search

phillipi / pix2pix

Image-to-image translation with conditional adversarial nets
https://phillipi.github.io/pix2pix/
Other
10.13k stars 1.71k forks source link

Input size 32x32 #175

Open vahid0001 opened 5 years ago

vahid0001 commented 5 years ago

Hi, I want to use pix2pix architecture with the Cifar10 dataset and the problem is the input size. I do not want to resize the images because it is computationally expensive and it takes a lot of time for each epoch. I want the architecture of the model(Generator and Discriminator) be suitable for input size, so I want to know is there any rule to modify the model for this purpose? Because I did it but the output result wasn't satisfying and like before the modifications. I was thinking that it might exist some logic behind designing the U-net and Patch-net architecture(number filters, layers, ...) that I hope you to help me understand redesigning the model that results in the same output as before for this dataset. Thank you so much.

junyanz commented 5 years ago

For the generator, you can remove the first two downsampling layers and their corresponding upsampling layers from defineG_unet_128. For the discriminator, remove one or two downsampling layers. If you are not familiar with Lua, you may want to modify our PyTorch code.