bfvae

• This is forked from seqam-lab/bfvae
• These original changes can also be found on the branch original
• The main paper is here and the supplement is here

This is the project page for Bayes-Factor-VAE: Hierarchical Bayesian Deep Auto-Encoder Models forFactor Disentanglement. The work was accepted by ICCV 2019 Oral. [Paper Link].

Further instructions added by Sachin Salim

Setup instructions

GPU drivers

If you don't have the GPU drivers installed:

• Install correct driver for the GPU, see here
• After this file is installed, if you have a GPU be sure to install conda forge cudnn

Dependencies

After installing the drivers:

• Install dependencies: conda env create -f reqs.yaml
• Activate conda env: conda activate bfvae_env

Datasets

• Download/synthesize dataset by python src/setup_dataset.py --dataset <dataset> --save_dir <save_dir>
• Currently supported datasets to be automatically downloaded/synthesized are:
  1. dsprites (Downloaded by script from here)
  2. latent2_0 (Synthesized)
  3. latent2_1 (Synthesized)

Execute instructions

• If in Greatlakes HPC, load cuda module: module load cuda cudnn
• Navigate to the model folder: cd src/bfvae<id>/
• Read cmdlines.txt for specific instructions on running various scripts
• Eg: To train the model on dsprites, execute python main.py --dataset dsprites --dset_dir <dset_dir>

Analyse training

• After the training, The training loss is saved in records/
• Navigate to cd src/ and Execute python analyse_training.py --model bfvae<id> --train_filename <train_filename>
• Pass the respective file inside records/ folder as the train_filename