ivo-gilles / tsit

TSIT: A Simple and Versatile Framework for Image-to-Image Translation

based on EndlessSora/TSIT

Installation

Following [original code], or just

git clone https://github.com/trngpg/TSIT.git
cd TSIT
conda env create -f conda_env.yml

The code requires batch-norm (consult original code)

cd models/networks/
git clone https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
cp -rf Synchronized-BatchNorm-PyTorch/sync_batchnorm .
rm -rf Synchronized-BatchNorm-PyTorch
cd ../../

Tasks and Datasets

We are going to train TSIT on day2night unpaired data. The two datasets in used are BDD100K and VinAI's Scale.

Dataset Preparation

1. Get BDD100K data to your machine

   MYHOME=lustre/scratch/client/vinai/users/trungpq3/datasets
   cp /vinai-public-dataset/BDD100K/bdd100k_images_100k.zip $MYHOME
   cd $MYHOME && unzip bdd100k_images_100k.zip
   

   After unzipping, there is a folder bdd100k with the following structure:

   ./bdd100k/images/100k/{train-val-test}/{image_files}
   

2. Get Scale data

   cp /vinai-autopilot/data/lane/scale_combined $MYHOME
   

   After unzipping, the scale_combined folder has the following structure:

   ./scale_combined/train_672w/images/{Downtown-Highway-Rural-Sub-urban}/{Clear-CLoudy-Night-etc}/{image_files}
   

These data will be read by data loader in ./data/scale_day2night_dataset.py

Training

Training file is run_train_superpod.sh.

1. First, provide bdd and scale database paths
2. Run with bash run_train_superpod.sh

To train with sbatch on superpod (submit queued job), see run_sbatch_tsit.sh