-
Notifications
You must be signed in to change notification settings - Fork 0
/
evaluation.py
419 lines (346 loc) · 16.7 KB
/
evaluation.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
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
import logging
import torch
import numpy as np
import time
import wandb
from scipy.spatial import distance
from util.metric import getScorer
from util.util import AverageMeter, LogCollector,Progbar
from sentence_transformers import SentenceTransformer,util
def l2norm(X):
"""L2-normalize columns of X
"""
norm = np.linalg.norm(X, axis=1, keepdims=True)
return 1.0 * X / norm
def cal_error(videos, captions, measure='cosine'):
if measure == 'cosine':
captions = l2norm(captions)
videos = l2norm(videos)
errors = -1*np.dot(captions, videos.T)
elif measure == 'euclidean':
errors = distance.cdist(captions, videos, 'euclidean')
return errors
def v2t(c2v, n_caption=5):
"""
Videos->Text (Video-to-Text Retrieval)
c2v: (5N, N) matrix of caption to video errors
"""
assert c2v.shape[0] / c2v.shape[1] == n_caption, c2v.shape
ranks = np.zeros(c2v.shape[1])
for i in range(len(ranks)):
d_i = c2v[:, i]
inds = np.argsort(d_i)
rank = np.where((inds/n_caption) == i)[0][0]
ranks[i] = rank
# Compute metrics
r1 = 100.0 * len(np.where(ranks < 1)[0]) / len(ranks)
r5 = 100.0 * len(np.where(ranks < 5)[0]) / len(ranks)
r10 = 100.0 * len(np.where(ranks < 10)[0]) / len(ranks)
medr = np.floor(np.median(ranks)) + 1
meanr = ranks.mean() + 1
return map(float, [r1, r5, r10, medr, meanr])
def t2v(c2v, n_caption=5):
"""
Text->Videos (Text-to-Video Retrieval)
c2v: (5N, N) matrix of caption to video errors
"""
# print("errors matrix shape: ", c2v.shape)
assert c2v.shape[0] / c2v.shape[1] == n_caption, c2v.shape
ranks = np.zeros(c2v.shape[0])
for i in range(len(ranks)):
d_i = c2v[i]
inds = np.argsort(d_i)
rank = np.where(inds == int(i/n_caption))[0][0]
ranks[i] = rank
# Compute metrics
r1 = 100.0 * len(np.where(ranks < 1)[0]) / len(ranks)
r5 = 100.0 * len(np.where(ranks < 5)[0]) / len(ranks)
r10 = 100.0 * len(np.where(ranks < 10)[0]) / len(ranks)
medr = np.floor(np.median(ranks)) + 1
meanr = ranks.mean() + 1
return map(float, [r1, r5, r10, medr, meanr])
# mAP for Text-to-Video Retrieval
def t2v_map(c2v, n_caption=5):
"""
Text->Videos (Text-to-Video Retrieval)
c2v: (5N, N) matrix of caption to video errors
"""
# print("errors matrix shape: ", c2v.shape)
assert c2v.shape[0] / c2v.shape[1] == n_caption, c2v.shape
scorer = getScorer('AP')
perf_list = []
for i in range(c2v.shape[0]):
d_i = c2v[i, :]
labels = [0]*len(d_i)
labels[int(i/n_caption)] = 1
sorted_labels = [labels[x] for x in np.argsort(d_i)]
current_score = scorer.score(sorted_labels)
perf_list.append(current_score)
return np.mean(perf_list)
# mAP for Video-to-Text Retrieval
def v2t_map(c2v, n_caption=5):
"""
Videos->Text (Video-to-Text Retrieval)
c2v: (5N, N) matrix of caption to video errors
"""
# print("errors matrix shape: ", c2v.shape)
assert c2v.shape[0] / c2v.shape[1] == n_caption, c2v.shape
scorer = getScorer('AP')
perf_list = []
for i in range(c2v.shape[1]):
d_i = c2v[:, i]
labels = [0]*len(d_i)
labels[i*n_caption:(i+1)*n_caption] = [1]*n_caption
sorted_labels = [labels[x] for x in np.argsort(d_i)]
current_score = scorer.score(sorted_labels)
perf_list.append(current_score)
return np.mean(perf_list)
def encode_data(model, data_loader, log_step=10, logging=print, return_ids=True):
"""Encode all videos and captions loadable by `data_loader`
"""
batch_time = AverageMeter()
val_logger = LogCollector()
# switch to evaluate mode
end = time.time()
# numpy array to keep all the data
video_embs = None
cap_embs = None
concept_vectors = None
class_vid_outs = None
class_text_outs = None
video_ids = ['']*len(data_loader.dataset)
