-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
188 lines (137 loc) · 5.79 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
from __future__ import absolute_import, division, print_function
import argparse
import cv2
import os
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchvision
# import torchvision.models as pt_models
import dataset as dataset
from vgg import *
from torch.optim.lr_scheduler import StepLR
from torch.utils.data import Dataset, DataLoader
from torch.autograd import Variable
from options import opt, device
from model import *
from misc import *
# from progress.bar import Bar
import re
import sys
import clip
import torchvision.transforms as T
from ssim import *
import wandb
def get_lr(optimizer):
for param_group in optimizer.param_groups:
return param_group['lr']
if __name__ == '__main__':
#scale = opt.scale
# run_name = 'CADAqua_lol_check'
# run = wandb.init(project="CADAqua", config=opt, name=run_name)
transform_re = T.Resize(size = (224,224))
transform_PIL = T.ToPILImage()
model, preprocess = clip.load("ViT-B/32", device=device)
print("Underwater Image Enhancement")
netG = U_Restormer()
netG.to(device)
L1_loss = nn.L1Loss()
mse_loss = nn.MSELoss()
ssim_loss = SSIMLoss(11)
vgg = Vgg16(requires_grad=False).to(device)
optim_g = optim.Adam(netG.parameters(),
lr=opt.learning_rate_g,
betas = (opt.beta1, opt.beta2),
weight_decay=opt.wd_g)
#scheduler = StepLR(optim_g, step_size=25, gamma=0.1)
dataset = dataset.Dataset_Load(data_path = opt.data_path,
transform=dataset.ToTensor()
)
batches = int(dataset.len / opt.batch_size)
dataloader = DataLoader(dataset, batch_size=opt.batch_size, shuffle=True)
if not os.path.exists(opt.checkpoints_dir):
os.makedirs(opt.checkpoints_dir)
models_loaded = getLatestCheckpointName()
latest_checkpoint_G = models_loaded
print('loading model for generator ', latest_checkpoint_G)
if latest_checkpoint_G == None :
start_epoch = 1
print('No checkpoints found for netG and netD! retraining')
else:
checkpoint_g = torch.load(os.path.join(opt.checkpoints_dir, latest_checkpoint_G))
start_epoch = checkpoint_g['epoch'] + 1
netG.load_state_dict(checkpoint_g['model_state_dict'])
optim_g.load_state_dict(checkpoint_g['optimizer_state_dict'])
for param_group in optim_g.param_groups:
param_group['lr'] = opt.learning_rate_g
print('Restoring model from checkpoint ' + str(start_epoch))
netG.train()
for epoch in range(start_epoch, opt.end_epoch + 1):
# bar = Bar('Training', max=batches)
opt.total_mse_loss = 0.0
opt.total_vgg_loss = 0.0
opt.total_ssim_loss = 0.0
opt.total_G_loss = 0.0
for i_batch, sample_batched in enumerate(dataloader):
hazy_batch = sample_batched['hazy']
clean_batch = sample_batched['clean']
hazy_batch = hazy_batch.to(device)
clean_batch = clean_batch.to(device)
optim_g.zero_grad()
pred_batch = netG(hazy_batch)
batch_mse_loss = torch.mul(opt.lambda_mse, L1_loss(pred_batch, clean_batch))
batch_mse_loss.backward(retain_graph=True)
batch_ssim_loss = torch.mul(opt.lambda_ssim, ssim_loss(pred_batch, clean_batch))
batch_ssim_loss.backward(retain_graph=True)
batch_vgg_loss = 0.0
for i in range(opt.batch_size):
pred_image_batch = transform_re(pred_batch[i])
pred_image_batch = transform_PIL(pred_image_batch)
pred_image_features = preprocess(pred_image_batch).unsqueeze(0).to(device)
clean_image_batch = transform_re(clean_batch[i])
clean_image_batch = transform_PIL(clean_image_batch)
clean_image_features = preprocess(clean_image_batch).unsqueeze(0).to(device)
pred_image_features = model.encode_image(normalize_batch(pred_image_features))
clean_image_features = model.encode_image(normalize_batch(clean_image_features))
batch_vgg_loss += torch.mul(opt.lambda_vgg, mse_loss(pred_image_features, clean_image_features))
batch_vgg_loss.backward(retain_graph=True)
opt.batch_mse_loss = batch_mse_loss.item()
opt.total_mse_loss += opt.batch_mse_loss
opt.batch_ssim_loss = batch_ssim_loss.item()
opt.total_ssim_loss += opt.batch_ssim_loss
opt.batch_vgg_loss = batch_vgg_loss.item()
opt.total_vgg_loss += opt.batch_vgg_loss
opt.batch_G_loss = opt.batch_mse_loss + opt.batch_vgg_loss + opt.batch_ssim_loss
opt.total_G_loss += opt.batch_G_loss
optim_g.step()
#scheduler.step()
# bar.suffix = f' Epoch : {epoch} | ({i_batch+1}/{batches}) | ETA: {bar.eta_td} | g_mse: {opt.batch_mse_loss} | g_vgg: {opt.batch_vgg_loss}'
print('\r Epoch : ' + str(epoch) + ' | (' + str(i_batch+1) + '/' + str(batches) + ') | mse: ' + str(opt.batch_mse_loss) + ' | clip: ' + str(opt.batch_vgg_loss) + ' | ssim: ' + str(opt.batch_ssim_loss), end='', flush=True)
# bar.next()
model_params = {}
for name, param in netG.named_parameters():
if param.requires_grad:
model_params[name] = param.data
print('\nFinished ep. %d, lr = %.6f, total_mse = %.6f, total_clip = %.6f, total_ssim = %.6f' % (epoch, get_lr(optim_g), opt.total_mse_loss, opt.total_vgg_loss, opt.total_ssim_loss))
# print('training epoch %d, %d / %d patches are finished, g_mse = %.6f' % (
# epoch, i_batch, batches, opt.batch_mse_loss))
# wandb.log({
# 'epoch':epoch,
# 'L1_loss':opt.total_mse_loss,
# 'vgg_loss':opt.total_vgg_loss,
# 'ssim_loss':opt.total_ssim_loss,
# 'total_loss':opt.total_G_loss,
# 'alpha_r':model_params['alph'],
# 'beta_g':model_params['beta'],
# 'gamma_b':model_params['gama']
# })
torch.save({'epoch':epoch,
'model_state_dict':netG.state_dict(),
'optimizer_state_dict':optim_g.state_dict(),
'L1_loss':opt.total_mse_loss,
'vgg_loss':opt.total_vgg_loss,
'ssim_loss':opt.total_ssim_loss,
'opt':opt,
'total_loss':opt.total_G_loss}, os.path.join(opt.checkpoints_dir, 'netG_' + str(epoch) + '.pt'))