-
Notifications
You must be signed in to change notification settings - Fork 8
/
train_proposal_pair2.py
367 lines (313 loc) · 14.5 KB
/
train_proposal_pair2.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
import os
import numpy as np
import cPickle as pickle
import tempfile
import h5py
import time
import json
import sys
os.environ['GLOG_minloglevel'] = '2'
import caffe
caffe.set_mode_gpu()
from utils import recall_vs_iou_thresholds, convert, nms, get_gt
def static_vars(**kwargs):
def decorate(func):
for k in kwargs:
setattr(func, k, kwargs[k])
return func
return decorate
def create_net():
f = tempfile.NamedTemporaryFile(mode='w+', delete=False)
f.write(
"""
name: 'pair_nn'
#level base
layer { name: "data" type: "DummyData" top: "lv" dummy_data_param { shape { dim: 1024 dim: 1 dim: 12 dim: 202 } } }
layer { name: 'conv' type: 'Convolution' bottom: "lv" top: "conv" convolution_param { engine: CAFFE num_output: 1 kernel_w: 1 kernel_h: 5 weight_filler {type: "xavier"} } }
layer { name: "relu" type: "ReLU" bottom: "conv" top: "conv" }
layer { name: "pool" type: "Pooling" bottom: "conv" top: "pool" pooling_param { pool: AVE kernel_h: 3 kernel_w: 1 stride_h: 1 stride_w: 1} }
#level upper
layer { name: "data_upper" type: "DummyData" top: "lv_upper" dummy_data_param { shape { dim: 1024 dim: 1 dim: 12 dim: 202 } } }
layer { name: 'conv_upper' type: 'Convolution' bottom: "lv_upper" top: "conv_upper" convolution_param { engine: CAFFE num_output: 1 kernel_w: 1 kernel_h: 5 weight_filler {type: "xavier"} } }
layer { name: "relu_upper" type: "ReLU" bottom: "conv_upper" top: "conv_upper" }
layer { name: "pool_upper" type: "Pooling" bottom: "conv_upper" top: "pool_upper" pooling_param { pool: AVE kernel_h: 3 kernel_w: 1 stride_h: 1 stride_w: 1} }
layer { name: "concat" type: "Concat" bottom: "pool" bottom: "pool_upper" top: "ip1" concat_param { axis: 1 } }
#proposal loss
layer { name: "label" type: "DummyData" top: "label" dummy_data_param { shape { dim: 1024 dim: 1 dim: 1 dim: 1 } } }
layer { name: "ip1" type: "InnerProduct" bottom: "ip1" top: "ip2" inner_product_param { num_output: 500 weight_filler {type: "xavier"} } }
layer { name: "ip2" type: "InnerProduct" bottom: "ip2" top: "rs" inner_product_param { num_output: 2 weight_filler {type: "xavier"} } }
layer { name: "loss" type: "SoftmaxWithLoss" bottom: "rs" bottom: "label" top: "loss" loss_weight: 1 loss_param {ignore_label: -1 normalize: true} }
""")
f.close()
return f.name
def create_deploy():
f = tempfile.NamedTemporaryFile(mode='w+', delete=False)
f.write(
"""
name: 'pair_nn'
#level base
layer { name: "data" type: "DummyData" top: "lv" dummy_data_param { shape { dim: 1 dim: 1 dim: 12 dim: 202 } } }
layer { name: 'conv' type: 'Convolution' bottom: "lv" top: "conv" convolution_param { engine: CAFFE num_output: 1 kernel_w: 1 kernel_h: 5 weight_filler {type: "xavier"} } }
layer { name: "relu" type: "ReLU" bottom: "conv" top: "conv" }
