forked from zhujiagang/DTPP
/
eval_tpp_net_kinetics.py
262 lines (197 loc) · 9.28 KB
/
eval_tpp_net_kinetics.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
import argparse
import os
import sys
import math
import cv2
import numpy as np
import multiprocessing
from sklearn.metrics import confusion_matrix
relative_path = "/home/lilin/my_code"
caffe_root = relative_path + '/deeptemporal/lib/caffe-tpp-net-python/'
sys.path.insert(0, caffe_root + 'python')
from pyActionRecog import parse_directory
from pyActionRecog import parse_split_file
from pyActionRecog.utils.video_funcs import default_aggregation_func
net_weights = "/home/lilin/my_code/ucf101_split_3_rgb_flow_models/ucf101_split_3_rgb_tpp_kinetics_st_0.01_lr_0.0001_iter_1500.caffemodel"
memory_enough = False
split = int(net_weights.split('split')[1].split('_')[1])
str = net_weights.split('/')[-1]
if 'rgb' in str:
modality = 'rgb'
else:
modality = 'flow'
if 'ucf101' in str:
dataset = 'ucf101'
else:
dataset = 'hmdb51'
if dataset == 'ucf101':
frame_path = "/home/lilin/UCF-101-result/"
else:
frame_path = relative_path + "/HMDB-51-result/"
# net_proto = relative_path + "/deeptemporal/models/"+ dataset + "/" + modality + "_tpp_delete_dropout_deploy.prototxt"
net_proto = relative_path + "/deeptemporal/models/"+ dataset + "/" + modality + "_kinetics_tpp_p124_deploy.prototxt"
rgb_prefix ='img_'
flow_x_prefix ='flow_x_'
flow_y_prefix ='flow_y_'
num_frame_per_video = 25
num_id = 1
num_worker = 1
num_scores = 1
gpus = [1]
print "dataset: ", dataset, "split: ", split, "modality: ", modality, "gpu: ", num_id
# save_scores = "hmdb51_split_1_flow_tpp_delete_dropout_lr_0.0001_iter_672" ### 66.34%
# save_scores = "hmdb51_split_1_rgb_tpp_delete_dropout_lr_0.0001_iter_112" ### 61.50%
# save_scores = "hmdb51_split_2_rgb_tpp_delete_dropout_lr_0.00001_iter_112" ### 61.18%
# save_scores = "hmdb51_split_2_flow_tpp_delete_dropout_lr_0.0001_iter_112" ### 69.15%
# save_scores = "hmdb51_split_3_rgb_tpp_delete_dropout_lr_0.001_iter_2016" ### 60.52%
# save_scores = "hmdb51_split_3_flow_tpp_delete_dropout_lr_0.0001_iter_224" ### 68.82%
# save_scores = "ucf101_split_1_rgb_tpp_delete_dropout_lr_0.00001_iter_600" ### 89.69%
# save_scores = "ucf101_split_1_flow_tpp_delete_dropout_lr_0.00001_iter_1500" ### 0.8911
# save_scores = "ucf101_split_2_flow_tpp_delete_dropout_sec_lr_0.00001_iter_1200" ### 91.56%
# save_scores = "ucf101_split_2_rgb_tpp_delete_dropout_lr_0.0001_iter_1800" ### 90.30%
# save_scores = "ucf101_split_3_rgb_tpp_delete_dropout_lr_0.0001_iter_900" ### 89.10%
# save_scores = "ucf101_split_3_flow_tpp_delete_dropout_lr_0.00001_iter_1800" ### 92.54%
### ablatation studies
# save_scores = "ucf101_split_1_rgb_tpp_p1_lr_0.00001_iter_600" ### 88.22%
# save_scores = "ucf101_split_1_flow_tpp_p1_lr_0.0001_iter_600" ### 88.16%
# save_scores = 'hmdb51_split_3_rgb_tpp_p1248_lr_0.001_iter_1568' ### 60.59%
# save_scores = 'hmdb51_split_3_rgb_tpp_p12_lr_0.001_iter_224' ### 