/
fan_ckplus_traintest.py
169 lines (149 loc) · 7.78 KB
/
fan_ckplus_traintest.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
import os
import argparse
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.backends.cudnn as cudnn
from basic_code import load, util, networks
DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
def main():
parser = argparse.ArgumentParser(description='PyTorch Frame Attention Network Training')
parser.add_argument('--at_type', '--attention', default=1, type=int, metavar='N',
help= '0 is self-attention; 1 is self + relation-attention')
parser.add_argument('--epochs', default=60, type=int, metavar='N',
help='number of total epochs to run')
parser.add_argument('-f', '--fold', default=10, type=int, help='which fold used for ck+ test')
parser.add_argument('--lr', '--learning-rate', default=1e-2, type=float,
metavar='LR', help='initial learning rate')
parser.add_argument('-e', '--evaluate', default=False, dest='evaluate', action='store_true',
help='evaluate model on validation set')
args = parser.parse_args()
best_acc = 0
at_type = ['self-attention', 'self_relation-attention'][args.at_type]
logger = util.Logger('./log/','fan_ckplus')
logger.print('The attention method is {:}, learning rate: {:}'.format(at_type, args.lr))
''' Load data '''
video_root = './data/face/ck_face'
video_list = './data/txt/CK+_10-fold_sample_IDascendorder_step10.txt'
batchsize_train= 48
batchsize_eval= 64
train_loader, val_loader = load.ckplus_faces_fan(video_root, video_list, args.fold, batchsize_train, batchsize_eval)
''' Load model '''
_structure = networks.resnet18_at(at_type=at_type)
_parameterDir = './pretrain_model/Resnet18_FER+_pytorch.pth.tar'
model = load.model_parameters(_structure, _parameterDir)
''' Loss & Optimizer '''
optimizer = torch.optim.SGD(filter(lambda p: p.requires_grad, model.parameters()), args.lr, momentum=0.9, weight_decay=1e-4)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=30, gamma=0.2)
cudnn.benchmark = True
''' Train & Eval '''
if args.evaluate == True:
logger.print('args.evaluate: {:}', args.evaluate)
val(val_loader, model, logger)
return
logger.print('frame attention network (fan) ck+ dataset, learning rate: {:}'.format(args.lr))
for epoch in range(args.epochs):
train(train_loader, model, optimizer, epoch)
acc_epoch = val(val_loader, model, at_type)
is_best = acc_epoch > best_acc
if is_best:
logger.print('better model!')
best_acc = max(acc_epoch, best_acc)
util.save_checkpoint({
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'accuracy': acc_epoch,
}, at_type=at_type)
lr_scheduler.step()
logger.print("epoch: {:} learning rate:{:}".format(epoch+1, optimizer.param_groups[0]['lr']))
def train(train_loader, model, optimizer, epoch):
losses = util.AverageMeter()
topframe = util.AverageMeter()
topVideo = util.AverageMeter()
# switch to train mode
output_store_fc = []
target_store = []
index_vector = []
model.train()
for i, (input_first, input_second, input_third, target_first, index) in enumerate(train_loader):
target_var = target_first.to(DEVICE)
input_var = torch.stack([input_first, input_second , input_third], dim=4).to(DEVICE)
# compute output
''' model & full_model'''
pred_score = model(input_var)
loss = F.cross_entropy(pred_score, target_var)
loss = loss.sum()
#
output_store_fc.append(pred_score)
target_store.append(target_var)
index_vector.append(index)
# measure accuracy and record loss
acc_iter = util.accuracy(pred_score.data, target_var, topk=(1,))
