#3 · open · 1 comments
Hi, thank you for making an amazing notebook that I can practice. I would like to change the size of the image. For example, in R1GAN and BEGAN, `IMAGE_DIM` is (64, 64, 3). However, I would like to change `IMAGE_DIM` as (512,512,3) . I want my result of GAN also to become (512,512,3). Can you give me the instruction about how am I supposed to change the network? Thank you
Hi gunahn, I didn't develop this code but I work with GAN development as well. From what I understand in the code you need to change the following parts: --- ## Generator: ```python class Generator(nn.Module): """ Convolutional Generator """ def __init__(self, out_channel=3, n_filters=128, n_noise=512): super(Generator, self).__init__() self.fc = nn.Linear(n_noise, 1024*4*4) self.G = nn.Sequential( ResidualBlock(1024, 512), nn.Upsample(scale_factor=2, mode='bilinear'), # (N, 512, 8, 8) ResidualBlock(512, 256), nn.Upsample(scale_factor=2, mode='bilinear'), # (N, 256, 16, 16) ResidualBlock(256, 128), nn.Upsample(scale_factor=2, mode='bilinear'), # (N, 128, 32, 32) ResidualBlock(128, 64), nn.Upsample(scale_factor=2, mode='bilinear'), # (N, 64, 64, 64) ResidualBlock(64, 64), nn.Upsample(scale_factor=2, mode='bilinear'), # (N, 64, 128, 128) ResidualBlock(64, 32), nn.Upsample(scale_factor=2, mode='bilinear'), # (N, 32, 256, 256) ResidualBlock(32, 16), nn.Upsample(scale_factor=2, mode='bilinear'), # (N, 16, 512, 512) ResidualBlock(16, 16), nn.Conv2d(16, out_channel, 3, padding=1) # (N, 3, 512, 512) ) def forward(self, z): B = z.size(0) h = self.fc(z) h = h.view(B, 1024, 4, 4) x = self.G(h) return x ``` --- ## Discriminator: ```python class Discriminator(nn.Module): """ Convolutional Discriminator """ def __init__(self, in_channel=3): super(Discriminator, self).__init__() self.D = nn.Sequential( nn.Conv2d(in_channel, 32, 3, padding=1), # (N, 32, 512, 512) ResidualBlock(32, 64), nn.AvgPool2d(3, 2, padding=1), # (N, 64, 256, 256) ResidualBlock(64, 128), nn.AvgPool2d(3, 2, padding=1), # (N, 128, 128, 128) ResidualBlock(128, 256), nn.AvgPool2d(3, 2, padding=1), # (N, 256, 64, 64) ResidualBlock(256, 512), nn.AvgPool2d(3, 2, padding=1), # (N, 512, 32, 32) ResidualBlock(512, 1024), nn.AvgPool2d(3, 2, padding=1), # (N, 1024, 16, 16) ResidualBlock(1024, 2048), nn.AvgPool2d(3, 2, padding=1) # (N, 2048, 8, 8) ) self.fc = nn.Linear(2048*8*8, 1) # (N, 1) def forward(self, x): B = x.size(0) h = self.D(x) h = h.view(B, -1) y = self.fc(h) return y ``` --- ## Other adaptations: ```python IMAGE_DIM = (512, 512, 3) def tensor2img(tensor): img = (np.transpose(tensor.detach().cpu().numpy(), [1,2,0])+1)/2. return img def get_sample_image(G, n_noise=512, n_samples=64): """ save sample 100 images """ n_rows = int(np.sqrt(n_samples)) z = (torch.rand(size=[n_samples, n_noise])*2-1).to(DEVICE) # U[-1, 1] x_fake = G(z) x_fake = torch.cat([torch.cat([x_fake[n_rows*j+i] for i in range(n_rows)], dim=2) for j in range(n_rows)], dim=1) result = tensor2img(x_fake) return result ``` --- I haven't tested it, but if you have any errors, share here and I can try to help.