-
Notifications
You must be signed in to change notification settings - Fork 6
/
model_utils.py
384 lines (291 loc) · 13.9 KB
/
model_utils.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
import argparse
import time
import os, sys
import torch
import torch.nn as nn
import torch.nn.init as init
import numpy as np
# from BNInception import bninception
from RGB_OFF import bninception_off
import re, deprecated
import pdb
class RBF(nn.Module):
def __init__(self, num_kernel, input_dim):
super(RBF, self).__init__()
self.num_kernel = num_kernel
self.input_dim = input_dim
self.center = nn.Parameter(torch.rand(self.num_kernel, self.input_dim) , requires_grad=True)
self.beta = nn.Parameter(torch.rand(self.num_kernel), requires_grad=True)
def forward(self, input):
x= (input-self.center).pow(2).sum(2, keepdim=False).sqrt()
x = torch.exp(-self.beta.mul(x))
return x
class topk_crossEntrophy(nn.Module):
def __init__(self, top_k=0.7):
super(topk_crossEntrophy, self).__init__()
self.loss = nn.NLLLoss()
self.top_k = top_k
self.softmax = nn.LogSoftmax()
def forward(self, input, target):
softmax_result = self.softmax(input)
loss = torch.autograd.Variable(torch.Tensor(1).zero_()).cuda()
for idx, row in enumerate(softmax_result):
gt = target[idx]
pred = torch.unsqueeze(row, 0)
gt = torch.unsqueeze(gt, 0)
cost = self.loss(pred, gt)
cost = torch.unsqueeze(cost, 0)
loss = torch.cat((loss, cost), 0)
loss = loss[1:]
if self.top_k == 1.0:
valid_loss = loss
# import pdb;pdb.set_trace()
index = torch.topk(loss, int(self.top_k * loss.size()[0]))
valid_loss = loss[index[1]]
return torch.mean(valid_loss)
def initNetWeights(net, require_grad=True):
for m in net.modules():
if m.requires_grad:
if isinstance(m, nn.Conv2d):
init.xavier_uniform(m.weight)
if m.bias:
init.constant(m.weight)
elif isinstance(m, nn.BatchNorm2d):
init.constant(m.weight, 1)
init.constant(m.bias, 0)
elif isinstance(m, nn.Linear):
init.normal(m.weight, std=1e-3)
if m.bias:
init.constant(m.bias, 0)
def pretrained_bninception():
sys.path.append('/home/zhufl/Temporal-Residual-Motion-Generation/tsn-pytorch')
from BNInception_model import bninception
# net = bninception(num_classes=101, pretrained=None)
net = bninception()
checkpoint = torch.load('/home/zhufl/Temporal-Residual-Motion-Generation/tsn-pytorch/ucf101_rgb.pth') # checkpoint only has 416 in length;
print(len(list(checkpoint)))
net.fc = nn.Linear(1024, 101)
count = 0
base_dict = {}
for k, v in checkpoint.items():
count = count + 1
if 415>count>18:
print count, k[18:]
base_dict.setdefault(k[18:], checkpoint[k])
elif count<19:
print count, k[11:]
base_dict.setdefault(k[11:], checkpoint[k])
# else:
# '''check un-recovered weights '''
# print count, k
# base_dict.setdefault(k[18:], checkpoint[k])
base_dict.setdefault('fc.weight', checkpoint['base_model.fc-action.1.weight'])
base_dict.setdefault('fc.bias', checkpoint['base_model.fc-action.1.bias'])
# For kinetics dataset:
# base_dict.setdefault('new_fc.weight', checkpoint['base_model.fc_action.1.weight'])
# base_dict.setdefault('new_fc.bias', checkpoint['base_model.fc_action.1.bias'])
'''
mode_state = net.state_dict()
re_base_dict = {k[11:]:v for k, v in base_dict.items() if k not in mode_state}
print(re_base_dict.keys())
base_dict = {k: v for k, v in base_dict.items() if k in mode_state}
print(len(list(base_dict)))
print(len(list(re_base_dict)))
base_dict.update(re_base_dict)
print(len(list(base_dict)))
'''
# mode_state = net.state_dict()
# re_base_dict = {k:v for k, v in mode_state.items() if k not in base_dict}
# print(re_base_dict.keys())
# mode_state.update(base_dict)
net.load_state_dict(base_dict)
print("Finish loading the caffemodel ucf101_rgb weights for plain BNInception model")
return net
# @deprecation.deprecated(deprecated_in="1.0", removed_in="2.0",
# current_version=__version__,
# details="Use the bar function instead")
# def pretrained_bninception_off():
# net = bninception_off(101, None)
# checkpoint = torch.load('/home/zhufl/videoPrediction/rgb_off_reference_split_1.caffemodel.pth')
# print("Number of parameters recovered from original caffemodel {}".format(len(checkpoint)))
# # for key,value in checkpoint.items():
# # if 'running_var' in key:
# # print key
# # fix inception & fc-action-motion layer name issue
# for key, value in checkpoint.items():
# new_key = key.replace("/", "_")
# new_key = new_key.replace("-", "_")
# # fix error for copying a param of torch.Size([x]) from checkpoint, where the shape is torch.Size([1, ]) in current model.