layer { name: "pool" type: "Pooling" bottom: "conv" top: "pool" pooling_param { pool: AVE kernel_h: 3 kernel_w: 1 stride_h: 1 stride_w: 1} }
#level upper
layer { name: "data_upper" type: "DummyData" top: "lv_upper" dummy_data_param { shape { dim: 1 dim: 1 dim: 12 dim: 202 } } }
layer { name: 'conv_upper' type: 'Convolution' bottom: "lv_upper" top: "conv_upper" convolution_param { engine: CAFFE num_output: 1 kernel_w: 1 kernel_h: 5 weight_filler {type: "xavier"} } }
layer { name: "relu_upper" type: "ReLU" bottom: "conv_upper" top: "conv_upper" }
layer { name: "pool_upper" type: "Pooling" bottom: "conv_upper" top: "pool_upper" pooling_param { pool: AVE kernel_h: 3 kernel_w: 1 stride_h: 1 stride_w: 1} }
layer { name: "concat" type: "Concat" bottom: "pool" bottom: "pool_upper" top: "ip1" concat_param { axis: 1 } }
#proposal
layer { name: "ip1" type: "InnerProduct" bottom: "ip1" top: "ip2" inner_product_param { num_output: 500 } }
layer { name: "ip2" type: "InnerProduct" bottom: "ip2" top: "rs" inner_product_param { num_output: 2 } }
layer { name: "loss" type: "Softmax" bottom: "rs" top: "loss"}
""")
f.close()
return f.name
def create_solver(netf):
f = tempfile.NamedTemporaryFile(mode='w+', delete=False)
f.write("""
net: '""" + netf + """'
base_lr: 0.1
momentum: 0.9
weight_decay: 0.00005
lr_policy: 'step'
stepsize: 1000
display: 20000
snapshot: 1000
max_iter: 20000
snapshot_prefix: "rank_pair2"
""")
f.close()
return f.name
def stats(label, pred, hist=False):
printout = ''
if hist==True:
y = label.squeeze()
ind = np.nonzero(y > 0)
neg = np.nonzero(y == 0)
_, ct = np.unique(ind[y.ndim-1], return_counts=True)
printout += 'Hist ({}|{}) '.format(len(neg[y.ndim-1]), ct)
fg = np.sum(label >= 1)
bg = np.sum(label == 0)
total = fg+bg
corr = np.sum(label == pred)
tfg = np.sum(np.logical_and(label>=1,pred>=1))
tbg = np.sum(np.logical_and(label==0,pred==0))
ffg = np.sum(np.logical_and(label==0,pred>=1))
fbg = np.sum(np.logical_and(label>=1,pred==0))
printout += 'Acc: {:.4} \n'.format(float(corr)/(fg+bg))
printout += 'TF {:.2} FF {:.2} \nFB {:.2} TB {:.2} \n'.format(float(tfg)/total, float(ffg)/total, float(fbg)/total, float(tbg)/total, )
return printout
def extract_feature(feature, proposal, num_frame, duration, PYRAMID=[1,2,4,8,16,16]):
from tsn import extract
data = []
for ind, num in enumerate(PYRAMID):
feat = extract(feature, num_frame, duration, proposal[ind,0], proposal[ind,1], num, zero_padding=True)
if feat is None:
return None
data.append(feat)
return np.vstack(data)
# Sample propsoal pairs with a pyramid around specific level (a list) if given
def sample_pairs(proposals, lv=None, scale=2):
pairs = []
nlv = proposals.shape[0]
if lv is not None:
plist = lv
else:
plist = range(nlv)
for i in plist:
top = np.zeros((2, 4))
top[0, :] = proposals[i, :]
center = (proposals[i, 0] + proposals[i, 1]) / 2
length = (proposals[i, 1] - proposals[i, 0]) / 2
top[1, 0] = center - length*scale
top[1, 1] = center + length*scale
pairs.append(top)
#pairs.append(proposals[i:i+2,:])
