DIT_GAN/train.py at main · willy808/DIT_GAN · 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
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
import itertools
import time
import sys
sys.path.append("./groups")
from loss import *
from util import *
from models.generator import *
from models.discriminator import *
from models.basic import *
from dataset import *
from torch.utils.data import ConcatDataset
import torch.optim as optim
from torchmetrics import PeakSignalNoiseRatio,StructuralSimilarityIndexMeasure,MetricCollection
from torchvision.utils import make_grid, save_image
import matplotlib.pyplot as plt
class CycleGANsformer(nn.Module):
    def __init__(self):
        super().__init__()
        os.environ["CUDA_VISIBLE_DEVICES"] = '0,1'
        torch.cuda.set_device(1)
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        print("device", torch.cuda.current_device(), torch.cuda.get_device_name(torch.cuda.current_device()))
        self.discX = TDiscriminator().to(device)
        self.device=device
        self.genX2Y = TGenerator().to(device)
        self.state_dict = {
            'self.discX': self.discX.state_dict(),
            'self.genX2Y': self.genX2Y.state_dict()
        }

        self.opt_discX = optim.AdamW(
            self.discX.parameters(),
            lr=1e-4,
            weight_decay=1e-5,
        )

        self.opt_genX2Y = optim.AdamW(
            self.genX2Y.parameters(),
            lr=5e-5,
            weight_decay=1e-5,
        )
        self.opt_genX2Y_scheduler = optim.lr_scheduler.ReduceLROnPlateau(self.opt_genX2Y, mode='min', factor=0.5, patience=10, verbose=True)#optim.lr_scheduler.StepLR(self.opt_genX2Y, step_size=20, gamma=0.5)
        self.opt_discX_scheduler = optim.lr_scheduler.ReduceLROnPlateau(self.opt_discX, mode='min', factor=0.5, patience=10, verbose=True)
        self.L1 = nn.L1Loss()
        self.mse = nn.MSELoss()
        self.GradientDifferenceLoss = GDL()
        self.HingeLoss = HingeLoss()
        self.Perceptual_loss = Perceptual_loss134()
        self.deblur_percept = Perceptual_loss()
        self.batch_size=2
    def fit(self, dataset, epochs=100):
        loader = DataLoader(dataset, batch_size=self.batch_size, shuffle=True, num_workers=4)
        # testloader=DataLoader(ImageDatasetLoader(mode="test",transformers=transforms_ ), batch_size=self.batch_size, shuffle=False, num_workers=16, pin_memory=True)
        evlloader=DataLoader(ImageDatasetLoader(mode="evl",transformers=transforms_ ), batch_size=self.batch_size, shuffle=False, num_workers=4)


        print("Start the training!")
        step = 0
        for epoch in range(epochs):
            for idx, (x_img, y_img) in enumerate(loader):
                step+=1
                y_img1=y_img.type(torch.cuda.FloatTensor).cuda(non_blocking=True)
                x_img1=x_img.type(torch.cuda.FloatTensor).cuda(non_blocking=True)
                self.genX2Y.train()
                self.discX.train()
#                 assert fake_x.size() == y_img1.size() , f"fake_imgs.size(): {fake_x.size()} real_imgs.size(): {y_img1.size()}"
#######################################################################
######################################################################
                self.opt_genX2Y.zero_grad()
                fake_y1 = self.genX2Y(y_img1, x_img1)
                D_Y_real1, segmap1 = self.discX(fake_y1, y_img1)
                # D_Y_real1, segmap1 = self.discX(fake_y1[2], y_img1)

                generator_loss = self.HingeLoss(D_Y_real1,segmap1,real = False)

                fake_y2 = self.genX2Y(y_img1, x_img1)
                perceptual_loss = self.Perceptual_loss(fake_y2,y_img1)
                # perceptual_loss = self.Perceptual_loss(fake_y2[2],y_img1)

                cycle_X = self.genX2Y(y_img1, x_img1)
                deblur_percept_loss =  self.deblur_percept(cycle_X,x_img1)*1 + self.GradientDifferenceLoss(cycle_X,x_img1)*1

                G_loss = generator_loss*1 + perceptual_loss*0.15 + deblur_percept_loss *0.25   #+ GDL_loss *5

                G_loss.backward(retain_graph=True)
                self.opt_genX2Y.step()

