-
Notifications
You must be signed in to change notification settings - Fork 3
/
DI_train.py
164 lines (138 loc) · 6.84 KB
/
DI_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
# -*- coding: utf-8 -*-
"""
Created on Sun Apr 21 15:23:11 2019
@author: lyfeng
"""
import torch,os
import CNN_R,CNN_H
from CASIAWebFace import CASIAWebFace_dataset
from torch.autograd import Variable
import argparse
import torch.optim as optim
import torch.nn as nn
import datetime
import torchvision.transforms as transforms
import torch.backends.cudnn as cudnn
parser = argparse.ArgumentParser(description='PyTorch sphereface')
parser.add_argument('--net_r', default='sphere36a', type=str)
parser.add_argument('--net_h', default='Hallu_Net', type=str)
parser.add_argument('--weights_r', default='./weights/CNN_R_sphere36a_19.pth', help='weights of the trained net_r')
parser.add_argument('--weights_h', default='./weights/CNN_H_Hallu_Net_19.pth', help='weights of the trained net_h')
parser.add_argument('--alpha', default=8, type=int, help='weights of L_SR & L_SI')
#parser.add_argument('--dataset', default='../../dataset/face/casia/casia.zip', type=str)
parser.add_argument('--lr', default=0.001, type=float, help='learning rate')
#parser.add_argument('--bs_r', default=256, type=int, help='')
parser.add_argument('--bs', default=32, type=int, help='')
parser.add_argument('--data_root',default='../datasets/CASIA-WebFace-aligned')
parser.add_argument('--file_root',default='../datasets/casia_landmark.txt')
args = parser.parse_args()
use_cuda = torch.cuda.is_available()
#cudnn.benchmark = True
transform = transforms.Compose([
transforms.ToTensor(), # range [0, 255] -> [0.0,1.0]
transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)) # range [0.0, 1.0] -> [-1.0,1.0]
])
dataset = CASIAWebFace_dataset(args.data_root, args.file_root, transform=transform, downsample=True)
train_loader = torch.utils.data.DataLoader(dataset, batch_size=args.bs, shuffle=True, num_workers=4, drop_last=False)
class L2Norm(nn.Module):
def __init__(self):
super(L2Norm,self).__init__()
self.eps = 1e-10
def forward(self, x):
norm = torch.sqrt(torch.sum(x * x, dim = 1) + self.eps)
x= x / norm.unsqueeze(-1).expand_as(x)
return x
class SuperIdentityLoss(nn.Module):
def __init__(self):
super(SuperIdentityLoss, self).__init__()
self.eps = 1e-6
def forward(self, X, Y): # X,Y are feats
diff = torch.add(L2Norm()(X), -L2Norm()(Y))
error = diff * diff + self.eps
loss = torch.mean(error) # wrong here!
return loss
# Domain-Integrated Training
def train(net_r,net_h,epoch,args,train_loader):
net_r.train()
net_h.train()
train_loss = 0
correct_I_SR = 0
correct_I_HR = 0
total = 0
batch_idx = 0
for batch_idx, data in enumerate(train_loader):
""" get the training data and label """
I_LR,I_HR,label = data # downsampled 1/8
if I_HR is None: break
I_LR = I_LR.float()
I_HR = I_HR.float()
label = label.long()
if use_cuda: I_LR, I_HR, label = I_LR.cuda(), I_HR.cuda(), label.cuda()
I_LR, I_HR, label = Variable(I_LR), Variable(I_HR), Variable(label) # bs=128
""" train the recognition model"""
I_SR = net_h(I_LR) # I_SR
net_r.feature = False # output prob!
outputs_I_SR = net_r(I_SR)
outputs_I_HR = net_r(I_HR)
loss_r1 = criterion_r(outputs_I_SR, label) # bs=128
loss_r2 = criterion_r(outputs_I_HR, label) # bs=128. totally bs=256 for CNN_R
loss_r = loss_r1+loss_r2
optimizer_r.zero_grad()
loss_r.backward(retain_graph=True)
optimizer_r.step()
""" train the hallucination model"""
loss_h1 = criterion_h(I_SR, I_HR) # I_SR = net_h(I_LR) bs=128 for CNN_H
net_r.feature = True # output feats!
I_SR_feats = net_r(I_SR) # norm !
I_HR_feats = net_r(I_HR) # norm !
loss_h2 = criterion_si(I_SR_feats,I_HR_feats)
loss_h = loss_h1 + args.alpha*loss_h2 # alpha!
optimizer_h.zero_grad()
loss_h.backward()
optimizer_h.step()
""" the recognition model training info """
# train_loss += loss.item()
outputs_I_SR = outputs_I_SR[0] # 0=cos_theta 1=phi_theta
outputs_I_HR = outputs_I_HR[0]
_, pre_I_SR = torch.max(outputs_I_SR.data, 1)
_, pre_I_HR = torch.max(outputs_I_HR.data, 1)
total += label.size(0)
correct_I_SR += pre_I_SR.eq(label.data).cpu().sum()
correct_I_HR += pre_I_HR.eq(label.data).cpu().sum()
if batch_idx%50==0:
# print the cnn_r info
print('Epoch %d (%d/%d) loss_r=%.4f = loss_r1=%.4f + loss_r2=%.4f Acc(I_SR)=%.4f%% (%d/%d) Acc(I_HR)=%.4f%% (%d/%d) lamb=%.2f it=%d'
% (epoch, batch_idx, len(train_loader), loss_r.item(), loss_r1.item(), loss_r2.item(),
100.0*correct_I_SR/total, correct_I_SR, total, 100.0*correct_I_HR/total, correct_I_HR, total,
criterion_r.lamb, criterion_r.it))
#
# print the cnn_h info
print(' loss_h=%.4f = loss_h1=%.4f + loss_h2=%.4f'
% (loss_h.item(), loss_h1.item(), (args.alpha*loss_h2).item()))
def dt():
return datetime.datetime.now().strftime('%H:%M:%S')
def save_model(model,filename):
state = model.state_dict()
for key in state: state[key] = state[key].clone().cpu()
torch.save(state, filename)
if __name__ == '__main__':
net_r = getattr(CNN_R,args.net_r)().cuda()
net_r.load_state_dict(torch.load(args.weights_r))
criterion_r = CNN_R.AngleLoss() # L_FR Recognition loss
optimizer_r = optim.SGD(net_r.parameters(), lr=args.lr, momentum=0.9, weight_decay=5e-4)
net_h = getattr(CNN_H,args.net_h)(64,32,4,0).cuda()
net_h.load_state_dict(torch.load(args.weights_h))
criterion_h = CNN_H.Super_Resolution_Loss() # L_SR super-resolution loss
optimizer_h = optim.SGD(net_h.parameters(), lr=args.lr, momentum=0.9, weight_decay=5e-4)
criterion_si = SuperIdentityLoss() # L_SIsuper-identity loss
print('start: time={}'.format(dt()))
for epoch in range(0, 20):
# print("Epoch---{:d}".format(epoch+1))
if epoch in [0,6,10,15]:
if epoch!=0: args.lr *= 0.1
optimizer_r = optim.SGD(net_r.parameters(), lr=args.lr, momentum=0.9, weight_decay=5e-4)
optimizer_h = optim.SGD(net_h.parameters(), lr=args.lr, momentum=0.9, weight_decay=5e-4)
train(net_r,net_h,epoch,args,train_loader)
save_model(net_r, 'FT_{}_{}.pth'.format(args.net_r,epoch)) # FT:finetune
save_model(net_h, 'FT_{}_{}.pth'.format(args.net_h,epoch))
print('finish: time={}\n'.format(dt()))