-
Notifications
You must be signed in to change notification settings - Fork 7
/
pretrain_as.py
186 lines (144 loc) · 6.41 KB
/
pretrain_as.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
import torch
import torch.nn.functional as F
from torch.autograd import Variable
import time
import os
import sys
from kmeanspp import *
from utils import AverageMeter, calculate_accuracy
def momentum_update(model, model_ema, m):
for p1, p2 in zip(model.parameters(), model_ema.parameters()):
p2.data.mul_(m).add_(1-m, p1[1].detach().data)
def init_dict(data_loader, model_clone, batch_num):
feature_v_dict = []
feature_a_dict = []
cuda = torch.device("cuda")
for idx, (video, audio) in enumerate(data_loader):
print('init dict batch', idx)
if idx == batch_num:
break
with torch.no_grad():
video = video.to(device=cuda)
audio = audio.to(device=cuda)
feature_v, feature_a = model_clone(video, audio)
feature_v_dict.append(feature_v)
feature_a_dict.append(feature_a)
feature_v_dict = torch.cat(feature_v_dict, dim=0)
feature_a_dict = torch.cat(feature_a_dict, dim=0)
return feature_v_dict, feature_a_dict
def train_epoch(epoch, data_loader, model, model_clone, feature_v_pool, feature_a_pool, nowidx_pool, feature_v_dict, feature_a_dict, nowidx_dict, criterion, optimizer, opt,
epoch_logger, batch_logger):
print('train at epoch {}'.format(epoch))
model.train()
model_clone.train()
cuda = torch.device("cuda")
batch_time = AverageMeter()
data_time = AverageMeter()
losses = AverageMeter()
accuracies = AverageMeter()
end_time = time.time()
for i, (video, audio) in enumerate(data_loader):
data_time.update(time.time() - end_time)
video = video.to(device=cuda)
audio = audio.to(device=cuda)
bs = video.size(0)
device = video.device
feature_v, feature_a = model(video, audio) # M * 512, M * 512
# sampling hard example based on gradient from pool
target = torch.arange(bs).to(device=device)
feature_a_pool_all = torch.cat([feature_a.detach(), feature_a_pool], dim=0) # (M+N) * 512
feature_v_pool_all = torch.cat([feature_v.detach(), feature_v_pool], dim=0) # (M+N) * 512
feature_a_pool_all.requires_grad = True
feature_v_pool_all.requires_grad = True
cosv2a = torch.mm(feature_a_pool_all, feature_v.t()).t() # M * (M+N)
cosa2v = torch.mm(feature_v_pool_all, feature_a.t()).t() # M * (M+N)
lossv2a = criterion(cosv2a, target) # M * (M+N)
lossa2v = criterion(cosa2v, target) # M * (M+N)
lossv2a = lossv2a.mean() # 1
lossa2v = lossa2v.mean() # 1
# have not tested
lossv2a.backward(retain_graph=True)
lossa2v.backward(retain_graph=True)
a_pool_gradient = feature_a_pool_all.grad.data # (M+N) * 512
v_pool_gradient = feature_v_pool_all.grad.data # (M+N) * 512
a_pool_gradient = a_pool_gradient.detach().cpu().numpy()
v_pool_gradient = v_pool_gradient.detach().cpu().numpy()
# a fake function
a_idx = kmeanspp_select(a_pool_gradient, bs) # M
v_idx = kmeanspp_select(v_pool_gradient, bs) # M
a_pool_sample = feature_a_pool_all[a_idx] # M * 512
v_pool_sample = feature_v_pool_all[v_idx] # M * 512
#pool_shape = a_idx.shape[0]
pool_shape = a_pool_sample.shape[0]
if nowidx_dict + pool_shape > feature_v_dict.shape[0]:
nowidx_dict = 0
feature_v_dict[nowidx_dict:nowidx_dict+pool_shape] = v_pool_sample.detach()
feature_a_dict[nowidx_dict:nowidx_dict+pool_shape] = a_pool_sample.detach()
nowidx_dict += pool_shape
#if nowidx_dict + pool_shape > feature_v_dict.shape[0]:
# nowidx_dict = 0
# -----
# compute loss with dict
feature_a_dict_all = torch.cat([feature_a.detach(), feature_a_dict], dim=0) # (M+K) * 512
feature_v_dict_all = torch.cat([feature_v.detach(), feature_v_dict], dim=0) # (M+K) * 512
cosv2a = torch.mm(feature_a_dict_all, feature_v.t()).t() # M * (M+K)
cosa2v = torch.mm(feature_v_dict_all, feature_a.t()).t() # M * (M+K)
lossv2a = criterion(cosv2a, target).mean() # 1
lossa2v = criterion(cosa2v, target).mean() # 1
loss = lossv2a + lossa2v
acc = calculate_accuracy(cosv2a, target)
losses.update(loss.data.item(), video.size(0))
accuracies.update(acc, video.size(0))
optimizer.zero_grad()
loss.backward()
optimizer.step()
#model_clone(v=cosv2a, a=cosa2v, swap_av=True)
momentum_update(model, model_clone, 0.9)
#model_clone(v=cosv2a, a=cosa2v, swap_av=True)
# update pool
with torch.no_grad():
feature_v, feature_a = model_clone(video, audio)
if nowidx_pool + bs > feature_v_pool.shape[0]:
nowidx_pool = 0
feature_v_pool[nowidx_pool:nowidx_pool+bs] = feature_v.detach()
feature_a_pool[nowidx_pool:nowidx_pool+bs] = feature_a.detach()
nowidx_pool += bs
batch_time.update(time.time() - end_time)
end_time = time.time()
batch_logger.log({
'epoch': epoch,
'batch': i + 1,
'iter': (epoch - 1) * len(data_loader) + (i + 1),
'loss': losses.val,
'acc': accuracies.val,
'lr': optimizer.param_groups[0]['lr']
})
print('Epoch: [{0}][{1}/{2}]\t'
'Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
'Data {data_time.val:.3f} ({data_time.avg:.3f})\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc {acc.val:.3f} ({acc.avg:.3f})'.format(
epoch,
i + 1,
len(data_loader),
batch_time=batch_time,
data_time=data_time,
loss=losses,
acc=accuracies))
epoch_logger.log({
'epoch': epoch,
'loss': losses.avg,
'acc': accuracies.avg,
'lr': optimizer.param_groups[0]['lr']
})
if epoch % opt.checkpoint == 0:
save_file_path = os.path.join(opt.result_path,
'save_{}.pth'.format(epoch))
states = {
'epoch': epoch + 1,
'arch': opt.arch,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
}
torch.save(states, save_file_path)
return feature_v_pool, feature_a_pool, nowidx_pool, feature_v_dict, feature_a_dict, nowidx_dict