##################################################
                self.opt_discX.zero_grad()
                D_X_real,seg_map2 = self.discX(x_img1,y_img1)

                fake_x = self.genX2Y(y_img1, x_img1)
                D_X_fake,seg_map3 = self.discX(fake_x,y_img1)
                # D_X_fake,seg_map3 = self.discX(fake_x[2],y_img1)
                D_real_loss = self.HingeLoss(D_X_real,seg_map2)
                D_fake_loss = self.HingeLoss(D_X_fake,seg_map3,real = False)
                D_X_loss = D_real_loss*1 + D_fake_loss*1
                D_X_loss.backward(retain_graph=True)
                self.opt_discX.step()
##########################################################################
##########################################################################
            self.opt_genX2Y_scheduler.step(G_loss)
            # self.opt_genX2Y_scheduler.step()
            self.opt_discX_scheduler.step(D_X_loss)


            dir_checkpoint ="./0429lite_checkpoint/"
            torch.save(self.state_dict, dir_checkpoint + f'0429lite_epoch{epoch}.pth')
            print(f'[Epoch {epoch+1}/{epochs}]')
            print(f'[G loss: {G_loss.item()} | generator_loss: {generator_loss.item()} |perceptual_loss: {perceptual_loss.item()} |deblur_percept_loss: {deblur_percept_loss.item()}]')
            print(f'[D loss: { D_X_loss.item()} ]')

            if (epoch) % 2 == 0:
                self.genX2Y.eval()
                with torch.no_grad():
                    ans_ssim = 0
                    ans_psnr_toB = 0
                    ans_psnr_toA = 0
                    for idx, (x_img, y_img) in enumerate(evlloader):
                        y_img1=y_img.type(torch.cuda.FloatTensor).cuda(non_blocking=True)
                        x_img1=x_img.type(torch.cuda.FloatTensor).cuda(non_blocking=True)
                        psnrA = PeakSignalNoiseRatio(data_range=torch.max(x_img1)-torch.min(x_img1)).cuda()
                        psnrB = PeakSignalNoiseRatio(data_range=torch.max(y_img1)-torch.min(y_img1)).cuda()
                        ssim = StructuralSimilarityIndexMeasure().cuda()
                        fake_B = self.genX2Y(y_img1, x_img1)
                        # fake_B = self.deblur(fake_B)

                        ans_ssim +=ssim(fake_B,x_img1)
                        ans_psnr_toB +=psnrB(fake_B,y_img1)
                        ans_psnr_toA +=psnrA(fake_B,x_img1)

                        if idx == len(evlloader)-1:

                            nrows = x_img1.size(0)
                            real_A = make_grid(x_img1, nrow=nrows, normalize=True)
                            fake_B = make_grid(fake_B, nrow=nrows, normalize=True)
                            real_B = make_grid(y_img1, nrow=nrows, normalize=True)

                            image_grid = torch.cat((real_A, fake_B, real_B), 1).cpu().permute(1, 2, 0)

                            plt.figure(figsize=(1.5*nrows, 1.5*3))
                            plt.imshow(image_grid)
                            plt.axis('off')
                            plt.savefig(f'./0429lite_checkpoint/plot/{epoch}.png')
                            plt.show()
                            plt.close(image_grid)


                    print(f"Epoch {epoch+1} | ssim : {ans_ssim/len(evlloader)}")
                    print(f"Epoch {epoch+1} | psnr_toB : {ans_psnr_toB/len(evlloader)}")
                    print(f"Epoch {epoch+1} | psnr_toA : {ans_psnr_toA/len(evlloader)}")


                    cbct_count=0
                    path="./0429/"+str(epoch)+"/"
                    if not os.path.exists(path):
                        os.mkdir(path)
                    cbct_data=[]
                    for i in range(len(data)):
                        ds=pydicom.dcmread(data[i][0],force=True)
                        cbct_data.append(ds)
                    cbctdata=cbct_data[24600:24704]


                    start_time = time.time()
                    for idx, (a_img, b_img) in enumerate(evlloader):

                        self.genX2Y.eval()
                        a=a_img.type(torch.cuda.FloatTensor)
                        b=b_img.type(torch.cuda.FloatTensor)

                        fake_B = self.genX2Y(b,a).detach()
                        b = fake_B.cpu().numpy()
                        c =b.copy()
                        for s in range(len(a_img)):
                            t=s+idx*self.batch_size
                            write_dicom(cbctdata[t],c[s],path,t)
                    end_time = time.time()
                    execution_time = end_time - start_time
                    print("Execution time:", execution_time, "seconds")
if __name__ == "__main__":
    # Create an instance of CycleGANsformer
    cg = CycleGANsformer()

    # Create instances of ImageDatasetLoader for seta and setb
    seta = ImageDatasetLoader(mode="train1", transformers=transforms_)
    setb = ImageDatasetLoader(mode="train", transformers=data_transform)

    # Concatenate the datasets into trainset
    trainset = ConcatDataset([seta, setb])

    # Train the CycleGAN model using seta
    cg.fit(seta)