flatnet/fns_all.py at flatnet-sep · siddiquesalman/flatnet · GitHub

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
import torch
import scipy.io as sio
import numpy as np
import os
from skimage.color import rgb2gray
import skimage.io
import random
from torch.autograd import Variable
import torch.nn as nn
import torch.nn.functional as F
from skimage.transform import resize as rsz
import torch.optim as optim
import os
from models import*
from torch_vgg import Vgg16


#make train within train for gen and dis
def train_discriminator_epoch(gen, dis, optim_dis, criterion, train_loader, epochs, disc_err,device):
	for param in gen.parameters():
		param.requires_grad = False
	for param in dis.parameters():
		param.requires_grad = True

	for ep in range(epochs):
		for X_train, Y_train in train_loader:
			batchsize = X_train.shape[0]
			target_real = Variable(torch.rand(batchsize,1)*0.5 + 0.7).to(device)
			target_fake = Variable(torch.rand(batchsize,1)*0.3).to(device)
			X_train, Y_train = X_train.to(device), Y_train.to(device)
			optim_dis.zero_grad()
			dis.train()
			high_res_real = Variable(Y_train)
			high_res_fake,_ = gen(X_train)

			dis_loss = criterion(dis(high_res_real), target_real) + criterion(dis(Variable(high_res_fake.data)), target_fake)
			dis_loss.backward()
			optim_dis.step()
			disc_err.append(dis_loss.item())
			print('At Epoch:'+str(ep)+', Pretraining Dis Loss:'+str(dis_loss.item()))
	return disc_err


def validate(gen, dis, vgg, wts, val_loader, gen_criterion, dis_criterion,device):
	k = 0
	tloss = 0
	gen.eval()
	with torch.no_grad():
		for X_val, Y_val in val_loader:
			batchsize = X_val.shape[0]
			ones_const = Variable(torch.ones(batchsize, 1)).to(device)
			# X_val, Y_val = batchGenerator(1, arr[i], h, phase_mask_fwd,device,pathstr)
			X_val, Y_val = X_val.to(device), Y_val.to(device)
			# print(Y_val.shape)
			X_valout,_ = gen(X_val)
			valfeatures_y = vgg(Y_val)
			valfeatures_x = vgg(X_valout)
			if k == 5:
				op = X_valout
			tloss += (wts[0]*(gen_criterion(Y_val, X_valout)+(wts[1]*gen_criterion(valfeatures_y.relu2_2, valfeatures_x.relu2_2))+(wts[1]*gen_criterion(valfeatures_y.relu4_3, valfeatures_x.relu4_3)))+wts[2]*dis_criterion(dis(X_valout), ones_const)).item()
			k += 1
		tloss = tloss/len(val_loader)
	return op, tloss


def train_full_epoch(gen, dis, vgg, wts, optim_gen, optim_dis, train_loader, val_loader,
gen_criterion, dis_criterion, device, vla, e, savedir, train_error, val_error, disc_err,ss,valFreq):
	i = 0

	for X_train, Y_train in train_loader:
		X_train, Y_train = X_train.to(device), Y_train.to(device)
		batchsize = X_train.shape[0]

		#Train discriminator
		ones_const = Variable(torch.ones(batchsize, 1)).to(device)
		target_real = Variable(torch.rand(batchsize,1)*0.5 + 0.7).to(device)
		target_fake = Variable(torch.rand(batchsize,1)*0.3).to(device)

		for param in gen.parameters():
			param.requires_grad = False

		for param in dis.parameters():
			param.requires_grad = True

		optim_dis.zero_grad()
		dis.train()
		high_res_real = Variable(Y_train)
		high_res_fake,_ = gen(X_train)
		dis_loss = dis_criterion(dis(high_res_real), target_real) + dis_criterion(dis(Variable(high_res_fake.data)), target_fake)
		dis_loss.backward()
		optim_dis.step()
		disc_err.append(dis_loss.item())

		#Train generator
		for param in gen.parameters():
			param.requires_grad = True

		for param in dis.parameters():
			param.requires_grad = False

		optim_gen.zero_grad()
		gen.train()
		Xout,_ = gen(X_train)
		features_y = vgg(Y_train)
		features_x = vgg(Xout)
		loss = wts[0]*gen_criterion(Y_train, Xout)+(wts[1]*gen_criterion(features_y.relu2_2, features_x.relu2_2))+(wts[1]*gen_criterion(features_y.relu4_3, features_x.relu4_3))+wts[2]*dis_criterion(dis(Xout), ones_const)
		loss.backward()
		optim_gen.step()
		train_error.append(loss.item())
		if i % valFreq == 0:
			Xvalout, vloss= validate(gen, dis, vgg, wts, val_loader, gen_criterion, dis_criterion, device)
			val_error.append(vloss)
			if vloss < vla:
				vla = vloss
				Xvalout = Xvalout.cpu()
				ims = Xvalout.detach().numpy()
				ims = ims[0, :, :, :]
				ims = np.swapaxes(np.swapaxes(ims,0,2),0,1)
				ims = (ims-np.min(ims))/(np.max(ims)-np.min(ims))
				skimage.io.imsave(savedir+'/best.png', ims)
				dict_save = {
				'gen_state_dict': gen.state_dict(),
				'dis_state_dict': dis.state_dict(),
				'optimizerG_state_dict': optim_gen.state_dict(),
				'optimizerD_state_dict': optim_dis.state_dict(),
				'train_err': train_error,
				'val_err': val_error,
				'disc_err': disc_err,
				'last_finished_epoch': e}
				torch.save(dict_save, savedir+'/best.tar')
				print('Saved best')
		print('Epoch and Iterations::'+str(e)+','+str(i))
		print('Train and Val Loss:'+str(loss.item())+','+str(vloss))
		ss.flush()
		i += 1
	return train_error, val_error, disc_err, vla, Xvalout