CapsVoxGAN/gan.py at master · CPUFronz/CapsVoxGAN · 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
import os
import torch
from tqdm import tqdm
from torch.utils.tensorboard import SummaryWriter

from voxeldata import VoxelData
from constants import DATASET_HDF
from constants import CUBE_SIZE
from constants import LEAK_VALUE
from constants import Z_SIZE
from constants import EPOCHS
from constants import BATCH_SIZE
from constants import DISCRIMINATOR_LEARNING_RATE
from constants import GENERATOR_LEARNING_RATE
from constants import DISCRIMINATOR_THRESHOLD
from constants import LOG_PATH
from constants import GENERATED_PATH
from constants import CLASSES
from constants import RANDOM_SEED
from constants import SAVED_DISCRIMINATOR
from constants import SAVED_GENERATOR

torch.manual_seed(RANDOM_SEED)


class Generator(torch.nn.Module):
    def __init__(self, cube_size=CUBE_SIZE, z_size=Z_SIZE):
        super(Generator, self).__init__()
        self.cube_size = cube_size
        self.z_size = z_size

        self.layer1 = torch.nn.Sequential(
            torch.nn.ConvTranspose3d(self.z_size, self.cube_size * 8, kernel_size=4, stride=2, padding=(1, 1, 1)),
            torch.nn.BatchNorm3d(self.cube_size * 8),
            torch.nn.ReLU()
        )
        self.layer2 = torch.nn.Sequential(
            torch.nn.ConvTranspose3d(self.cube_size * 8, self.cube_size * 4, kernel_size=4, stride=2, padding=(1, 1, 1)),
            torch.nn.BatchNorm3d(self.cube_size * 4),
            torch.nn.ReLU()
        )
        self.layer3 = torch.nn.Sequential(
            torch.nn.ConvTranspose3d(self.cube_size * 4, self.cube_size * 2, kernel_size=4, stride=2, padding=(1, 1, 1)),
            torch.nn.BatchNorm3d(self.cube_size * 2),
            torch.nn.ReLU()
        )
        self.layer4 = torch.nn.Sequential(
            torch.nn.ConvTranspose3d(self.cube_size * 2, self.cube_size, kernel_size=4, stride=2, padding=(1, 1, 1)),
            torch.nn.BatchNorm3d(self.cube_size),
            torch.nn.ReLU()
        )
        self.layer5 = torch.nn.Sequential(
            torch.nn.ConvTranspose3d(self.cube_size, 1, kernel_size=4, stride=2, padding=(1, 1, 1)),
            torch.nn.Sigmoid()
        )

    def forward(self, x):
        out = x.view(-1, self.z_size, 1, 1, 1)  # torch.Size([BATCH_SIZE, 200, 1, 1, 1])
        out = self.layer1(out)                  # torch.Size([BATCH_SIZE, 256, 2, 2, 2])
        out = self.layer2(out)                  # torch.Size([BATCH_SIZE, 128, 4, 4, 4])
        out = self.layer3(out)                  # torch.Size([BATCH_SIZE, 64, 8, 8, 8])
        out = self.layer4(out)                  # torch.Size([BATCH_SIZE, 32, 16, 16, 16])
        out = self.layer5(out)                  # torch.Size([BATCH_SIZE, 1, 32, 32, 32])
        return out


class Discriminator(torch.nn.Module):
    def __init__(self, cube_size=CUBE_SIZE, leak_value=LEAK_VALUE):
        super(Discriminator, self).__init__()
        self.cube_size = cube_size
        self.leak_value = leak_value

        self.layer1 = torch.nn.Sequential(
            torch.nn.Conv3d(1, self.cube_size, kernel_size=4, stride=2, padding=(1, 1, 1)),
            torch.nn.BatchNorm3d(self.cube_size),
            torch.nn.LeakyReLU(self.leak_value)
        )
        self.layer2 = torch.nn.Sequential(
            torch.nn.Conv3d(self.cube_size, self.cube_size * 2, kernel_size=4, stride=2, padding=(1, 1, 1)),
            torch.nn.BatchNorm3d(self.cube_size * 2),
            torch.nn.LeakyReLU(self.leak_value)
        )
        self.layer3 = torch.nn.Sequential(
            torch.nn.Conv3d(self.cube_size * 2, self.cube_size * 4, kernel_size=4, stride=2, padding=(1, 1, 1)),
            torch.nn.BatchNorm3d(self.cube_size * 4),
            torch.nn.LeakyReLU(self.leak_value)
        )
        self.layer4 = torch.nn.Sequential(
            torch.nn.Conv3d(self.cube_size * 4, self.cube_size * 8, kernel_size=4, stride=2, padding=(1, 1, 1)),
            torch.nn.BatchNorm3d(self.cube_size * 8),
            torch.nn.LeakyReLU(self.leak_value)
        )
        self.layer5 = torch.nn.Sequential(
            torch.nn.Conv3d(self.cube_size * 8, 1, kernel_size=4, stride=2, padding=(1, 1, 1)),
            torch.nn.Sigmoid()
        )