return pairs
# Dynamic sample batch size, it is little tricky to do considering a pyramid has n levels.
# Use some hard negatives (all levels 0), the effective number will multiple by n levels
@static_vars(list={})
def load_data_pairs(gt, proposals, batch=1024, random=256, hard_neg=256, maxnump=32, PYRAMID=[12,12], thr_ious=[0.7,0.3], subset='training'):
x = np.zeros((batch, 1, np.sum(PYRAMID), 202))
y = np.ones((batch, 1, 1, 1))*-1
target_pos = batch-hard_neg-random
pos_count = 0
neg_count = 0
rand_count = 0
count = 0
if not load_data_pairs.list.has_key(subset):
# behavior like caffe preload
load_data_pairs.list[subset] = []
for vid, vitem in gt.iteritems():
if vitem['subset'] == subset:
load_data_pairs.list[subset].append(vid)
#file_ct = 0
list = load_data_pairs.list[subset]
for pos, vid in enumerate(list):
vitem = gt[vid]
if not os.path.exists(FEATURE + '/feat/%s.h5' % vid) or not proposals.has_key(vid) or len(proposals[vid])==0:
continue
with h5py.File(FEATURE + '/feat/%s.h5' % vid, 'r') as hf:
fg = np.asarray(hf['fg'])
bg = np.asarray(hf['bg'])
feat = np.hstack([fg,bg])
with h5py.File(FEATURE + '/flow/%s.h5' % vid, 'r') as hf:
fg2 = np.asarray(hf['fg'])
bg2 = np.asarray(hf['bg'])
feat2 = np.hstack([fg2,bg2])
feat = feat + feat2
num_frame = vitem['numf']
duration = vitem['duration']
proposal = np.asarray(proposals[vid])
max_ious = proposal[:, :, 2].max(axis=1)
max_ious_lv = proposal[:, :, 2].argmax(axis=1)
#positive per level
pos_indexes = np.nonzero(max_ious > thr_ious[0])[0]
#hard negative
neg_indexes = np.nonzero(max_ious < thr_ious[1])[0]
# sample postives around the best level
pos_pairs = []
for i in range(len(pos_indexes)):
ind = pos_indexes[i]
lv = max_ious_lv[ind]
pos_pairs += sample_pairs(proposal[ind,], [lv])
np.random.shuffle(pos_pairs)
for pair in pos_pairs[:min(target_pos,maxnump)]:
if pos_count < target_pos and count < batch:
data = extract_feature(feat, pair, num_frame, duration, PYRAMID=PYRAMID)
if data is not None and data.shape[0]==np.sum(PYRAMID):
x[count, ...].flat = data.flat
y[count, ...] = 1
pos_count += 1
count +=1
# sample random negative around around positive
rand_pairs = []
for i in range(len(pos_indexes)):
ind = pos_indexes[i]
lv = max_ious_lv[ind]
rand_pairs += sample_pairs(proposal[ind,], [i for i in range(6) if i!=lv])
np.random.shuffle(rand_pairs)
for pair in rand_pairs[:min(random,maxnump)]:
if rand_count<random and count < batch:
data = extract_feature(feat, pair, num_frame, duration, PYRAMID=PYRAMID)
if data is not None and data.shape[0] == np.sum(PYRAMID):
x[count, ...].flat = data.flat
y[count, ...] = 0
rand_count += 1
count += 1
# sample hard negatives
neg_pairs = []
for i in range(len(neg_indexes)):
ind = neg_indexes[i]
neg_pairs += sample_pairs(proposal[ind,])
np.random.shuffle(neg_pairs)
for pair in neg_pairs[:min(hard_neg,maxnump)]:
if neg_count < hard_neg and count < batch:
data = extract_feature(feat, pair, num_frame, duration, PYRAMID=PYRAMID)
if data is not None and data.shape[0]==np.sum(PYRAMID):
x[count, ...].flat = data.flat
y[count, ...] = 0
neg_count += 1
count += 1
#file_ct += 1
if count==batch:
roll = list[pos:] + list[:pos]
load_data_pairs.list[subset] = roll
break
#print 'Files opened %d'%file_ct
#print subset, neg_count
#print max_ious
return x, y
def rank_propsal_pairs(gt, proposals, model, PYRAMID=[12,12], feature_path='/home/DATASETS/actnet/tsn_score/'):
# load model
nf = create_deploy()
sf = create_solver(nf)
solver = caffe.get_solver(sf)
solver.net.copy_from(model)
# classifying
ACC = []
ranked_proposals = {}
for vid, proposal in proposals.iteritems():
# print vid
if not os.path.exists(feature_path + '/feat/%s.h5' % vid) or len(proposals)==0:
continue
with h5py.File(feature_path + '/feat/%s.h5' % vid, 'r') as hf:
fg = np.asarray(hf['fg'])
bg = np.asarray(hf['bg'])
feat = np.hstack([fg,bg])
with h5py.File(feature_path + '/flow/%s.h5' % vid, 'r') as hf:
fg2 = np.asarray(hf['fg'])
bg2 = np.asarray(hf['bg'])
feat2 = np.hstack([fg2,bg2])
feat = feat + feat2
num_frame = gt[vid]['numf']
duration = gt[vid]['duration']
proposal = np.asarray(proposal)
ranked_proposals[vid] = []
for i in range(proposal.shape[0]):
current = np.copy(proposal[i,...])
