-
Notifications
You must be signed in to change notification settings - Fork 1
/
fuse_8.py
239 lines (210 loc) · 9.33 KB
/
fuse_8.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
import torch
import torch.nn as nn
from pytorch_i3d import InceptionI3d
from pytorch_i3d_attn import InceptionI3d_skl
from torchsummary import summary
from dataset import *
from main import *
import numpy as np
class Cont_Loss(torch.nn.Module):
def __init__(self):
super(Cont_Loss, self).__init__()
def forward(self, x, y, n_pos, n_neg):
#y = y[:x.shape[0]]
bs = x.shape[0]
dist = torch.empty(bs)
for i in range(bs):
dist[i] = torch.exp(torch.matmul(
x[i].reshape(1, -1), y[i].reshape(1, -1).t()))
pos_target = torch.ones([n_pos])
neg_target = torch.zeros([n_neg])
target = torch.cat([pos_target, neg_target])
return torch.nn.functional.binary_cross_entropy_with_logits(dist, target).cuda()
class Fusion_Loss(torch.nn.Module):
def __init__(self):
super(Fusion_Loss, self).__init__()
def forward(self, x, y):
y = y[:x.shape[0]]
return torch.sum(torch.pow(torch.abs(torch.sub(x, y)), 2))
def init_i3d():
# INITIALIZE I3D WITH PRE-TRAINED WEIGHTS
i3d = InceptionI3d(400, in_channels=3)
i3d.replace_logits(31)
# print(i3d)
i3d.load_state_dict(torch.load('weights/weights_i3d.pt'))
#summary(i3d, (3, 64, 224, 224))
i3d.fuse_bin = True
i3d.fuse_layer()
i3d.cuda()
#summary(i3d, (3, 64, 224, 224))
i3d = nn.DataParallel(i3d)
#summary(i3d, (3, 64, 224, 224))
return i3d
def init_i3d_skel():
# INITIALIZE I3D WITH PRE-TRAINED WEIGHTS
i3d = InceptionI3d_skl(400, in_channels=3)
i3d.replace_logits(31)
# print(i3d)
i3d.load_state_dict(torch.load('weights/weights_i3d.pt'))
i3d.fuse_layer()
i3d.cuda()
i3d = nn.DataParallel(i3d)
#summary(i3d, (3, 64, 224, 224))
return i3d
def init_agcn(flag):
# INITIALIZE AGCN WITH PRE-TRAINED WEIGHTS
if flag == 1:
parser = get_parser1()
print('attention')
else:
parser = get_parser2()
print('contrastive')
# load arg form config file
p = parser.parse_args()
if p.config is not None:
with open(p.config, 'r') as f:
default_arg = yaml.load(f)
key = vars(p).keys()
for k in default_arg.keys():
if k not in key:
print('WRONG ARG: {}'.format(k))
assert (k in key)
parser.set_defaults(**default_arg)
arg = parser.parse_args()
init_seed(12)
proc = Processor(arg)
agcn = proc.load_model()
#summary(agcn, (3, 400, 15, 1))
return proc, agcn
if __name__ == '__main__':
i3d = init_i3d()
i3d_skel = init_i3d_skel()
proc, agcn = init_agcn(1)
proc_cont, agcn_cont = init_agcn(2)
batch_size = 16
n_pos = 8
n_neg = batch_size - n_pos
weight_factor = 0.01
split = 'new'
root = '/data/stars/user/rdai/smarthomes/Blurred_smarthome_clipped_SSD/'
dataset = Dataset('/data/stars/user/sdas/smarthomes_data/splits/train_'+split+'_CS.txt', 'train', root, 'rgb', None,
'/data/stars/user/sdas/PhD_work/poses_attention/2s-AGCN-For-Daily-Living/data/xsub/train_data_joint_'+split+'.npy',
'/data/stars/user/sdas/PhD_work/poses_attention/2s-AGCN-For-Daily-Living/data/xsub/train_label_'+split+'.pkl',
random_choose=4000, random_shift=False, random_move=False,
window_size=400, normalization=False, debug=False, use_mmap=True)
dataloader = torch.utils.data.DataLoader(
dataset, batch_size=n_pos, shuffle=True, num_workers=36, drop_last=True, pin_memory=True)
val_dataset = Dataset('/data/stars/user/sdas/smarthomes_data/splits/validation_'+split+'_CS.txt', 'val', root, 'rgb', None,
'/data/stars/user/sdas/PhD_work/poses_attention/2s-AGCN-For-Daily-Living/data/xsub/val_data_joint_'+split+'.npy',
'/data/stars/user/sdas/PhD_work/poses_attention/2s-AGCN-For-Daily-Living/data/xsub/val_label_'+split+'.pkl',
random_choose=4000, random_shift=False, random_move=False,
window_size=400, normalization=False, debug=False, use_mmap=True)
val_dataloader = torch.utils.data.DataLoader(
val_dataset, batch_size=n_pos, shuffle=False, num_workers=36, drop_last=True, pin_memory=True)
dataloaders = {'train': dataloader, 'val': val_dataloader}
datasets = {'train': dataset, 'val': val_dataset}
ske_dataloader = proc_cont.load_data(n_neg)
save_model = 'weights_fused_new/'
max_steps = 100
lr = 0.01
optimizer = optim.SGD(i3d.parameters(), lr=lr, momentum=0.9)
#lr_sched = optim.lr_scheduler.MultiStepLR(optimizer, [20, 50])
lr_sched = optim.lr_scheduler.ReduceLROnPlateau(
optimizer, factor=0.1, patience=10, verbose=True)
num_steps_per_update = 1 # accum gradient
steps = 0
# train it
while steps < max_steps: # for epoch in range(num_epochs):
print ('Step {}/{}'.format(steps, max_steps))
print ('-' * 10)
# Each epoch has a training and validation phase
for phase in ['train', 'val']:
if phase == 'train':
i3d.train(True)
i3d_skel.train(True)
agcn.train(True)
else:
i3d.train(False) # Set model to evaluate mode
i3d_skel.train(False)
agcn.train(False)
# print(phase)
tot_loss = 0.0
tot_acc = 0.0
num_iter = 0
optimizer.zero_grad()
#process = tqdm(ske_dataloader[phase])