caption_ids = ['']*len(data_loader.dataset)
captions_ori = [''] * len(data_loader.dataset)
for i, (videos, captions,concept_bows,caption_ori, idxs, cap_ids, vid_ids) in enumerate(data_loader):
# make sure val logger is used
model.logger = val_logger
# compute the embeddings and interpretations
vid_emb, cap_emb = model.forward_matching(videos, captions, True)
class_vid_out,class_text_out = model.forward_classification(vid_emb,cap_emb, True)
# initialize the numpy arrays given the size of the embeddings
if video_embs is None:
video_embs = np.zeros((len(data_loader.dataset), vid_emb.size(1)))
cap_embs = np.zeros((len(data_loader.dataset), cap_emb.size(1)))
concept_vectors = np.zeros((len(data_loader.dataset), concept_bows.size(1)))
class_vid_outs = np.zeros((len(data_loader.dataset), concept_bows.size(1)))
class_text_outs = np.zeros((len(data_loader.dataset), concept_bows.size(1)))
# preserve the embeddings by copying from gpu and converting to numpy
video_embs[idxs] = vid_emb.data.cpu().numpy().copy()
cap_embs[idxs] = cap_emb.data.cpu().numpy().copy()
concept_vectors[idxs] = concept_bows.data.cpu().numpy().copy()
class_vid_outs[idxs] = class_vid_out.data.cpu().numpy().copy()
class_text_outs[idxs] = class_text_out.data.cpu().numpy().copy()
for j, idx in enumerate(idxs):
caption_ids[idx] = cap_ids[j]
video_ids[idx] = vid_ids[j]
captions_ori[idx] = caption_ori[j]
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
del videos, captions
if return_ids == True:
return video_embs, cap_embs,class_vid_outs,class_text_outs, concept_vectors,captions_ori,video_ids, caption_ids
else:
return video_embs, cap_embs,class_vid_outs,class_text_outs,concept_vectors,captions_ori
def eval_multi_label_classifiction_perf(outs,labels,topn=10):
recall = []
pred_class_num = []
gt_num = []
match_num_sum = []
for i in range(outs.shape[0]):
topn_pred_class = np.zeros([outs.shape[1]])
outs_i = np.squeeze(outs[i, :])
label_i = np.squeeze(labels[i, :])
if np.isnan(label_i.sum()):
label_i = np.zeros([len(label_i)])
##get the performance of topn
rankidx = np.argsort(outs_i)[::-1]
rankidx = rankidx[0:topn]
##get the number of predicted class
outs_pred_class_index = np.where(outs_i > 0.5)[0]
outs_pred_num = outs_pred_class_index.size
pred_class_num.append(outs_pred_num)
i_gt_number = np.sum(label_i)
if np.isnan(i_gt_number):
i_gt_number = 0
gt_num.append(i_gt_number)
recallAtk_i = 0.0
if i_gt_number==0:
print('i_gt_number==0')
topn_pred_class[rankidx] = 1
match_candidate = np.multiply(topn_pred_class, label_i)
if len(np.where(match_candidate > 0))>0:
match_index = np.where(match_candidate > 0)[0]
match_num = len(match_index)*1.0
else:
match_num =0
if (match_num > 0):
if i_gt_number>topn:
recallAtk_i = match_num*1.0 / topn
else:
recallAtk_i = match_num * 1.0 / i_gt_number
match_num_sum.append(match_num)
recall.append(recallAtk_i)
return match_num_sum,recall,gt_num,pred_class_num
def eval_ITV(opt, val_loader, model, concept2vec,measure='cosine',topn=10,return_perf=True,capid2conceptvecs=None):
logging.basicConfig(format='%(asctime)s %(message)s', level=logging.INFO)
# compute the encoding for all the validation video and captions
video_embs, cap_embs, class_vid_outs_ori,class_text_outs_ori,concept_vectors_ori,captions_list_ori,video_ids, caption_ids = encode_data(model, val_loader, opt.log_step,
logging.info)
##first evaluation encoder
# we load data as video-sentence pairs
# but we only need to forward each video once for evaluation
# so we get the video set and mask out same videos with feature_mask
feature_mask = []
evaluate_videos = set()
for video_id in video_ids:
feature_mask.append(video_id not in evaluate_videos)
evaluate_videos.add(video_id)
video_embs = video_embs[feature_mask]
