-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
170 lines (143 loc) · 7.81 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
# -*- coding: utf-8 -*-
import os
import numpy as np
import torch
from torch import nn
from torch import optim
import torch.backends.cudnn as cudnn
from torch.utils.data import DataLoader
from tensorboardX import SummaryWriter
from options import args
from networks import VGG, EncoderDecoder
from losses import percept_loss
from optimizer import build_optimizer # ##
from config import _C # ##
from utils import build_dataset, mkdir, get_lr, adjust_lr, Visualizer
from evaluation import create_window, calc_psnr, calc_ssim
# Use GPU
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_ids
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
cudnn.benchmark = True
np.random.seed(0)
# Load dataset
trainbase = build_dataset(args.dataset, args.data_dir, args.scale_size, isTraining=True)
train_loader = DataLoader(trainbase, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=args.pin_memory)
valbase = build_dataset(args.dataset, args.data_dir, args.scale_size, isTraining=False)
val_loader = DataLoader(valbase, batch_size=1, shuffle=False, num_workers=args.num_workers, pin_memory=args.pin_memory)
# Build model
Denoise_Net = EncoderDecoder(window_size=args.window_size, isRes=args.isRes, isSwinT=args.isSwinT).to(device)
if args.load_model != "":
Denoise_Net = torch.load(os.path.join(args.models_dir, args.dataset, args.name, args.load_model), map_location=device)
Denoise_Net = nn.DataParallel(Denoise_Net)
fe_layer = VGG(args.model).to(device)
fe_layer = nn.DataParallel(fe_layer)
# Set up optimization
# optimizer = optim.Adam(Denoise_Net.parameters(), lr=args.init_lr, weight_decay=args.weight_decay) ##
optimizer = build_optimizer(_C, Denoise_Net) # ##
criterion = nn.L1Loss()
# Create logs
mkdir(os.path.join(args.models_dir, args.dataset, args.name))
viz = Visualizer(env=args.name + "_" + args.dataset)
writer = SummaryWriter(os.path.join(args.logs_dir, args.dataset, args.name))
dt_size = len(train_loader.dataset)
checkpoint_path = os.path.join(args.models_dir, args.dataset, "{net}", "{epoch}-{loss}.pth")
f = open(os.path.join(args.logs_dir, args.dataset, args.name, "options.txt"), "w")
f.write(str(args))
f.write("\n")
f.write(str(_C))
# print(_C, f) <_io.TextIOWrapper name='logs/CT/STUNet/options.txt' mode='w' encoding='UTF-8'>##
f.close()
f = open(os.path.join(args.logs_dir, args.dataset, args.name, "network.txt"), "w")
f.write(str(Denoise_Net))
f.close()
# Start training
best_loss = float("inf")
best_epoch = 0
window = create_window(11, 1).to(device)
print("Start training...")
for epoch in range(args.load_epoch, args.num_epochs):
epoch_loss = 0.0
step = 0
Denoise_Net.train(mode=True)
# fe_layer.train(mode=True)##
for (low, high) in train_loader:
step += 1
img_low = low.to(device)
img_high = high.to(device)
# zero the parameter gradients
optimizer.zero_grad()
pred_high = Denoise_Net(img_low)
loss = percept_loss(pred_high, img_high, fe_layer, criterion=criterion, eps=args.eps,
weight_rank=args.weight_rank, weight_l1=args.weight_l1, weight_percept=args.weight_percept) # ##
# loss = criterion(pred_high, img_high) ##
viz.img(name="in_low", img_=(img_low[0, :, :, :]+1.0)/2.0)
viz.img(name="in_high", img_=(img_high[0, :, :, :]+1.0)/2.0)
viz.img(name="out_high", img_=(pred_high[0, :, :, :]+1.0)/2.0)
loss.backward()
optimizer.step()
epoch_loss += loss.item()
# Loss of current batch
niter = epoch * len(train_loader) + step