# Iterate over data.
for neg_ske_data, data in tqdm(zip(ske_dataloader[phase], dataloaders[phase])):
num_iter += 1
# get the inputs
#inputs, labels = i3d_data
inputs, labels, ske_input, ske_label = data
inputs = torch.cat([inputs, inputs])
labels = torch.cat([labels, labels])
ske_label = torch.cat([ske_label, ske_label])
ske_input_att = torch.cat([ske_input, ske_input])
ske_input_cont = torch.cat([ske_input, neg_ske_data])
#print(torch.max(labels, dim=1)[1].long(), ske_label)
# wrap them in Variable
inputs = Variable(inputs.cuda())
labels = Variable(labels.cuda())
attention_weights, per_frame_logits, fuse_logits = i3d(inputs)
ske_input_att = Variable(ske_input_att.float().cuda())
ske_label = Variable(ske_label.long().cuda())
ske_output, ske_emb = agcn(ske_input_att)
per_frame_logits_skel = i3d_skel(inputs, ske_output)
#print(per_frame_logits.shape, fuse_logits.shape)
# Skeleton model output
#ske_input, ske_label, index = ske_data
with torch.no_grad():
ske_input_cont = Variable(
ske_input_cont.float().cuda(),
requires_grad=False)
ske_label = Variable(
ske_label.long().cuda(),
requires_grad=False)
ske_output = agcn_cont(ske_input_cont)
#print(fuse_logits.shape, ske_output.shape)
# compute classification loss (with max-pooling along time B x C x T)
criterion = nn.CrossEntropyLoss().cuda()
loss1 = criterion(per_frame_logits, torch.max(
labels, dim=1)[1].long())
loss2 = criterion(per_frame_logits_skel,
torch.max(labels, dim=1)[1].long())
tot_loss += loss1.data
tot_loss += 0.1*loss2.data
# INCLUDE FUSION LOSS
fusion_loss = Fusion_Loss().cuda()
fuse_loss = fusion_loss(attention_weights, ske_emb)
cont_loss = Cont_Loss().cuda()
cont_loss = cont_loss(fuse_logits, ske_output, n_pos, n_neg)
#fuse_loss = torch.sum(torch.square(torch.abs(torch.sub(per_frame_logits, ske_output))))
# print(fuse_loss)
tot_loss += fuse_loss.data * 0.001
t_loss = (0.9*loss1) + (0.1*loss2) + \
(fuse_loss*0.001) + (0.001*cont_loss)
t_loss.backward()
acc = calculate_accuracy(
per_frame_logits, torch.max(labels, dim=1)[1])
'''
if phase == 'val':
print(torch.max(per_frame_logits, dim=1)[1].long(), torch.max(labels, dim=1)[1].long())
'''
# print(acc)
tot_acc += acc
if phase == 'train':
optimizer.step()
optimizer.zero_grad()
# lr_sched.step()
if phase == 'train':
print ('{} Tot Loss: {:.4f}, Acc: {:.4f}'.format(
phase, tot_loss/num_iter, tot_acc/num_iter))
# save model
torch.save(i3d.module.state_dict(),
save_model+str(steps).zfill(6)+'.pt')
tot_loss = tot_acc = 0.
steps += 1
if phase == 'val':
lr_sched.step(tot_loss/num_iter)
print ('{} Tot Loss: {:.4f}, Acc: {:.4f}'.format(
phase, (tot_loss)/num_iter, tot_acc/num_iter))