Image-to-Image Translation with Conditional Adversarial Networks

Install

• install fantastic pytorch and pytorch.vision

Datasets

• Download images from author's implementation
• Suppose you downloaded the "facades" dataset in /path/to/facades

Train with facades dataset (mode: B2A)

• CUDA_VISIBLE_DEVICES=x python main_pix2pixgan.py --dataset pix2pix --dataroot /path/to/facades/train --valDataroot /path/to/facades/val --mode B2A --exp ./facades --display 5 --evalIter 500
• Resulting model is saved in ./facades directory named like net[D|G]_epoch_xx.pth

Train with edges2shoes dataset (mode: A2B)

• CUDA_VISIBLE_DEVICES=x python main_pix2pixgan.py --dataset pix2pix --dataroot /path/to/edges2shoes/train --valDataroot /path/to/edges2shoes/val --mode A2B --exp ./edges2shoes --batchSize 4 --display 5

Results

• Randomly selected input samples
• Corresponding real target samples
• Corresponding generated samples

Note

• We modified pytorch.vision.folder and transform.py as to follow the format of train images in the datasets
• Most of the parameters are the same as the paper.
• You can easily reproduce results of the paper with other dataets
• Try B2A or A2B translation as your need

Reference

• pix2pix.torch
• pix2pix-pytorch (Another pytorch implemention of the pix2pix)
• dcgan.pytorch
• FANTASTIC pytorch pytorch doc
• genhacks from soumith