forked from ytpeng-aimlab/Histoformer
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
236 lines (177 loc) · 8.36 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
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
import os
import cv2
import numpy as np
import skimage.io
import torch
import torch.nn as nn
import torch.nn.functional as F
import matplotlib.pyplot as plt
import argparse
from torch.autograd import Variable
from timm.models.layers import DropPath, trunc_normal_
from torchvision import models
from torchvision import transforms
from datasets import *
from utils import *
from Generator import *
from Discriminator import *
from loss import *
from model_histoformer import *
# Training settings
parser = argparse.ArgumentParser(description='histogram_network')
# global settings
parser.add_argument('--epochs', type=int, default=150, help='the starting epoch count')
parser.add_argument('--lr', type=float, default=0.0002, help='initial learning rate')
parser.add_argument('--weight_decay', type=float, default=0.02, help='weight decay')
# parser.add_argument('--train_dir', type=str, default ='./data/UIEB/train', help='dir of train data')
# parser.add_argument('--gt_dir', type=str, default ='./data/UIEB/train_gt', help='dir of gt data')
# parser.add_argument('--val_dir', type=str, default ='./data/UIEB/test', help='dir of test data (use test data as val)')
# args for Histoformer
parser.add_argument('--norm_layer', type=str, default ='nn.LayerNorm', help='normalize layer in transformer')
parser.add_argument('--embed_dim', type=int, default=32, help='dim of emdeding features')
parser.add_argument('--token_projection', type=str,default='linear', help='linear/conv token projection')
parser.add_argument('--token_mlp', type=str,default='TwoDCFF', help='TwoDCFF/ffn token mlp')
parser.add_argument('--save_dir', type=str, default ='./checkpoints/onlyinter/', help='save dir')
parser.add_argument('--save_image_dir', type=str, default ='./results/', help='save image dir')
# parser.add_argument('--weight_gan', type=float, default=0.4, help='the weight of gan')
opt = parser.parse_args()
### Data ###
trainloader= get_training_set()
valloader= get_val_set()
testloader= get_test_set()
### Model ###
model = Histoformer(embed_dim=opt.embed_dim,token_projection='linear',token_mlp='TwoDCFF')
### Cuda/GPU ###
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
cuda = True if torch.cuda.is_available() else False
model = model.cuda()
net_d = Discriminator().cuda()
net_g = Generator().cuda()
net_d.apply(weights_init_normal)
net_g.apply(weights_init_normal)
#Load model
# model = torch.load('/home/roger/rong/checkpoints/onlyinter/Histoformer-PQR_200_modifyloss.pth')
# net_d = torch.load('/home/roger/rong/checkpoints/onlyinter/Histoformer-PQR_netD_200_modifyloss.pth')
# net_g = torch.load('/home/roger/rong/checkpoints/onlyinter/Histoformer-PQR_netG_200_modifyloss.pth')
### Optimizer ###
optimizer = torch.optim.AdamW(model.parameters(), lr=opt.lr, betas=(0.9, 0.999),eps=1e-8, weight_decay=opt.weight_decay) #
optimizer_d = torch.optim.Adam(net_d.parameters(), lr=opt.lr, betas=(0.5, 0.999))
optimizer_g = torch.optim.Adam(net_g.parameters(), lr=opt.lr, betas=(0.5, 0.999))
### Loss ###
criterionGAN = GANLoss().cuda()
criterionL1 = nn.L1Loss().cuda()
perceploss = VGG19_PercepLoss().cuda()
contentloss = VGG19_Content().cuda()#.eval()
weight_gan = 0.4
### Train ###
best_loss = 1e10
best_loss_v = 1e10
Loss_avg=[]
Loss_avg_v=[]
for e in range(opt.epochs):
total_loss_g = 0
torch.cuda.empty_cache()
model.train()
for i, (input_img, label_img, ori_img, hs_img) in enumerate(trainloader): #, hs_img
