DCGAN

预备知识

• nn.ConvTranspose2d

                                             

class torch.nn.ConvTranspose2d(in_channels, out_channels, kernel_size, stride=1, padding=0, output_padding=0, groups=1, bias=True, dilation=1)

         

DCGAN

• Generator / Discriminator

               

• Remark

1. Discriminator中使用stride=2而不是pooling来进行降维；
2. 用fully convolution 代替 fully connected layer;
3. Generator最后一个 nn.ConvTranspose2d层后的激活函数为nn.Tanh()；
4. Discriminator最后一个卷积层之后的激活函数为nn.Sigmod().

Experiment

• config

1. dataset : align_celeba [202599-images];
2. batch_size = 128, image_size = 64, init_dims  = 100, s_channels = 64;
3. num_epochs = 10, base_lr  = 2e-4, beta = 0.5;
4. optimizer : optim.Adam

• loss

                 

• visual