-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
186 lines (155 loc) · 8.49 KB
/
train.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 os
import random
import argparse
import logging
import torch
import cv2
import torch.nn.functional as F
from torchvision import transforms
from models.emonet_split import EmoNet
from datasets.data import dataloader
from datasets.sampler import ImbalancedDatasetSampler
from utils.utils_logging import AverageMeter, init_logging
from utils.utils_callbacks import CallBackEvaluation, CallBackLogging
from losses.losses import pearsonr, concordance_cc
from utils.utlis_pts import *
torch.backends.cudnn.benchmark = True
parser = argparse.ArgumentParser(description='parameters for face emotion recognition network')
parser.add_argument('--batchsize', type=int, default=64, metavar='N', help='batch size')
parser.add_argument('--epochs', type=int, default=50, metavar='N', help='training epochs')
parser.add_argument('--nclasses', type=int, default=5, choices=[5,8])
parser.add_argument('--num_workers', type=int, default=4, metavar='N', help='5,8')
parser.add_argument('--wd', type=float, default=0)
parser.add_argument('--lr', type=float, default=0.001)
parser.add_argument('--s', type=float, default=260.0)
parser.add_argument('--pts', action='store_true', help='plot alignment on face images')
parser.add_argument('--kd', action='store_true', help='train with kd')
parser.add_argument('--kd_w', type=float,choices=[0.3,0.6])
parser.add_argument('--path', type=str, default='', help='teacher mode path')
class ImageFlip(object):
def __call__(self, img):
img = transforms.functional.hflip(img)
return img
def kl_loss(x_s, y_t, T = 4):
p = F.log_softmax(x_s / T, dim=1)
q = F.softmax(y_t / T, dim=1)
l_kl = F.kl_div(p, q, reduction='sum') * (T ** 2) / x_s.shape[0]
return l_kl
def main():
args = parser.parse_args()
manualSeed = random.randint(1, 10000)
random.seed(manualSeed)
torch.manual_seed(manualSeed)
torch.cuda.manual_seed_all(manualSeed)
cur_path = os.path.abspath(os.curdir)
n_expression = args.nclasses
if args.kd:
args.output = cur_path.replace('emonet_code','Results')+'/emonet/KD/'+'E_'+str(n_expression)+'_optim_'+str(args.optim)+\
'wd'+str(args.wd)+'lr'+str(args.lr)+'_scale_'+str(args.s)+'_kdw_'+str(args.kd_w)
else:
args.output = cur_path.replace('emonet_code','Results')+'/emonet/'+'E_'+str(n_expression)+'_optim_'+str(args.optim)+\
'wd'+str(args.wd)+'lr'+str(args.lr)+'_scale_'+str(args.s)
if not os.path.exists(args.output):
os.makedirs(args.output, exist_ok=True)
log_root = logging.getLogger()
init_logging(log_root, args.output)
transform_train = transforms.Compose([transforms.RandomHorizontalFlip(), transforms.ToTensor()])
transform_valid_noflip = transforms.Compose([transforms.ToTensor()])
transform_valid_flip = transforms.Compose([ImageFlip(), transforms.ToTensor()])
print('loading training set')
train_data = dataloader(subset='train', transform_image=transform_train, scale=args.s, n_expression=n_expression, seed=manualSeed)
train_loader = torch.utils.data.DataLoader(train_data, sampler=ImbalancedDatasetSampler(train_data),batch_size=args.batchsize, num_workers=args.num_workers, pin_memory=True,drop_last=True)
print('loading validation set')
valid_data = dataloader(subset='valid', transform_image=transform_valid_noflip, n_expression=n_expression)
valid_loader = torch.utils.data.DataLoader(valid_data, batch_size=args.batchsize, shuffle=False, num_workers=args.num_workers, pin_memory=True)
valid_data = dataloader(subset='valid', transform_image=transform_valid_flip, n_expression=n_expression)
valid_loader_flip = torch.utils.data.DataLoader(valid_data, batch_size=args.batchsize, shuffle=False, num_workers=args.num_workers, pin_memory=True)
print('loading testing set')
test_data = dataloader(subset='test', transform_image=transform_valid_noflip, n_expression=n_expression)