losses.update(loss.item(), input_var.size(0))
topframe.update(acc_iter[0], input_var.size(0))
# compute gradient and do SGD step
optimizer.zero_grad()
loss.backward()
optimizer.step()
if i % 200 == 0:
logger.print('Epoch: [{:3d}][{:3d}/{:3d}]\t'
'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Acc@1 {topframe.val:.3f} ({topframe.avg:.3f})\t'
.format(
epoch, i, len(train_loader), loss=losses, topframe=topframe))
index_vector = torch.cat(index_vector, dim=0) # [256] ... [256] ---> [21570]
index_matrix = []
for i in range(int(max(index_vector)) + 1):
index_matrix.append(index_vector == i)
index_matrix = torch.stack(index_matrix, dim=0).to(DEVICE).float() # [21570] ---> [380, 21570]
output_store_fc = torch.cat(output_store_fc, dim=0) # [256,7] ... [256,7] ---> [21570, 7]
target_store = torch.cat(target_store, dim=0).float() # [256] ... [256] ---> [21570]
pred_matrix_fc = index_matrix.mm(output_store_fc) # [380,21570] * [21570, 7] = [380,7]
target_vector = index_matrix.mm(target_store.unsqueeze(1)).squeeze(1).div(
index_matrix.sum(1)).long() # [380,21570] * [21570,1] -> [380,1] / sum([21570,1]) -> [380]
acc_video = util.accuracy(pred_matrix_fc.cpu(), target_vector.cpu(), topk=(1,))
topVideo.update(acc_video[0], i + 1)
logger.print(' *Acc@Video {topVideo.avg:.3f} *Acc@Frame {topframe.avg:.3f} '.format(topVideo=topVideo, topframe=topframe))
def val(val_loader, model, at_type):
topVideo = util.AverageMeter()
# switch to evaluate mode
model.eval()
output_store_fc = []
output_alpha = []
target_store = []
index_vector = []
with torch.no_grad():
for i, (input_var, target, index) in enumerate(val_loader):
# compute output
target = target.to(DEVICE)
input_var = input_var.to(DEVICE)
''' model & full_model'''
f, alphas = model(input_var, phrase = 'eval')
output_store_fc.append(f)
output_alpha.append(alphas)
target_store.append(target)
index_vector.append(index)
index_vector = torch.cat(index_vector, dim=0) # [256] ... [256] ---> [21570]
index_matrix = []
for i in range(int(max(index_vector)) + 1):
index_matrix.append(index_vector == i)
index_matrix = torch.stack(index_matrix, dim=0).to(DEVICE).float() # [21570] ---> [380, 21570]
output_store_fc = torch.cat(output_store_fc, dim=0) # [256,7] ... [256,7] ---> [21570, 7]
output_alpha = torch.cat(output_alpha, dim=0) # [256,1] ... [256,1] ---> [21570, 1]
target_store = torch.cat(target_store, dim=0).float() # [256] ... [256] ---> [21570]
''' keywords: mean_fc ; weight_sourcefc; sum_alpha; weightmean_sourcefc '''
weight_sourcefc = output_store_fc.mul(output_alpha) #[21570,512] * [21570,1] --->[21570,512]
sum_alpha = index_matrix.mm(output_alpha) # [380,21570] * [21570,1] -> [380,1]
weightmean_sourcefc = index_matrix.mm(weight_sourcefc).div(sum_alpha)
target_vector = index_matrix.mm(target_store.unsqueeze(1)).squeeze(1).div(
index_matrix.sum(1)).long() # [380,21570] * [21570,1] -> [380,1] / sum([21570,1]) -> [380]
if at_type == 'self-attention':
pred_score = model(vm=weightmean_sourcefc, phrase='eval', AT_level='pred')
if at_type == 'self_relation-attention':
pred_score = model(vectors=output_store_fc, vm=weightmean_sourcefc, alphas_from1=output_alpha, index_matrix=index_matrix, phrase='eval', AT_level='second_level')
acc_video = util.accuracy(pred_score.cpu(), target_vector.cpu(), topk=(1,))
topVideo.update(acc_video[0], i + 1)
logger.print(' *Acc@Video {topVideo.avg:.3f} '.format(topVideo=topVideo))
return topVideo.avg
if __name__ == '__main__':
main()