current[:,2] = 0
pairs = sample_pairs(current)
for pos, pair in enumerate(pairs):
x = extract_feature(feat, pair, num_frame, duration, PYRAMID=PYRAMID)
if x is not None and x.shape[0]==np.sum(PYRAMID):
x = x[np.newaxis, np.newaxis, ...]
solver.net.blobs['lv'].data[...] = x[:,:,0:12,:]
solver.net.blobs['lv_upper'].data[...] = x[:,:,12:24,:]
rs = solver.net.forward()
prob = rs['loss'][0,0,...]
ACC.append([[pair[1,3]>0.5],[prob<0.5]])
current[pos,2] = 1-prob
else:
current[pos,2] = 0
ranked_proposals[vid].append(current)
ACC = np.hstack(ACC)
print 'Test on validation set ' + stats(ACC[0,:], ACC[1,:])
return ranked_proposals
if __name__ == '__main__':
nf = create_net()
sf = create_solver(nf)
solver = caffe.get_solver(sf)
FEATURE = os.environ['ACTNET_HOME'] + '/tsn_score/'
with open('actNet200-V1-3.pkl', 'rb') as f:
gt = pickle.load(f)['database']
with open('train_proposals.pkl', 'rb') as f:
train_proposals = pickle.load(f)
with open('val_proposals.pkl', 'rb') as f:
val_proposals = pickle.load(f)
max_epoch = 1 #50
max_iter = 1000 #1000
print 'Train Proposal Ranker'
for ep in range(max_epoch):
loss = 0
for it in range(max_iter):
############# Train #############
x, y = load_data_pairs(gt, train_proposals)
solver.net.blobs['lv'].data[...] = x[:, :, 0:12, :]
solver.net.blobs['lv_upper'].data[...] = x[:, :, 12:24, :]
solver.net.blobs['label'].data[...] = y
solver.step(1)
loss += solver.net.blobs['loss'].data
if (it+1)%50 == 0:
print time.strftime('[%Y%-m-%d %H:%M:%S] ') + 'Iter %d '%(max_iter*ep+it+1),
print 'Train Loss: %f | %d %d'%(loss/50, np.sum(y==0), np.sum(y>=1))
loss = 0
model = 'rank_pair2_iter_{}.caffemodel'.format(max_iter*(ep+1))
ranked_val_proposals = rank_propsal_pairs(gt, val_proposals, model, feature_path=FEATURE)
with open('ranked_val_proposals.pkl', 'wb') as f:
pickle.dump(ranked_val_proposals, f)
proposal_at_1 = {'s-init':[],'s-end':[],'score':[],'label':[],'video-id':[]}
for vid, proposal in ranked_val_proposals.iteritems():
proposal = np.asarray(proposal)
proposal = proposal.reshape((proposal.shape[0]*proposal.shape[1], proposal.shape[2]))
keep_ind = nms(proposal, proposal[:,2], 0.45)
proposal = proposal[keep_ind,:]
proposal_at_1['s-init'].append(proposal[0,0])
proposal_at_1['s-end'].append(proposal[0,1])
proposal_at_1['score'].append(proposal[0,2])
proposal_at_1['video-id'].append('v_'+vid)
gt = get_gt(gt)
iou_thrs = np.arange(0.1, 1.0, 0.1)
recall_at_1 = recall_vs_iou_thresholds(convert(proposal_at_1), gt, iou_threshold=iou_thrs)
print np.array_str(np.vstack([iou_thrs, recall_at_1]), precision=4, suppress_small=True)