| 65 | |
| 66 | |
| 67 | class _netD(nn.Module): |
| 68 | def __init__(self, ngpu, num_classes=10): |
| 69 | super(_netD, self).__init__() |
| 70 | self.ngpu = ngpu |
| 71 | |
| 72 | # Convolution 1 |
| 73 | self.conv1 = nn.Sequential( |
| 74 | nn.Conv2d(3, 16, 3, 2, 1, bias=False), |
| 75 | nn.LeakyReLU(0.2, inplace=True), |
| 76 | nn.Dropout(0.5, inplace=False), |
| 77 | ) |
| 78 | # Convolution 2 |
| 79 | self.conv2 = nn.Sequential( |
| 80 | nn.Conv2d(16, 32, 3, 1, 0, bias=False), |
| 81 | nn.BatchNorm2d(32), |
| 82 | nn.LeakyReLU(0.2, inplace=True), |
| 83 | nn.Dropout(0.5, inplace=False), |
| 84 | ) |
| 85 | # Convolution 3 |
| 86 | self.conv3 = nn.Sequential( |
| 87 | nn.Conv2d(32, 64, 3, 2, 1, bias=False), |
| 88 | nn.BatchNorm2d(64), |
| 89 | nn.LeakyReLU(0.2, inplace=True), |
| 90 | nn.Dropout(0.5, inplace=False), |
| 91 | ) |
| 92 | # Convolution 4 |
| 93 | self.conv4 = nn.Sequential( |
| 94 | nn.Conv2d(64, 128, 3, 1, 0, bias=False), |
| 95 | nn.BatchNorm2d(128), |
| 96 | nn.LeakyReLU(0.2, inplace=True), |
| 97 | nn.Dropout(0.5, inplace=False), |
| 98 | ) |
| 99 | # Convolution 5 |
| 100 | self.conv5 = nn.Sequential( |
| 101 | nn.Conv2d(128, 256, 3, 2, 1, bias=False), |
| 102 | nn.BatchNorm2d(256), |
| 103 | nn.LeakyReLU(0.2, inplace=True), |
| 104 | nn.Dropout(0.5, inplace=False), |
| 105 | ) |
| 106 | # Convolution 6 |
| 107 | self.conv6 = nn.Sequential( |
| 108 | nn.Conv2d(256, 512, 3, 1, 0, bias=False), |
| 109 | nn.BatchNorm2d(512), |
| 110 | nn.LeakyReLU(0.2, inplace=True), |
| 111 | nn.Dropout(0.5, inplace=False), |
| 112 | ) |
| 113 | # discriminator fc |
| 114 | self.fc_dis = nn.Linear(13*13*512, 1) |
| 115 | # aux-classifier fc |
| 116 | self.fc_aux = nn.Linear(13*13*512, num_classes) |
| 117 | # softmax and sigmoid |
| 118 | self.softmax = nn.Softmax() |
| 119 | self.sigmoid = nn.Sigmoid() |
| 120 | |
| 121 | def forward(self, input): |
| 122 | if isinstance(input.data, torch.cuda.FloatTensor) and self.ngpu > 1: |
| 123 | conv1 = nn.parallel.data_parallel(self.conv1, input, range(self.ngpu)) |
| 124 | conv2 = nn.parallel.data_parallel(self.conv2, conv1, range(self.ngpu)) |