Understanding the Distillation Process from Deep Generative Models to Tractable Probabilistic Circuits

Xuejie Liu, Anji Liu, Guy Van den Broeck, Yitao Liang

This is the official implementation of our ICML 2023 paper.

Running the experiments

ImageNet32

#specify the data path
data_path=""
#train teacher DGM
cd exps/LVD_for_imagenet/
python train_vqvae2_model.py -id --data_path $data_path
#get LVD augmented dataset
python get_data_for_PG.py -id --data_path $data_path
#train cluster-conditioned PCs with progressive_growing
cd ../progressive_growing/
bash pg.sh "imagenet32"
#finetune PCs
cd ../LVD_for_imagenet/
python-jl progressive_growing_top.py -id --data_path $data_path

ImageNet64

#specify the data path
data_path=""
#train teacher DGM
cd exps/LVD_for_imagenet/
python train_vqvae2_model.py -id -img 64 -p 8 --data_path $data_path
#get LVD augmented dataset
python get_data_for_PG.py -id -img 64 -p 8 --data_path $data_path
#train cluster-conditioned PCs with progressive_growing
cd ../progressive_growing/
bash pg.sh "imagenet64"
#finetune PCs
cd ../LVD_for_imagenet/
python-jl progressive_growing_top.py -id -img 64 -p 8 --data_path $data_path