spyrosavl / composition-for-causal-generative-networks

Composition for Causal Generative Networks

This repository is an extension the replication implementation of Counterfactual Generative Networks.

Reproducing our experiments

Please download the datasets and the weights first:

conda activate cgn-cpu

python setup/download_datasets.py

python setup/download_weights.py

Then run the following commands to reproduce the experiments on slurm cluster: If you want to run the experiments on your local machine, please run the bash code in each of the .job files.

Fake-CGN baseline

cd u2net_cropping

sbatch run_imagenet.job

Poisson blending (Pretrained)

cd deterministic_refinement

sbatch run_imagenet.job

Poisson blending (Finetuned)

cd cgn_framework

sbatch run_train_cgn_poisson.job

cd deterministic_refinement_retrain

sbatch run_imagenet.job

GP-GAN

cd gp_gan_refinement

sbatch run_imagenet.job

sbatch run_inception_score.job