# # Dimension error for all bn layer
# if '_bn' in key:
# checkpoint[key] = torch.squeeze(checkpoint[key])
# checkpoint[new_key] = checkpoint.pop(key)
# checkpoint['last_linear.weight'] = checkpoint.pop('fc_action.weight')
# checkpoint['last_linear.bias'] = checkpoint.pop('fc_action.bias')
# # checkpoint['fc_action_motion.weight'] = checkpoint.pop('fc-action-motion.weight')
# # checkpoint['fc_action_motion.bias'] = checkpoint.pop('fc-action-motion.bias')
# # checkpoint['fc_action_motion_28.weight'] = checkpoint.pop('fc-action-motion_28.weight')
# # checkpoint['fc_action_motion_28.bias'] = checkpoint.pop('fc-action-motion_28.bias')
# # checkpoint['fc_action_motion_14.weight'] = checkpoint.pop('fc-action-motion_14.weight')
# # checkpoint['fc_action_motion_14.bias'] = checkpoint.pop('fc-action-motion_14.bias')
# model_state = net.state_dict()
# # print("Number of parameters of oringinal model is {}".format(len(model_state)))
# # for key, value in model_state.items():
# # print(key)
# base_dict = {k:v for k, v in checkpoint.items() if k in model_state}
# # for key, value in base_dict.items():
# # print(key)
# missing_dict = {k:v for k, v in model_state.items() if k not in checkpoint}
# for key, value in missing_dict.items():
# print("Missing {}".format(key))
# print("Number of parameters loaded from well trained model {}".format(len(base_dict)))
# model_state.update(base_dict)
# net.load_state_dict(model_state)
# print("Load weights and bias from RGB_OFF_caffemodel")
# return net
def pretrained_bninception_off(batch, num_seg):
'''
Load converted rgb_off_ucf101_caffemodel
bninception_off input: num_class, num_batch/num_crop, num_seg
'''
net = bninception_off(101, batch, num_seg)
checkpoint = torch.load('/home/zhufl/Data2/caffe2pytorch-tsn/converted_rgb_off_ucf101_caffemodel.pth')
print("Number of parameters recovered from original caffemodel {}".format(len(checkpoint)))
model_state = net.state_dict()
'''
Choose to init motion branch or not
If only fine-tune last layer, then need to load motion branch;
If fine-tune whole motion branch, then no need;
'''
# base_dict = {k:v for k, v in checkpoint.items() if k in model_state}
base_dict = {k:v for k, v in checkpoint.items() if 'fc-action' not in k }
# print(base_dict.keys())
# import pdb;pdb.set_trace()
# missing_dict = {k:v for k, v in model_state.items() if k not in base_dict}
# for key, value in missing_dict.items():
# print("Missing motion branch param {}".format(key))
missing_dict = {k:v for k, v in model_state.items() if k not in checkpoint}
loading_dict = {k:v for k, v in model_state.items() if k in checkpoint}
for key, value in missing_dict.items():
print("Missing {}".format(key))
for key, value in loading_dict.items():
print("Loading {}".format(key))
model_state.update(base_dict)
net.load_state_dict(model_state)
print("Finish Load weights and bias from RGB_OFF_caffemodel")
return net
def fine_tune_bninception_off(batch, num_seg):
'''
Loading most recent pre-trained model for fine-tuning;
bninception_off input: num_class, num_batch/num_crop, num_seg
motion_spatial_grad is fixed; Should not be able to train;
'''
net = bninception_off(101, batch, num_seg)
checkpoint = torch.load('/home/zhufl/Data2/caffe2pytorch-tsn/converted_rgb_off_ucf101_caffemodel.pth')
# checkpoint = torch.load('/home/zhufl/Temporal-Residual-Motion-Generation/tsn-pytorch/ucf101_rgb.pth') # checkpoint only has 416 i
# checkpoint = torch.load('/home/zhufl/videoPrediction/train/' + '2019-01-08_22-42-50.pth')
print("Number of parameters recovered from original caffemodel {}".format(len(checkpoint)))
model_state = net.state_dict()
'''
Choose to init motion branch or not
If only fine-tune last layer, then need to load motion branch;
If fine-tune whole motion branch, then no need;
'''
# base_dict = {k:v for k, v in checkpoint.items() if k in model_state}
# base_dict = {k:v for k, v in checkpoint.items() if 'fc' not in k }
base_dict = {k:v for k, v in checkpoint.items() if 'motion' not in k }
# print(base_dict.keys())
# import pdb;pdb.set_trace()
missing_dict = {k:v for k, v in model_state.items() if k not in base_dict}
for key, value in missing_dict.items():