    def forward(self, x):
        out = x.view(-1, 1, self.cube_size, self.cube_size, self.cube_size)  # torch.Size([BATCH_SIZE, 1, 32, 32, 32])
        out = self.layer1(out)                                               # torch.Size([BATCH_SIZE, 32, 16, 16, 16])
        out = self.layer2(out)                                               # torch.Size([BATCH_SIZE, 64, 8, 8, 8])
        out = self.layer3(out)                                               # torch.Size([BATCH_SIZE, 128, 4, 4, 4])
        out = self.layer4(out)                                               # torch.Size([BATCH_SIZE, 256, 2, 2, 2])
        out = self.layer5(out)                                               # torch.Size([BATCH_SIZE, 1, 1, 1, 1])
        return out


class GAN:
    def __init__(self, z_size=Z_SIZE):
        self.discriminator = Discriminator()
        self.generator = Generator()
        self.z_size = z_size

        if os.path.exists(SAVED_DISCRIMINATOR) and os.path.exists(SAVED_GENERATOR):
            self.discriminator = torch.load(SAVED_DISCRIMINATOR)
            self.generator = torch.load(SAVED_GENERATOR)

    def train(self, data):
        betas = (0.5, 0.5)
        discriminator_optimizer = torch.optim.Adam(self.discriminator.parameters(), lr=DISCRIMINATOR_LEARNING_RATE, betas=betas)
        generator_optimizer = torch.optim.Adam(self.generator.parameters(), lr=GENERATOR_LEARNING_RATE, betas=betas)
        criterion = torch.nn.BCELoss()

        if not os.path.exists(LOG_PATH):
            os.mkdir(LOG_PATH)
        writer = SummaryWriter(LOG_PATH)

        for epoch in range(EPOCHS):
            for i, X in enumerate(tqdm(data)):
                current_batch_size = X.size()[0]
                if current_batch_size != BATCH_SIZE:
                    # drop last batch due to incompatible size
                    continue

                # ================== train discriminator ==================
                noise = torch.randn(current_batch_size, self.z_size)
                real_labels = torch.ones(current_batch_size)
                fake_labels = torch.zeros(current_batch_size)

                discriminator_real_prediction = self.discriminator(X)
                discriminator_real_loss = criterion(discriminator_real_prediction, real_labels)

                fake_models = self.generator(noise)
                discriminator_fake_prediction = self.discriminator(fake_models)
                discriminator_fake_loss = criterion(discriminator_fake_prediction, fake_labels)

                discriminator_loss = discriminator_real_loss + discriminator_fake_loss

                discriminator_accuracy_real = torch.ge(discriminator_real_prediction.squeeze(), 0.5).float()
                discriminator_accuracy_fake = torch.le(discriminator_fake_prediction.squeeze(), 0.5).float()
                discriminator_accuracy = torch.mean(torch.cat((discriminator_accuracy_real, discriminator_accuracy_fake), 0))

                if discriminator_accuracy <= DISCRIMINATOR_THRESHOLD:
                    self.discriminator.zero_grad()
                    discriminator_loss.backward()
                    discriminator_optimizer.step()

                # ================== train generator ==================
                noise = torch.randn(current_batch_size, self.z_size)

                fake_models = self.generator(noise)
                discriminator_fake_prediction = self.discriminator(fake_models)
                generator_loss = criterion(discriminator_fake_prediction, real_labels)

                self.discriminator.zero_grad()
                self.generator.zero_grad()
                generator_loss.backward()
                generator_optimizer.step()

            # # ================== print and log progress ==================
            writer.add_scalar('Loss Discriminator', discriminator_loss.item(), epoch)
            writer.add_scalar('Loss Generator', generator_loss.item(), epoch)
            writer.add_scalar('Accuracy Discriminator', discriminator_accuracy, epoch)

            if (epoch % 10) == 0:
                noise = torch.randn(current_batch_size, self.z_size)
                generated = self.generator(noise)
                torch.save(generated, GENERATED_PATH + 'example_after_{:04d}_iterations.pt'.format(epoch))

            print('Iter: {0:5d}, D_loss: {1:.4}, G_loss: {2:.4}, D_acu: {3:.4}'.format(epoch, discriminator_loss.item(), generator_loss.item(), discriminator_accuracy))

        torch.save(self.discriminator, SAVED_DISCRIMINATOR)
        torch.save(self.generator, SAVED_GENERATOR)


if __name__ == '__main__':
    if torch.cuda.is_available():
        torch.set_default_tensor_type('torch.cuda.FloatTensor')

    dataset = VoxelData(DATASET_HDF, CLASSES)
    data_loader = torch.utils.data.DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=True)

    gan = GAN()
    gan.train(data_loader)