class_vid_outs = class_vid_outs_ori[feature_mask]
concept_vectors = concept_vectors_ori[feature_mask]
captions_list = list(np.array(captions_list_ori)[feature_mask])
video_ids = [x for idx, x in enumerate(video_ids) if feature_mask[idx] is True]
class_text_outs_filter = class_text_outs_ori
class_text_outs_filter[class_text_outs_filter<=0.99]=0
class_vid_outs_filter=class_vid_outs
class_vid_outs_filter[class_vid_outs_filter<=0.5]=0
sim_martix=cap_embs.dot(video_embs.T)
c2v_all_errors_emb = cal_error(video_embs, cap_embs, measure)
c2v_all_errors_concept = cal_error(class_vid_outs_filter, class_text_outs_filter, measure)
##first evaluation encoder
if opt.val_metric == "recall" and (not opt.testCollection=='msvd'):
# video retrieval
(r1i, r5i, r10i, medri, meanri) = t2v(c2v_all_errors_emb, n_caption=opt.n_caption)
print(" * Embedding matching")
print(" * Text to video:")
print(" * r_1_5_10: {}".format([round(r1i, 3), round(r5i, 3), round(r10i, 3)]))
print(" * medr, meanr: {}".format([round(medri, 3), round(meanri, 3)]))
print(" * " + '-' * 10)
(c_r1i, c_r5i, c_r10i, c_medri, c_meanri) = t2v(c2v_all_errors_concept, n_caption=opt.n_caption)
print(" * Concept matching")
print(" * Text to video:")
print(" * r_1_5_10: {}".format([round(c_r1i, 3), round(c_r5i, 3), round(c_r10i, 3)]))
print(" * medr, meanr: {}".format([round(c_medri, 3), round(c_meanri, 3)]))
print(" * " + '-' * 10)
# caption retrieval
(r1, r5, r10, medr, meanr) = v2t(c2v_all_errors_emb, n_caption=opt.n_caption)
print(" * Embedding Matching")
print(" * Video to text:")
print(" * r_1_5_10: {}".format([round(r1, 3), round(r5, 3), round(r10, 3)]))
print(" * medr, meanr: {}".format([round(medr, 3), round(meanr, 3)]))
print(" * " + '-' * 10)
(c_r1, c_r5, c_r10, c_medr, c_meanr) = v2t(c2v_all_errors_concept, n_caption=opt.n_caption)
print(" * Concept Matching")
print(" * Video to text:")
print(" * r_1_5_10: {}".format([round(c_r1, 3), round(c_r5, 3), round(c_r10, 3)]))
print(" * medr, meanr: {}".format([round(c_medr, 3), round(c_meanr, 3)]))
print(" * " + '-' * 10)
# record metrics in wandb
wandb.log({"val/r1": r1}, step=model.Eiters)
wandb.log({"val/r5": r5}, step=model.Eiters)
wandb.log({"val/r10": r10}, step=model.Eiters)
wandb.log({"val/medr": medr}, step=model.Eiters)
wandb.log({"val/meanr": meanr}, step=model.Eiters)
wandb.log({"val/r1i": r1i}, step=model.Eiters)
wandb.log({"val/r5i": r5i}, step=model.Eiters)
wandb.log({"val/r10i": r10i}, step=model.Eiters)
wandb.log({"val/medri": medri}, step=model.Eiters)
wandb.log({"val/meanri": meanri}, step=model.Eiters)
elif opt.val_metric == "map":
v2t_map_score =v2t_map(c2v_all_errors_emb, n_caption=opt.n_caption)
t2v_map_score = t2v_map(c2v_all_errors_emb, n_caption=opt.n_caption)
con_v2t_map_score =v2t_map(c2v_all_errors_concept, n_caption=opt.n_caption)
con_t2v_map_score = t2v_map(c2v_all_errors_concept, n_caption=opt.n_caption)
print('embedding v2t_map', v2t_map_score)
print('embedding t2v_map', t2v_map_score)
print('concept v2t_map', con_v2t_map_score)
print('concept t2v_map', con_t2v_map_score)
wandb.log({"val/v2t_map": v2t_map_score}, step=model.Eiters)
wandb.log({"val/t2v_map": t2v_map_score}, step=model.Eiters)
encoder_currscore = 0
if opt.val_metric == "recall" and (not opt.testCollection=='msvd'):
if opt.direction == 'v2t' or opt.direction == 'all':
encoder_currscore += (r1 + r5 + r10)
if opt.direction == 't2v' or opt.direction == 'all':
encoder_currscore += (r1i + r5i + r10i)
elif opt.val_metric == "map":
if opt.direction == 'v2t' or opt.direction == 'all':
encoder_currscore += v2t_map_score
if opt.direction == 't2v' or opt.direction == 'all':
encoder_currscore += t2v_map_score
concept_vectors_text_ori = concept_vectors_ori
if capid2conceptvecs is not None:
