-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_pix2pix.py
181 lines (142 loc) · 5.97 KB
/
train_pix2pix.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
"""
Pix2Pix with U^2-Net Generator Trainer.
Code is adopted from: https://github.com/togheppi/pix2pix
"""
import torch
from torchvision import transforms
from torch.autograd import Variable
from dataset.dataloader import DatasetFromFolder
from models.u2net import U2NET
from models.discriminator import Discriminator
from utilities import plot
import argparse
import os, sys, glob
from torch import nn
import logging
from utilities.iou import jaccard
parser = argparse.ArgumentParser()
parser.add_argument('--batch_size', type=int, default=12, help='train batch size')
parser.add_argument('--num_epochs', type=int, default=200, help='number of train epochs')
parser.add_argument('--lrG', type=float, default=0.0002, help='learning rate for generator, default=0.0002')
parser.add_argument('--lrD', type=float, default=0.0002, help='learning rate for discriminator, default=0.0002')
parser.add_argument('--lamb', type=float, default=100, help='lambda for L1 loss')
parser.add_argument('--beta1', type=float, default=0.5, help='beta1 for Adam optimizer')
parser.add_argument('--beta2', type=float, default=0.999, help='beta2 for Adam optimizer')
params = parser.parse_args()
print(params)
save_dir = 'results/'
if not os.path.exists(save_dir):
os.mkdir(save_dir)
bce_loss_ = nn.BCELoss(size_average=True)
def muti_bce_loss_fusion(d0, d1, d2, d3, d4, d5, d6, labels_v):
loss0 = bce_loss_(d0,labels_v)
loss1 = bce_loss_(d1,labels_v)
loss2 = bce_loss_(d2,labels_v)
loss3 = bce_loss_(d3,labels_v)
loss4 = bce_loss_(d4,labels_v)
loss5 = bce_loss_(d5,labels_v)
loss6 = bce_loss_(d6,labels_v)
loss = loss0 + loss1 + loss2 + loss3 + loss4 + loss5 + loss6
return loss
# Train data
train_data = DatasetFromFolder(folder='train')
train_data_loader = torch.utils.data.DataLoader(dataset=train_data,
batch_size=params.batch_size,
shuffle=True)
# Test data
test_data = DatasetFromFolder(folder='val')
test_data_loader = torch.utils.data.DataLoader(dataset=test_data,
batch_size=8,
shuffle=False)
test_input, test_target = test_data_loader.__iter__().__next__()
chks = sorted(glob.glob('*.pth'), key=os.path.getmtime)
G = U2NET()
G.load_state_dict(torch.load(chks[-1])['model'])
for name, child in G.named_children():
if name in ['last_conv', 'fc']:
for param in child.parameters():
param.requires_grad = False
G.cuda()
D = Discriminator(9, 64, 1)
D.cuda()
D.normal_weight_init(mean=0.0, std=0.02)
# Loss function
BCE_loss = torch.nn.BCELoss().cuda()
L1_loss = torch.nn.L1Loss().cuda()
# Optimizers
G_optimizer = torch.optim.Adam(G.parameters(), lr=params.lrG, betas=(params.beta1, params.beta2))
D_optimizer = torch.optim.Adam(D.parameters(), lr=params.lrD, betas=(params.beta1, params.beta2))
# Training GAN
D_avg_losses = []
G_avg_losses = []
step = 0
for epoch in range(params.num_epochs):
D_losses = []
G_losses = []
# training
for i, (input, target) in enumerate(train_data_loader):
# input & target image data
x_ = Variable(input.cuda())
y_ = Variable(target.cuda())
# Train discriminator with real data
D_real_decision = D(x_, y_).squeeze()
real_ = Variable(torch.ones(D_real_decision.size()).cuda())
D_real_loss = BCE_loss(D_real_decision, real_)
# Train discriminator with fake data
gen_image, d1, d2, d3, d4, d5, d6 = G(x_, classify=False)
D_fake_decision = D(x_, gen_image).squeeze()
fake_ = Variable(torch.zeros(D_fake_decision.size()).cuda())
D_fake_loss = BCE_loss(D_fake_decision, fake_)
# Back propagation
D_loss = (D_real_loss + D_fake_loss) * 0.5
D.zero_grad()
D_loss.backward()
D_optimizer.step()
# Train generator
gen_image, d1, d2, d3, d4, d5, d6 = G(x_, classify=False)
auxiliary = muti_bce_loss_fusion(gen_image, d1, d2, d3, d4, d5, d6, y_)
D_fake_decision = D(x_, gen_image).squeeze()
G_fake_loss = BCE_loss(D_fake_decision, real_)
# L1 loss
l1_loss = params.lamb * L1_loss(gen_image, y_)
# Back propagation
G_loss = G_fake_loss + l1_loss + auxiliary
G.zero_grad()
G_loss.backward()
G_optimizer.step()
# loss values
D_losses.append(D_loss.item())
G_losses.append(G_loss.item())
if step % 50 == 0:
print('Epoch [%d/%d], Step [%d/%d], D_loss: %.4f, G_loss: %.4f'
% (epoch+1, params.num_epochs, i+1, len(train_data_loader), D_loss.item(), G_loss.item()))
sys.stdout.flush()
step += 1
D_avg_loss = torch.mean(torch.FloatTensor(D_losses))
G_avg_loss = torch.mean(torch.FloatTensor(G_losses))
if epoch % 5 == 0:
torch.save({
'model_g': G.state_dict(),
'model_d': D.state_dict(),
}, f'model_epoch_{epoch}.pth')
G.eval()
iou_stats = torch.zeros(5)
for i, (input, target) in enumerate(test_data_loader):
pred, d1, d2, d3, d4, d5, d6 = G(input.cuda(), classify=False)
pred = pred.detach().cpu().int()
target = target.int()
for x in range(pred.shape[0]):
calculated_iou = jaccard(pred[x],target[x])
iou_stats += calculated_iou/pred.shape[0]
print(f'Mean IoU: {torch.mean(iou_stats/i)*100:.2f}')
print(f'Classwise IoU: ')
for x in range(5):
print(f'{x} - {iou_stats[x]/i*100:.2f}')
G.train()
# avg loss values for plot
D_avg_losses.append(D_avg_loss)
G_avg_losses.append(G_avg_loss)
# Show result for test image
gen_image,d1, d2, d3, d4, d5, d6 = G(Variable(test_input.cuda()), classify=False)
gen_image = gen_image.cpu().data
plot.plot_test_result(test_input, test_target, gen_image, epoch, save=True, save_dir=save_dir)