# print('i',i)
input_img = input_img.to(device)
label_img = label_img.to(device)
optimizer.zero_grad() #retain_graph=True
pred_img = model(input_img)
# exit()
loss_list=[]
percep_loss_list=[]
hist_img_list=[]
gt_img_list=[]
R_out = pred_img[:,0]
G_out = pred_img[:,1]
B_out = pred_img[:,2]
R_labels = label_img[:,0]
G_labels = label_img[:,1]
B_labels = label_img[:,2]
R_loss = L2_histo(R_out,R_labels)
G_loss = L2_histo(G_out,G_labels)
B_loss = L2_histo(B_out,B_labels)
for j in range(len(ori_img)):
RGB_hs_img0, gt0 = hist_match(ori_img[j],hs_img[j], R_out[j], G_out[j], B_out[j]) #RGB_hs_img:numpy (460,620,3)
RGB_hs_img = transforms.ToTensor()(RGB_hs_img0)
RGB_hs_img = RGB_hs_img.unsqueeze(0).cuda() #torch.Size([1, 3, 460, 620])
gt = transforms.ToTensor()(gt0)
gt = gt.unsqueeze(0).cuda()
hist_img_list.append(RGB_hs_img0)
gt_img_list.append(gt0)
loss_mae = criterionL1(RGB_hs_img, gt)
loss_percep = perceploss(RGB_hs_img,gt)
loss_list.append(loss_mae)
percep_loss_list.append(loss_percep) ###
mae_loss = sum(loss_list)
mae_loss = mae_loss/len(ori_img)
percep_loss = sum(percep_loss_list) ###
percep_loss = percep_loss/len(ori_img) ###
RGB_loss = (2*R_loss)+(0.5*G_loss)+(1*B_loss)
loss = torch.mean(RGB_loss)
loss = loss + (0.6*mae_loss) + (0.6*percep_loss)
loss.backward()
optimizer.step()
for k in range(len(hist_img_list)):
RGB_hs_img0 = hist_img_list[k]
gt0 = gt_img_list[k]
RGB_hs_img1 = align_to_four(RGB_hs_img0)
RGB_hs_img1 = npTOtensor(RGB_hs_img1)
gt1 = align_to_four(gt0)
gt1 = npTOtensor(gt1)
### forward
real_a = RGB_hs_img1
real_b = gt1
# print('real_a.shape:',real_a.shape) # ([1, 3, 460, 620])
# print('real_b.shape:',real_b.shape)
fake_b = net_g(real_a)
fake_b.data.clamp_(-1, 1)
######################
# (1) Update D network
######################
optimizer_d.zero_grad()
# train with fake
pred_fake = net_d(fake_b) #,real_a
loss_d_fake = criterionGAN(pred_fake, False)
# train with real
pred_real = net_d(real_b)
loss_d_real = criterionGAN(pred_real, True)
# Combined D loss
loss_d = (loss_d_fake + loss_d_real) * 0.5
loss_d.backward(retain_graph=True) #retain_graph=True
optimizer_d.step()
######################
# (2) Update G network
######################
content_loss1 = contentloss(fake_b,real_b,'relu1_1')
content_loss2 = contentloss(fake_b,real_b,'relu1_2')
content_loss3 = contentloss(fake_b,real_b,'relu2_1')
content_loss4 = contentloss(fake_b,real_b,'relu2_2')
loss_content = content_loss1 + content_loss2 + content_loss3 + content_loss4
optimizer_g.zero_grad()
# First, G(A) should fake the discriminator
pred_fake = net_d(fake_b)
loss_g_gan = criterionGAN(pred_fake, True)
# Second, G(A) = B
loss_g_l1 = criterionL1(fake_b, real_b)
# print('fake_b',fake_b.shape,real_b.shape)
loss_g = ((loss_g_gan + loss_g_l1)*weight_gan + loss_content*0.3) #+ mae_percep_loss*(1-weight_gan)
total_loss_g += loss_g.item()
loss_g.backward()
optimizer_g.step()
loss_gan = torch.tensor(total_loss_g)/len(ori_img)
loss = loss + loss_gan #
# print('i',i)
if i%20==0:
print('epoch: {} , batch: {}, loss: {}'.format(e + 1+200, i + 1, loss.data))
if (e+1)%10 == 0 or (e+1+200)>150 or e==0:
torch.save({'state_dict': model.state_dict(),
'optimizer' : optimizer.state_dict()
}, os.path.join(opt.save_dir,"Histoformer-PQR_{}_modifyloss.pth".format(e+1)))
torch.save({'state_dict': net_g.state_dict(),
'optimizer' : optimizer_g.state_dict()
}, os.path.join(opt.save_dir,"Histoformer-PQR_netG_{}_modifyloss.pth".format(e+1)))
torch.save({'state_dict': net_d.state_dict(),
'optimizer' : optimizer_d.state_dict()
}, os.path.join(opt.save_dir,"Histoformer-PQR_netD_{}_modifyloss.pth".format(e+1)))