-
Notifications
You must be signed in to change notification settings - Fork 4
/
train.py
214 lines (195 loc) · 8.96 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
# -*- coding: utf-8 -*-
"""
@Time ： 3/18/24 4:22 AM
@Auth ： woldier wong
@File ：train.py
@IDE ：PyCharm
@DESCRIPTION：train
"""
import torch
import numpy as np
from tqdm import tqdm
from torch.utils.data import DataLoader
import time
from utils.eval_utils import SSIM, PSNR, RMSE, denormalize, trunc
from utils.train_valid_utils import get_config, check_dir, config_backpack, init_model, load_dataset, init_optimizer
import datetime
from model import AbstractDenoiser
from torchvision.transforms import Compose, RandomHorizontalFlip, RandomVerticalFlip, ToTensor
import pytz
from utils.img_utils import show_img
import random
from PIL import Image
from accelerate import Accelerator
import sys
sys.path.append('/')
# ================================================================================
# =========================load config=========================================
config_paths = [
r'./config/dn_cnn/config.yml', # _______________0 DnCNN
r'./config/red_cnn/config.yml', # ______________1 REDCNN
r'./config/adnet/config.yml', # ________________2 ADNet
r'./config/ct_former/config.yml', # ____________3 CTformer
r'./config/nb_net/config.yml', # _______________4 NBNet
r'./config/uformer/config.yml', # ______________5 Uformer
r'./config/WiTUnet/config.yml', # ______________6 WiTUnet-Tiny
]
config_path = config_paths[6]
# ================================================================================
tz = pytz.timezone('Asia/Shanghai')
date_str = datetime.datetime.now(tz).strftime("%Y_%m_%d_%H")
def train_loop(net: AbstractDenoiser, train_set, test_set, optimizer: torch.optim.Optimizer, config: dict):
print("===========================woldier Deep Learning Distribution Framework====================================")
data_loader_train = DataLoader(train_set, shuffle=True, batch_size=config["train"]["batch_size"])
data_loader_test = DataLoader(test_set, shuffle=False, batch_size=config["train"]["batch_size"])
# ==================================Accelerator Distribution Training===========================================
accelerator = Accelerator()
device = accelerator.device
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=30, gamma=0.1)
net, optimizer, data_loader_train, data_loader_test, scheduler = accelerator.prepare(
net, optimizer, data_loader_train, data_loader_test, scheduler
)
# =====================================================================================
best_test_loss = 1000
epochs = config["train"]["epochs"]
start_e = config["train"].get("start-epochs", 0)
base_path = date_str + "_" + config["logs"]["name"] # YYYY_mm_DD_HH_XXX
for epoch in range(start_e, start_e + epochs):
start = time.time()
# initialize loss value for every epoch
train_loss, test_loss = 0, 0
# =============================train====================================
train_loss = train_loop_one(
data_loader_train,
epoch,
net,
optimizer,
train_loss,
accelerator
)
# =============================test====================================
ssim, psnr, rmse, test_loss = valid_loop_one(base_path,
data_loader_test,
epoch,
net,
test_loss,
accelerator,
(epoch + 1) % config["train"]["save_img_rate"] == 0)
# =============================logging====================================
best_test_loss = logging(base_path,
best_test_loss,
ssim,
epoch,
epochs,
net,
psnr,
rmse,
start,
test_loss,
train_loss,
accelerator
)
scheduler.step()
def train_loop_one(data_loader_train, epoch, net: AbstractDenoiser, optimizer, train_loss, accelerator):
with tqdm(total=len(data_loader_train), position=0, leave=True) as pbar:
