mhr / dgp-vae

Disentangled GP-VAE

DGP-VAE: Disentangled GP-VAE

Tensorflow implementation of the DGP-VAE model. Accompanying code for our paper.

Overview

We investigate the performance of a GP-VAE type model to learning disentangled representations from time series by providing a slightly modified model: the DGP-VAE.

Our model learns disentangled representations from sequential data by modeling each independent latent channel with a Gaussian Process and employing a structured variational distribution that can capture long-term dependencies in time.

We show the efficacy of our approach in two experiments:

1. A benchmark experiment comparing against state-of-the-art disentanglement models on synthetic data.

2. An experiment on real medical time series data, where we provide a detailed comparison with the state-of-the-art model for disentanglement that can exploit the structure of sequential data.

Dependencies

• Python >= 3.6
• TensorFlow = 1.15
• disentanglement_lib (Available here)
• Some more packages: see requirements.txt

Run

1. Clone or download this repo. cd yourself to its root directory.
2. Grab or build a working python enviromnent. Anaconda works fine.
3. Install dependencies, using pip install -r requirements.txt
4. Clone or download the disentanglement_lib and add its root directory to your PYTHONPATH.
5. Download data:
   1. Synthetic data: First, download the underlying factors of variation by running data/load_{dsprites, smallnorb, cars3d, shapes3d}.sh. Then synthesize the actual data set from these factors by running python data/create_dataset.py --data_type {dsprites, smallnorb, cars3d, shapes3d}.
   2. Real medical data: Download the HiRID data set and save the Parquet partitons of the merged stage of the pre-processed data in a directory. To create the training data from the partitions, run python create_hirid.py --hirid_merged_dir /path/to/hirid/partitions. Finally, to download the assignment matrix required for the DCI score calculation, run data/load_hirid_assign.sh.
6. Run the command python run_experiment.py --model_type dgp-vae --data_type {dsprites, smallnorb, cars3d, shapes3d, hirid} --exp_name <your_name> ...

To see all available flags run: python train.py --help

Reproducibility

The exact hyperparameters used for the experiments are reported in our paper. To reproduce our final results run the following commands:

• dSprites, SmallNORB, Cars3D, Shapes3D: python run_experiment.py --model_type dgp-vae --data_type {dsprites, smallnorb, cars3d, shapes3d} --time_len 5 --testing --batch_size 32 --exp_name reproduce_{dsprites, smallnorb, cars3d, shapes3d} --seed $RANDOM --banded_covar --latent_dim 64 --encoder_sizes=32,256,256 --decoder_sizes=256,256,256 --window_size 3 --sigma 1 --length_scale 2 --beta 1.0 --data_type_dci {dsprites, smallnorb, cars3d, shapes3d} --shuffle --save_score --visualize_score

• HiRID: python run_experiment.py --model_type dgp-vae --data_type hirid --time_len 25 --testing --batch_size 64 --exp_name reproduce_hirid --seed $RANDOM --banded_covar --latent_dim 8 --len_init scaled --kernel_scales 4 --encoder_sizes=128,128 --print_interval 1 --decoder_sizes=256,256 --window_size 12 --sigma 1.005 --length_scale 20.0 --beta 1.0 --shuffle --data_type_dci hirid --save_score --visualize_score