-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
179 lines (149 loc) · 4.84 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
import numpy as np
from numpy.fft import fft2, ifft2, ifftshift
import matplotlib.pyplot as plt
from scipy.io import savemat, loadmat
from tqdm import tqdm
import argparse
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
from torchvision.transforms import Compose, Resize, RandomCrop, Grayscale, ToTensor, RandomVerticalFlip
from torchvision.transforms import RandomResizedCrop, Grayscale, ToTensor, RandomVerticalFlip
from utils.utils_deblur import gauss_kernel
from utils.dataloader import Flickr2K, PoissBlur_List
from models.network_p4ip import P4IP_Net
LEARNING_RATE = 1e-4
NUM_EPOCHS = 101
BATCH_SIZE = 5
N_TRAIN = 128
N_VAL = 256
"""
Initiate a model, and transfer to gpu
"""
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = P4IP_Net()
model.to(device)
print("Number of GPUS available: ", torch.cuda.device_count())
"""
Setting up training data - blur kernels and photon levels
"""
# Adding Gaussian kernels
kernel_list = []
counts = [1,1,1,1,1,1,1,1,1,1]
idx= 0
for sigma in np.linspace(0.1, 2.5, 10):
for count in range(counts[idx]):
kernel_list.append(gauss_kernel(64,sigma))
idx +=1
# Adding Blur Kernels
struct = loadmat('data/motion_kernels.mat')
motion_kernels = struct['PSF_list'][0]
for idx in range(len(motion_kernels)):
kernel = motion_kernels[idx]
kernel = np.clip(kernel,0,np.inf)
kernel = kernel/np.sum(kernel.ravel())
kernel_list.append(kernel)
"""
Transform image and blur operations
"""
transform_img_train = Compose([Resize([N_VAL,N_VAL]),
RandomCrop([N_TRAIN,N_TRAIN]),
Grayscale(),
RandomVerticalFlip(),
ToTensor()])
transform_blur_train = PoissBlur_List(kernel_list, [1,60], N_TRAIN, True)
transform_img_val = Compose([Resize([N_VAL,N_VAL]),
Grayscale(),
ToTensor()])
transform_blur_val = PoissBlur_List(kernel_list, [1,60], N_VAL, False)
# Dataloaders
data_train = Flickr2K(True, transform_img_train, transform_blur_train)
data_val = Flickr2K(False, transform_img_val, transform_blur_val)
train_loader = DataLoader(data_train, batch_size=BATCH_SIZE, shuffle=False, num_workers=0)
val_loader = DataLoader(data_val, batch_size=BATCH_SIZE, shuffle=False, num_workers=0)
"""
Setting up training with :
1. L1 Loss
2. AdamOptimizer
"""
criterion_list = [torch.nn.L1Loss()]
wt_list = [1.0]
criterion_l2 = torch.nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)
"""
Training starts here
"""
for epoch in range(NUM_EPOCHS):
epoch_loss = 0
model.train()
"""
Training Epoch
"""
with tqdm(total=len(data_train), desc=f'Epoch {epoch + 1}/{NUM_EPOCHS}', unit='its') as pbar:
for i, data in enumerate(train_loader):
"""
Get training data - [true image, noisy_blurred, kernel, photon level]
"""
x, y, kernels, M = data
M = M.view( x.size(0), 1, 1, 1)
x, y, kernels, M = x.to(device), y.to(device), kernels.to(device), M.to(device)
"""
Forward Pass => calculating loss => computing gradients => Adam optimizer update
"""
# Forward Pass
optimizer.zero_grad()
x = x.type(torch.cuda.FloatTensor)
out_list = model(y, kernels, M)
out = out_list[-1]
# Calculating training loss
loss = 0
for idx in range(len(wt_list)):
loss += wt_list[idx]*criterion_list[idx](out.float(), x.float())
# Backprop
loss.backward()
# Adam optimizer step
optimizer.step()
epoch_loss += loss.item()
pbar.update(BATCH_SIZE)
pbar.set_postfix(**{'loss (batch)': loss.item()})
epoch_loss = epoch_loss*BATCH_SIZE/len(data_train)
print('Epoch: {}, Training Loss: {}, Current Learning Rate: {}'.format(epoch+1,epoch_loss,LEARNING_RATE))
"""
Validation Epoch
"""
val_loss, mse = 0, 0
model.eval()
with torch.no_grad(): # Don't maintain computation graph since no backprop reqd., saves GPU memory
for i, data in enumerate(val_loader):
"""
Getting validation pair
"""
x, y, kernels, M = data
x, y, kernels, M = x.type(torch.DoubleTensor).to(device), y.to(device), kernels.to(device), M.to(device)
M = M.view( x.size(0), 1, 1, 1)
"""
Forward Pass
"""
out_list = model(y, kernels, M)
out = out_list[-1]
"""
Calculating L2 loss and training loss on the validation set
"""
loss = 0
for idx in range(len(wt_list)):
loss += wt_list[idx]*criterion_list[idx](out.float(), x.float())
loss_l2 = criterion_l2(out, x)
val_loss += loss.item()
mse += loss_l2.item()
val_loss = val_loss*BATCH_SIZE/len(data_val)
mse = mse*BATCH_SIZE/len(data_val)
psnr = -10*np.log10(mse)
"""
Writing the epoch loss, validation loss to tensorboard for visualization
"""
print('Validation PSNR: %0.3f, Validation Loss: %0.6f'%(psnr, val_loss))
for param_group in optimizer.param_groups:
LEARNING_RATE = param_group['lr']
if epoch % 10 ==0:
torch.save(model.state_dict(), 'model_zoo/p4ip_net_%depoch.pth'%(epoch))