print("Missing motion branch param {}".format(key))
# missing_dict = {k:v for k, v in model_state.items() if k not in checkpoint}
# for key, value in missing_dict.items():
# print("Missing {}".format(key))
model_state.update(base_dict)
net.load_state_dict(model_state)
print("Load weights and bias from RGB_OFF_caffemodel")
return net
def fine_tune_bninception_off_sobel(batch, num_seg):
'''
Loading most recent pre-trained model for fine-tuning;
bninception_off input: num_class, num_batch/num_crop, num_seg
motion_spatial_grad is fixed; Should not be able to train;
'''
from RGB_OFF_v2 import bninception_off
net = bninception_off(101, batch, num_seg)
# init all trainable variable
# initNetWeights(net)
# model_name = '2019-01-13_00-44-51.pth'
model_name = '2019-01-14_12-39-41.pth'
# checkpoint = torch.load('/home/zhufl/Data2/caffe2pytorch-tsn/converted_rgb_off_ucf101_caffemodel.pth')
checkpoint = torch.load('/home/zhufl/videoPrediction/train/' + model_name)
print("Number of parameters recovered from original caffemodel {}".format(len(checkpoint)))
model_state = net.state_dict()
'''
Choose to init motion branch or not
If only fine-tune last layer, then need to load motion branch;
If fine-tune whole motion branch, then no need;
'''
base_dict = {k:v for k, v in checkpoint.items() if k in model_state}
# base_dict = {k:v for k, v in checkpoint.items() if 'motion' not in k }
# print(base_dict.keys())
# import pdb;pdb.set_trace()
missing_dict = {k:v for k, v in model_state.items() if k not in base_dict}
for key, value in missing_dict.items():
print("Missing motion branch param {}".format(key))
# missing_dict = {k:v for k, v in model_state.items() if k not in checkpoint}
# for key, value in missing_dict.items():
# print("Missing {}".format(key))
model_state.update(base_dict)
net.load_state_dict(model_state)
print("Load weights and bias from " + model_name)
return net
def selftrained_bninception_off(batch, num_seg, model='RGB_OFF_2019-02-13_15-40-01.pth'):
'''
Load self-trained OFF weights;
bninception_off input: num_class, num_batch/num_crop, num_seg;
'''
net = bninception_off(101,batch, num_seg)
checkpoint = torch.load('/home/zhufl/Temporal-Residual-Motion-Generation/videoPrediction/train/' + model)
print("Number of parameters recovered from original caffemodel {}".format(len(checkpoint)))
model_state = net.state_dict()
base_dict = {k:v for k, v in checkpoint.items() if k in model_state}
# print(base_dict.keys())
missing_dict = {k:v for k, v in model_state.items() if k not in checkpoint}
for key, value in missing_dict.items():
print("Missing {}".format(key))
model_state.update(base_dict)
net.load_state_dict(model_state)
print("Load weights and bias from '/home/zhufl/videoPrediction/train/" + model)
return net
def selftrained_bninception_off_sobel(batch, num_seg, model='2019-02-10_12-57-17.pth'):
'''
Load self-trained OFF weights;
bninception_off input: num_class, num_batch/num_crop, num_seg;
'''
from RGB_OFF_v2 import bninception_off
net = bninception_off(101, batch, num_seg)
net = torch.nn.DataParallel(net, device_ids=[0]).cuda()
checkpoint = torch.load('/home/zhufl/Temporal-Residual-Motion-Generation/videoPrediction/train/' + model)
# print("Number of parameters recovered from original caffemodel {}".format(len(checkpoint)))
# model_state = net.state_dict()
# base_dict = {k:v for k, v in checkpoint.items() if k in model_state}
# # print(base_dict.keys())
# missing_dict = {k:v for k, v in model_state.items() if k not in checkpoint}
# for key, value in missing_dict.items():
# print("Missing {}".format(key))
# model_state.update(base_dict)
# net.load_state_dict(model_state)
net.load_state_dict(checkpoint)
print("Load weights and bias from /home/zhufl/Temporal-Residual-Motion-Generation/videoPrediction/train/{}".format(model))
return net
def compare_two_model():
# bn = pretrained_bninception()
bn_off = pretrained_bninception_off()
# weight = bn.conv1_7x7_s2.weight
# weight_off = bn_off.conv1_7x7_s2.weight
# print(np.sum(weight.data.numpy() - weight_off.data.numpy())**2)
if __name__ == '__main__':
# net = pretrained_bninception_off()
compare_two_model()