net.train()
for step, batch in enumerate(data_loader_train):
inputs = batch["LDCT"]
labels = batch["FDCT"]
# c = batch['c'].cuda()
outputs, m_loss = net(inputs, labels)
train_loss += m_loss.item()
optimizer.zero_grad()
# m_loss.backward() # backward
accelerator.backward(m_loss) # backward
optimizer.step() # optimizer
pbar.update()
pbar.set_description("epoch%03d: Train Loss %.12f" % (epoch, train_loss / (step + 1))) # 设置描述
pbar.close()
train_loss = train_loss / float(len(data_loader_train))
return train_loss
def valid_loop_one(base_path, data_loader_test, epoch, net: AbstractDenoiser, test_loss, accelerator,
save_img: bool = False):
index = np.random.randint(0, int(len(data_loader_test)), dtype="int")
ssim, psnr, rmse = .0, .0, .0
with tqdm(total=int(len(data_loader_test)), position=0, leave=True) as pbar, torch.no_grad():
net.eval()
for step, batch in enumerate(data_loader_test):
inputs = batch["LDCT"]
labels = batch["FDCT"]
outputs, m_loss = net(inputs, labels)
outputs, labels = accelerator.gather_for_metrics((outputs, labels))
if index == step and save_img:
show_img(labels,
inputs,
outputs,
name="./results/{}/img/{}.jpg".format(base_path, epoch),
)
pass
test_loss += m_loss.item()
# =============回到原始尺度=================
gt = trunc(denormalize(labels.cpu().detach().numpy()))
pre = trunc(denormalize(outputs.cpu().detach().numpy()))
ssim += SSIM(gt, pre)
psnr += PSNR(gt, pre)
rmse += RMSE(gt, pre)
pbar.update()
pbar.set_description("epoch%03d: Valid Loss %.12f" % (epoch, test_loss / (step + 1))) # 设置描述
pbar.close()
test_loss = test_loss / float(len(data_loader_test))
ssim = ssim / float(len(data_loader_test))
psnr = psnr / float(len(data_loader_test))
rmse = rmse / float(len(data_loader_test))
return ssim, psnr, rmse, test_loss
def logging(base_path, best_test_loss, ssim, epoch, epochs, net, psnr, rmse, start, test_loss, train_loss, accelerator):
if best_test_loss > test_loss:
accelerator.wait_for_everyone()
torch.save(accelerator.unwrap_model(net).state_dict(),
"./results/{}/weight/".format(base_path) + "best" + ".pth")
best_test_loss = test_loss
if epoch % 1 == 0:
accelerator.wait_for_everyone()
torch.save(accelerator.unwrap_model(net).state_dict(),
"./results/{}/weight/".format(base_path) + "EPOCH" + str(
epoch) + ".pth")
log_str = '''Epoch #: {}/{}, Time taken: {} secs,\n\tLosses: train_MSE= {},test_MSE={}\n\tSSIM= {}, PSNR= {}, RMSE={}\n'''.format(
epoch, epochs, time.time() - start, train_loss, test_loss, ssim, psnr, rmse)
print(log_str)
f = open("./results/{}/logs/log.txt".format(base_path), "a")
f.writelines(log_str)
f.close() # close file
return best_test_loss
def run():
print(f"loading {config_path}")
config = get_config(config_path)
# check dir
base_dir = "./results/{}/".format(date_str + "_" + config["logs"]["name"])
check_dir(base_dir)
# save config backpack
config_backpack(config_path, base_dir)
# init model
model = init_model(config)
# load dataset
train_dataset, test_dataset = load_dataset(config)
# set transform
def transforms(examples):
random_h = random.random()
random_v = random.random()
compose = Compose(
[
RandomHorizontalFlip(float(random_h > 0.5)),
RandomVerticalFlip(float(random_v > 0.5)),
ToTensor(),
]
)
examples["LDCT"] = [compose(Image.fromarray((np.array(image) * 255).astype(np.uint8))) for image in
examples["LDCT"]]
examples["FDCT"] = [compose(Image.fromarray((np.array(image) * 255).astype(np.uint8))) for image in
examples["FDCT"]]
return examples
train_dataset.set_transform(transforms)
test_dataset.set_transform(transforms)
optim = init_optimizer(model, config)
train_loop(model, train_dataset, test_dataset, optim, config)
if __name__ == '__main__':
run()