concept_vectors_text_ori= np.zeros([concept_vectors_ori.shape[0],concept_vectors_ori.shape[1]])
prob = Progbar(len(caption_ids))
for icap_idx,icap_id in enumerate(caption_ids):
prob.add(1)
concept_vectors_text_ori[icap_idx] = capid2conceptvecs[icap_id]
match_num_sum_vid,recall_vid,gt_vid_num,decoded_concept_vid_num = eval_multi_label_classifiction_perf(class_vid_outs,concept_vectors,topn=topn)
match_num_sum_text,recall_text,gt_text_num,decoded_concept_text_num = eval_multi_label_classifiction_perf(class_text_outs_ori,concept_vectors_text_ori,topn=topn)
print(" * Video: multi-label classification:")
print(" * average matchnum@10(vid,text): {},{}".format(np.average(match_num_sum_vid),np.average(match_num_sum_text)))
print(" * average recall@10(vid,text):%.4f,%.4f"%(np.average(recall_vid),np.average(recall_text)))
print(" * average #decoded_concepts(vid,text): {},{}".format(np.average(decoded_concept_vid_num),np.average(decoded_concept_text_num)))
print(" * average gt_num(vid,text): {},{}".format(np.average(gt_vid_num),np.average(gt_text_num)))
print(" * "+'-'*10)
decoder_cur_vid_recall = np.average(recall_vid)
decoder_cur_text_recall = np.average(recall_text)
wandb.log({"val/#decoded_concepts_vid": np.average(decoded_concept_vid_num)}, step=model.Eiters)
wandb.log({"val/#decoded_concepts_text": np.average(decoded_concept_text_num)}, step=model.Eiters)
wandb.log({"val/decoder_vid_recall": np.average(recall_vid)}, step=model.Eiters)
wandb.log({"val/decoder_text_recall": np.average(recall_text)}, step=model.Eiters)
wandb.log({"val/rsum": encoder_currscore}, step=model.Eiters)
if return_perf:
return encoder_currscore,decoder_cur_vid_recall, decoder_cur_text_recall
else:
return captions_list,captions_list_ori, class_vid_outs,class_text_outs_ori, concept_vectors,concept_vectors_text_ori, sim_martix,video_ids,caption_ids
def get_concept_vector_BySim(query_file,concept2vec=None,concept_bank=None):
"""map all captions into concept vector
concept selection: BERT embedding similarity
"""
stop_word_file = 'stopwords_en.txt'
stop_words = []
with open(stop_word_file, 'rb') as reader:
for word in reader:
word = word.decode().strip()
stop_words.append(word)
stop_words.append('one')
embedder = SentenceTransformer('distilbert-base-nli-mean-tokens')
if concept_bank is None:
concept_bank=[item for item in concept2vec.vocab.concept2idx.keys()]
corpus_embeddings = embedder.encode(concept_bank, convert_to_tensor=True,show_progress_bar=False)
captions = {}
cap_ids = []
with open(query_file, 'r', encoding='iso-8859-1') as cap_reader:
for line in cap_reader.readlines():
cap_id, caption = line.strip().split(' ', 1)
captions[cap_id] = caption
cap_ids.append(cap_id)
if concept2vec is not None:
concept_vectors = np.zeros([len(cap_ids),concept2vec.ndims])
query_selections = [""]*len(cap_ids)
index = 0
pbar = Progbar(len(cap_ids))
for cap_id in cap_ids:
caption = captions[cap_id]
for cap in caption.split():
if cap not in stop_words:
query_embedding = embedder.encode(cap, convert_to_tensor=True,show_progress_bar=False)
cos_scores = util.pytorch_cos_sim(query_embedding, corpus_embeddings)
cos_scores = cos_scores.data.cpu().numpy().squeeze()
top_results = np.argpartition(-cos_scores, range(5))[0:5]
most_similar_idx = top_results[0]
sim_score = cos_scores[most_similar_idx]
if sim_score>0.9:
matched_concept=concept_bank[most_similar_idx]
# print("%s match to concept:%s"%(cap,matched_concept))
if concept2vec is not None:
concept_vectors[index,concept2vec.vocab.concept2idx[matched_concept]]=1
query_selections[index]=query_selections[index]+','+matched_concept
index = index+1
pbar.add(1)
if concept2vec is not None:
return concept_vectors,cap_ids,query_selections
else:
return cap_ids, query_selections