README.md

Overview

This repository provides a script and a set of instructions to construct datasets (e.g. Male-to-Female or Glasses Removal) suitable for training of the unpaired image-to-image translation models (CycleGAN, CouncilGAN, UVCGAN, etc).

Instructions

1. Download the CelebA dataset. The following files are required:

   1. Train/Val/Test Partitions. File name list_eval_partition.txt
   2. Attributes Annotations. File name list_attr_celeba.txt
   3. Aligned and Cropped Images. Archive name img_align_celeba_png.7z

2. Unpack the CelebA image archive img_align_celeba_png.7z. For example,

7z x img_align_celeba_png.7z        # requires 7zip installed

3. Use the provided convert_celeba.py script to convert the raw CelebA dataset into the CycleGAN form.

To create a Male-to-Female dataset you can use the following command:

python3 convert_celeba.py \
    --list-attr PATH_TO_list_attr_celeba.txt \
    --list-part PATH_TO_list_eval_partition.txt \
    --attr Male \
    PATH_TO_EXTRACTED_CELEBA_IMAGES \
    OUTPUT_DIRECTORY

Or, to create a Glasses removal dataset, you can run the following command:

python3 convert_celeba.py \
    --list-attr PATH_TO_list_attr_celeba.txt \
    --list-part PATH_TO_list_eval_partition.txt \
    --attr Eyeglasses \
    PATH_TO_EXTRACTED_CELEBA_IMAGES \
    OUTPUT_DIRECTORY

Requirements

To run the convert_celeba.py script one needs to have a working python3 interpreter and the following additional packages installed:

pandas
tqdm

To unpack the original CelebA dataset one needs to have a 7zip installed.

Checksums

The checksums directory contains the reference checksums of the Male-to-Female and Glasses removal datasets. These checksums were calculated over the datasets provided by CouncilGAN (CouncilGAN file names changed to match the CelebA).

The checksums directory also contans the reference checksums of the original CelebA dataset archive.