writer.add_scalars("train_loss", {"train_loss": loss.item()}, niter)
print("Epoch %d, Step %d / %d, Train loss: %0.4f"
% (epoch, step, (dt_size - 1) // train_loader.batch_size + 1, loss.item()))
with open(os.path.join(args.logs_dir, args.dataset, args.name, "loss_log.txt"), "a") as log_file: # Save the message
log_file.write("Epoch %d, Step %d / %d, Train loss: %0.4f\n"
% (epoch, step, (dt_size - 1) // train_loader.batch_size + 1, loss.item()))
viz.plot("train loss", loss.item())
# Write current lr
current_lr = get_lr(optimizer)
viz.plot("learning rate", current_lr)
writer.add_scalars("learning_rate", {"lr": current_lr}, niter)
print("epoch %d loss: %0.4f\nlearning rate: %f" % (epoch, epoch_loss / step, current_lr))
with open(os.path.join(args.logs_dir, args.dataset, args.name, "loss_log.txt"), "a") as log_file: # Save the message
log_file.write("epoch %d loss: %0.4f\nlearning rate: %f\n" % (epoch, epoch_loss / step, current_lr))
adjust_lr(optimizer, args.init_lr, epoch, args.num_epochs, power=args.power)
if (epoch + 1) % args.val_epoch_freq == 0 or epoch + 1 == args.num_epochs:
Denoise_Net.eval()
# fe_layer.eval()##
loss_lst = []
psnr_lst = []
ssim_lst = []
with torch.no_grad():
for (low, high) in val_loader:
img_low = low.to(device)
img_high = high.to(device)
pred_high = Denoise_Net(img_low)
viz.img(name="te_low", img_=(img_low[0, :, :, :]+1.0)/2.0)
viz.img(name="te_high", img_=(img_high[0, :, :, :]+1.0)/2.0)
viz.img(name="pred_high", img_=(pred_high[0, :, :, :]+1.0)/2.0)
loss = percept_loss(pred_high, img_high, fe_layer, criterion=criterion, eps=args.eps,
weight_rank=args.weight_rank, weight_l1=args.weight_l1, weight_percept=args.weight_percept) # ##
# loss = criterion(pred_high, img_high) ##
psnr = calc_psnr(pred_high, img_high)
ssim = calc_ssim(pred_high, img_high, window, 11, 1)
loss_lst.append(loss.item())
psnr_lst.append(psnr.mean(-1).item())
ssim_lst.append(ssim.mean(-1).item())
loss_arr = np.array(loss_lst)
psnr_arr = np.array(psnr_lst)
ssim_arr = np.array(ssim_lst)
writer.add_scalars("val_loss", {"val_loss": loss_arr.mean()}, epoch + 1)
viz.plot("val loss", loss_arr.mean())
print("Val Loss: " + str(loss_arr.mean()) + " ± " + str(loss_arr.std()))
writer.add_scalars("val_psnr", {"val_psnr": psnr_arr.mean()}, epoch + 1)
viz.plot("val psnr", psnr_arr.mean())
print("Val PSNR: " + str(psnr_arr.mean()) + " ± " + str(psnr_arr.std()))
writer.add_scalars("val_ssim", {"val_ssim": ssim_arr.mean()}, epoch + 1)
viz.plot("val ssim", ssim_arr.mean())
print("Val SSIM: " + str(ssim_arr.mean()) + " ± " + str(ssim_arr.std()))
torch.save(Denoise_Net, checkpoint_path.format(net=args.name, epoch=epoch+1, loss=loss_arr.mean()))
print("Model saved - Epoch: %d, Val Loss: %f, Val PSNR: %f, Val SSIM: %f"
% (epoch + 1, loss_arr.mean(), psnr_arr.mean(), ssim_arr.mean()))
with open(os.path.join(args.logs_dir, args.dataset, args.name, "loss_log.txt"), "a") as log_file: # Save the message
log_file.write("Model saved - Epoch: %d, Val Loss: %f, Val PSNR: %f, Val SSIM: %f\n"
% (epoch + 1, loss_arr.mean(), psnr_arr.mean(), ssim_arr.mean()))
if best_loss >= loss_arr.mean():
best_loss = loss_arr.mean()
best_epoch = epoch + 1
torch.save(Denoise_Net, checkpoint_path.format(net=args.name, epoch="best", loss="best"))
print("Best model saved - Epoch: %d, Val Loss: %f" % (best_epoch, best_loss))
with open(os.path.join(args.logs_dir, args.dataset, args.name, "loss_log.txt"), "a") as log_file: # Save the message
log_file.write("Best model saved - Epoch: %d, Val Loss: %f\n" % (best_epoch, best_loss))
print("Training finished. ")