test_loader = torch.utils.data.DataLoader(test_data, batch_size=args.batchsize, shuffle=False, num_workers=1, pin_memory=True)
test_data = dataloader(subset='test', transform_image=transform_valid_flip, n_expression=n_expression)
test_loader_flip = torch.utils.data.DataLoader(test_data, batch_size=args.batchsize, shuffle=False, num_workers=1, pin_memory=True)
# model initilization
if args.kd:
model_T = EmoNet(n_expression=n_expression)
state_dict = torch.load(args.path, map_location='cpu')
model_T.predictor.load_state_dict(state_dict)
model_T = torch.nn.DataParallel(model_T).cuda()
model_T.eval()
model = EmoNet(n_expression=n_expression)
model = torch.nn.DataParallel(model).cuda()
params = list(model.parameters())
sub_params = [p for p in params if p.requires_grad]
print('num of params', sum(p.numel() for p in sub_params))
optimizer = torch.optim.AdamW(sub_params, lr=args.lr, betas=(0.9, 0.999), weight_decay=args.wd)
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[15, 30, 45], gamma=0.1, last_epoch=-1)
total_step = int(len(train_data) / args.batchsize * args.epochs)
callback_logging = CallBackLogging(400,total_step, args.batchsize, None)
callback_validation_valid = CallBackEvaluation(valid_loader,valid_loader_flip, subset='valid')
callback_validation_test = CallBackEvaluation(test_loader,test_loader_flip, subset='test')
if args.kd:
_ = callback_validation_test(-1, model_T)
model_T.eval()
# training
global_step = 0
losses = AverageMeter()
acces = AverageMeter()
losses_kd = AverageMeter()
for epoch in range(args.epochs):
model.train()
model.module.feature.eval()
for index, data in enumerate(train_loader):
global_step += 1
images = data['image'].cuda()
valence = data['valence'].squeeze().cuda()
arousal = data['arousal'].squeeze().cuda()
label = data['expression'].cuda()
outputs = model(images)
if args.pts and index<10 and epoch==0:
pts = get_pts(outputs['heatmap'][0])
preds = pts * 4
img_pts = preds[17:, :]
img_show_ = np.asarray(images[0].permute(1,2,0).cpu().data)*255
img_show = LineDrawer_51(img_show_, img_pts.astype('int32'))
impath = args.output+'/debug/'
if not os.path.exists(impath):
os.makedirs(impath)
cv2.imwrite('%s/%d.jpg' % (impath, index), img_show[:,:,-1::])
loss_c = F.cross_entropy(outputs['expression'], label)
loss_mse = F.mse_loss(outputs['valence'], valence.detach())+F.mse_loss(outputs['arousal'], arousal.detach())
loss_pcc = 1-(pearsonr(outputs['valence'], valence.detach())+pearsonr(outputs['arousal'], arousal.detach()))/2
loss_ccc = 1-(concordance_cc(outputs['valence'].squeeze(), valence.detach())+concordance_cc(outputs['arousal'].squeeze(), arousal.detach()))/2
if args.kd:
with torch.no_grad():
outputs_T = model_T(images)
loss_kd = kl_loss(outputs['expression'], outputs_T['expression'].detach(), T=1)
# shake shake
alpha = torch.rand(1).cuda()
alpha.requires_grad = False
beta = torch.rand(1).cuda()
beta.requires_grad = False
gamma = torch.rand(1).cuda()
gamma.requires_grad = False
loss = loss_c + alpha/(alpha+beta+gamma)*loss_mse + beta/(alpha+beta+gamma)*loss_pcc + gamma/(alpha+beta+gamma)*loss_ccc
if args.kd:
loss = loss+loss_kd*args.kd_w
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses.update(loss_c.detach().item(), 1)
if args.kd:
losses_kd.update(loss_kd.detach().item(), 1)
else:
losses_kd.update(loss_pcc.detach().item(), 1)
batch_size = images.size(0)
_, pred = torch.max(outputs['expression'].detach(), 1)
correct = torch.eq(pred, label)
acc = correct.float().sum()/float(batch_size)
acces.update(acc.detach().item(), batch_size)
callback_logging(global_step, losses, losses_kd, acces, epoch, optimizer)
val_results = callback_validation_valid(epoch, model)
test_results = callback_validation_test(epoch, model)
torch.save(model.module.predictor.state_dict(), os.path.join(args.output, "predictor_%d.pth"%epoch))
scheduler.step()
if __name__ == '__main__':
main()