idx]
##union with similarity search
for word in query_selections_bert_sim[i].split(','):
if word in concept.concept2contractconcept.keys():
for icontrary in concept.concept2contractconcept[word].split(','):
if not icontrary in query_selections_bert_sim[i].split(','):
contrary_words = contrary_words + [icontrary]
if len(word)>0 and (word not in query_concept_decoding):
phraseidx = concept2vec.vocab.concept2idx[word]
concept_mapping=concept_mapping+[concept2vec.vocab.idx2concept[phraseidx]+':%.3f'%query_sigmoid[phraseidx]]
query_concept_sigmoid_combined_new[i, phraseidx] = query_sigmoid[phraseidx]
# query_concept_sigmoid_minusContrary[i, phraseidx] = query_sigmoid[phraseidx]
contrary_words = list(set(contrary_words))
for contrary_word in contrary_words:
if contrary_word not in query_selections_bert_sim[i].split(','):
if contrary_word in concept.concept2idx:
contrary_idx = concept.concept2idx[contrary_word]
contrary_prob = query_sigmoid[contrary_idx]
# query_sigmoidy_concept_sigmoid_minusContrary[i, contrary_idx] = contrary_prob-1.0
concept_mapping = concept_mapping + ['--%s:%.3f' % (contrary_word, contrary_prob)]
print("#%s:%s" % (iquery_id, ','.join(concept_mapping)))
lines.append(iquery_id+' '+','.join(concept_mapping)+'\n')
savename=query_file+'.decoded_egt%fAndContraryConcept.lemma.enhanced.%s'%(query_sigmoid_threshold,modelname)
with open(savename,'w') as writer:
writer.writelines(lines)
print('save in %s\n'%savename)
queryidxs = np.arange(queryidxstart, queryidxstart + len(query_ids))
queryset2queryidxs[query_set] =queryidxs
query_embs_all[queryidxs,:]=query_embs
bert_sim_concept_vectors_all[queryidxs,:]=bert_sim_concept_vectors
if concept_selection is not None:
nonUL_query_decoded_combined_concept_vectors_all[queryidxs,:]=nonUL_concept_vectors
UL_query_decoded_combined_concept_vectors_all[queryidxs,:]=query_concept_sigmoid_combined_new
queryidxstart=queryidxstart+len(query_ids)
query_ids_all = query_ids_all+query_ids
##make sure query_ids and query_ids2 are the same
embedding_matrix_all = None
start = time.time()
if embedding_matrix_all is None:
embedding_matrix_all, bert_sim_concept_matrix_all,iw2v_concept_matrix_all,w2v_sim_concept_matrix_all,nonUL_concept_decoded_matrix_all,UL_concept_decoded_combined_matrix_all,vis_ids = compute_distances(model.embed_vis, visual_loader,
query_embs_all, bert_sim_concept_vectors_all,iw2v_concept_vectors_all,w2vsim_query_concept_vectors_all,nonUL_query_decoded_combined_concept_vectors_all,
UL_query_decoded_combined_concept_vectors_all,
sigmoid=True, dim=concept_dim)
print("encode image time: %.3f s" % (time.time() - start))
for query_set in query_sets:
output_dir_tmp = output_dir.replace(valCollection, '%s/%s/%s' % (query_set, trainCollection, valCollection))
output_dir_tmp = output_dir_tmp.replace('/%s/' % options.cv_name, '/results/')
query_idx =[]
for i,sample in enumerate(queryset2queryidxs[query_set]):
query_idx.append(int(sample))
query_ids = np.array(query_ids_all)[query_idx]
embedding_matrix = embedding_matrix_all[query_idx,:]
nanidx = np.isnan(embedding_matrix)
embedding_matrix[nanidx] = 0
rows_min = np.min(embedding_matrix, 1)[:, np.newaxis]
rows_max = np.max(embedding_matrix, 1)[:, np.newaxis]
print('embedding matrix min:%.2f, max:%.2f\n'%(np.min(rows_min),np.max(rows_max)))
embedding_matrix_norm = (embedding_matrix - rows_min) / ((rows_max - rows_min))
del embedding_matrix
#UL concept sim
UL_concept_decoded_matrix= UL_concept_decoded_combined_matrix_all[query_idx,:]
nanidx = np.isnan(UL_concept_decoded_matrix)
UL_concept_decoded_matrix[nanidx] = 0
rows_min = np.min(UL_concept_decoded_matrix, 1)[:, np.newaxis]
rows_max = np.max(UL_concept_decoded_matrix, 1)[:, np.newaxis]
print('concept matrix min:%.2f, max:%.2f\n'%(np.min(rows_min),np.max(rows_max)))
UL_concept_decoded_matrix_norm = (UL_concept_decoded_matrix - rows_min) / ((rows_max - rows_min))
del UL_concept_decoded_matrix
print("mapping concept time: %.3f s" % (time.time() - start))
for theta in thetas:
pred_result_file = os.path.join(output_dir_tmp, 'id.sent.sim.%.2f.combinedDecodedConcept_theta'%(query_sigmoid_threshold) + str(theta).replace('.', '_') + '_score')
print(pred_result_file)
combined_matrix = (1 - theta) * embedding_matrix_norm + (theta) * UL_concept_decoded_matrix_norm
combined_inds = np.argsort(combined_matrix, axis=1)
with open(pred_result_file, 'w') as fout:
for index in range(combined_inds.shape[0]):
ind = combined_inds[index][::-1]
fout.write(query_ids[index] + ' ' + ' '.join(
[vis_ids[i] + ' %s' % combined_matrix[index][i] for i in ind]) + '\n')
if __name__ == '__main__':
main()