60.52%
# save_scores = 'hmdb51_split_3_rgb_tpp_p1_lr_0.001_iter_224' ### 58.69%
# save_scores = 'ucf101_split_1_rgb_seg3_lr_0.00001_iter_300' ### 89.29%
# save_scores = 'ucf101_split_1_flow_tpp_p12_lr_0.00001_iter_1200' ### 88.94%
# save_scores = 'ucf101_split_1_rgb_tpp_p1248_lr_0.001_iter_600' ### 89.70%
# save_scores = "ucf101_split_1_rgb_tpp_p12_lr_0.00001_iter_1200" ### 89.6%
# save_scores = 'ucf101_split_1_flow_seg3_lr_0.0001_iter_600' ### 88.61%
# save_scores = 'ucf101_split_1_flow_tpp_p124_ave_lr_0.00001_iter_600' ### 88.39%
# save_scores = 'ucf101_split_1_flow_tpp_p1248_lr_0.00001_iter_1200'
### kinetics
### UCF101
# save_scores = 'ucf101_split_1_rgb_tpp_kinetics_lr_0.0001_iter_1200' ### 93.02%
# save_scores = 'ucf101_split_1_flow_tpp_kinetics_lr_0.0001_iter_600' ## 94.50%
# save_scores = 'ucf101_split_2_rgb_tpp_kinetics_lr_0.00001_iter_300' ### 92.5%
# save_scores = 'ucf101_split_2_flow_tpp_kinetics_lr_0.0001_iter_600' ### 96.7%
# save_scores = 'ucf101_split_3_rgb_tpp_kinetics_lr_0.0001_iter_300' ### 92.92%
# save_scores = 'ucf101_split_3_flow_tpp_kinetics_lr_0.0001_iter_1800' ### 96.85%
save_scores = 'ucf101_split_3_rgb_tpp_kinetics_st_0.01_lr_0.0001_iter_1500' ### 93.39%
######
# save_scores = 'ucf101_split_1_rgb_tpp_kinetics_lr_0.0001_iter_1200' ### 93.02%
# save_scores = 'ucf101_split_1_flow_tpp_kinetics_st_0.001_lr_0.0001_iter_900' ### 95.6%
# save_scores = 'ucf101_split_2_rgb_tpp_kinetics_st_0.001_lr_0.00001_iter_600' ### 93.81%
# save_scores = 'ucf101_split_2_flow_tpp_kinetics_st_0.001_lr_0.0001_iter_600' ### 97.39%
# save_scores = 'ucf101_split_3_rgb_tpp_kinetics_lr_0.0001_iter_300' ### 92.92%
save_scores = 'ucf101_split_3_rgb_tpp_kinetics_st_0.01_lr_0.0001_iter_1500' ### 93.39%
# save_scores = 'ucf101_split_3_flow_tpp_kinetics_st_0.001_lr_0.00001_iter_300' ### 97.08%
#### HMDB51
# save_scores = 'hmdb51_split_1_rgb_tpp_kinetics_lr_0.00001_iter_336' ### 71.90%
# save_scores = 'hmdb51_split_1_flow_tpp_kinetics_lr_0.0001_iter_112' ### 80.20%
# save_scores = 'hmdb51_split_2_rgb_tpp_kinetics_snapshot_lr_0.00001_iter_224' ### 68.63%
# save_scores = 'hmdb51_split_2_flow_tpp_kinetics_lr_0.00001_iter_224' ### 79.02%
# save_scores = 'hmdb51_split_3_rgb_tpp_kinetics_lr_0.0001_iter_672' ### 70.85%
# save_scores = 'hmdb51_split_3_flow_tpp_kinetics_lr_0.00001_iter_224' ### 79.02%
# save_scores = 'hmdb51_split_3_flow_tpp_kinetics_st_0.001_lr_0.0001_iter_336' ### 78.69%
print save_scores
from pyActionRecog.action_caffe import CaffeNet
# build neccessary information
print dataset
split_tp = parse_split_file(dataset)
f_info = parse_directory(frame_path,
rgb_prefix, flow_x_prefix, flow_y_prefix)
gpu_list = gpus
eval_video_list = split_tp[split - 1][1]
score_name = 'pool_fusion'
global ii
ii = 0
def build_net():
global net
my_id = multiprocessing.current_process()._identity[0] \
if num_worker > 1 else 1
print "my", my_id
if num_worker == 1:
net = CaffeNet(net_proto, net_weights, num_id)
else:
if gpu_list is None:
net = CaffeNet(net_proto, net_weights, my_id-1)
else:
net = CaffeNet(net_proto, net_weights, gpu_list[my_id - 1])
def eval_video(video):
global net
label = video[1]
vid = video[0]
video_frame_path = f_info[0][vid]
if modality == 'rgb':
cnt_indexer = 1
elif modality == 'flow':
cnt_indexer = 2
else:
raise ValueError(modality)
frame_cnt = f_info[cnt_indexer][vid]
stack_depth = 0
if modality == 'rgb':
stack_depth = 1
elif modality == 'flow':
stack_depth = 5
step = 1.0 * (frame_cnt - stack_depth) / (num_frame_per_video-1)
if step > 0:
frame_ticks = []
for i in range(num_frame_per_video):
frame_ticks.append(int(1 + i * step))
else:
frame_ticks = [1] * num_frame_per_video
assert(len(frame_ticks) == num_frame_per_video)
frame_scores = []
if modality == 'rgb':
rgb_stack = []
for iii in range(num_frame_per_video):
tick = frame_ticks[iii]
name = '{}{:05d}.jpg'.format(rgb_prefix, tick)
frame = cv2.imread(os.path.join(video_frame_path, name), cv2.IMREAD_COLOR)
rgb_stack.append(frame)
if memory_enough is False:
scores = net.predict_single_rgb_stack_memory(rgb_stack, score_name, frame_size=(340, 256), stack_len=num_frame_per_video)
else:
scores = net.predict_single_rgb_stack(rgb_stack, score_name, frame_size=(340, 256),
stack_len=num_frame_per_video)
frame_scores.append(scores)
if modality == 'flow':
flow_stack = []
for iii in range(num_frame_per_video):
tick = frame_ticks[iii]
frame_idx = [min(frame_cnt, tick+offset) for offset in xrange(stack_depth)]
flow_one = np.empty((256, 340, 10), dtype=np.float32)
ix = 0
for idx in frame_idx:
x_name = '{}{:05d}.jpg'.format(flow_x_prefix, idx)
y_name = '{}{:05d}.jpg'.format(flow_y_prefix, idx)
frame_1 = cv2.imread(os.path.join(video_frame_path, x_name), cv2.IMREAD_GRAYSCALE)
frame_2 = cv2.imread(os.path.join(video_frame_path, y_name), cv2.IMREAD_GRAYSCALE)
flow_one[:,:,ix] = frame_1
ix += 1
flow_one[:,:,ix] = frame_2
ix += 1
flow_stack.append(flow_one)
if memory_enough is False:
scores = net.predict_single_flow_stack_test_memory(flow_stack, score_name, frame_size=(340, 256), stack_len=num_frame_per_video)
else:
scores = net.predict_single_flow_stack_test(flow_stack, score_name, frame_size=(340, 256))
frame_scores.append(scores)
global ii
ii += 1
print ii, 'video {} done'.format(vid)
sys.stdin.flush()
return np.array(frame_scores), label
if num_worker > 1:
pool = multiprocessing.Pool(num_worker, initializer=build_net)
video_scores = pool.map(eval_video, eval_video_list)
else:
build_net()
video_scores = map(eval_video, eval_video_list)
video_pred = [np.argmax(default_aggregation_func(x[0])) for x in video_scores]
video_labels = [x[1] for x in video_scores]
cf = confusion_matrix(video_labels, video_pred).astype(float)
cls_cnt = cf.sum(axis=1)
cls_hit = np.diag(cf)
cls_acc = cls_hit/cls_cnt
print cls_acc
print 'Accuracy {:.02f}%'.format(np.mean(cls_acc)*100)
if save_scores is not None:
np.savez(save_scores, scores=video_scores, labels=video_labels)