PyTorch code for Geometric Inductive Biases for Identifiable Unsupervised Learning of Disentangled Representations.

Experiments

Hardware & Software Dependency

• Hardware

  All experiments are done using a single NVIDIA RTX-2080Ti GPU.

• Software

  We use Ubuntu 18.04 platform with Python 3.7 and CUDA 10.1.

  For all dependent libraries, please refer to requirements.txt.

Installation

1. Clone this repository.

   git clone https://github.com/PanZiqiAI/gdrae.git
   

2. Create python environment and install dependent libraries via pip.

   yes | conda create --prefix=./venv python=3.7
   conda activate ./venv
   pip install -r requirements.txt
   

Training

Preparing Datasets

All image datasets should be in directory Datasets. Due to copyright reasons, we do not provide datasets here, please download the dataset files yourself and prepare datasets as the following instructions.

• 2DShapes

  Put the file dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz in directory Datasets/dsprites.

• 3DShapes

  Put the file 3dshapes.h5 in directory Datasets/3dshapes/raw.

• 3DFaces

  Put the file basel_face_renders.pth in directory Datasets/3dfaces.

• 3DCars

  Put the file car_001_mesh.mat - car_199_mesh.mat in directory Datasets/cars3d/raw.

• SmallNorb

  Put the file smallnorb-5x01235x9x18x6x2x96x96-testing-cat.mat, smallnorb-5x01235x9x18x6x2x96x96-testing-dat.mat, smallnorb-5x01235x9x18x6x2x96x96-testing-info.mat, smallnorb-5x46789x9x18x6x2x96x96-training-cat.mat, smallnorb-5x46789x9x18x6x2x96x96-training-dat.mat, smallnorb-5x46789x9x18x6x2x96x96-training-info.mat in directory Datasets/small_norb/raw.

Running Experiments

• Train on Sector2D Synthetic Manifold

  cd gdrae-sector-manifold
  python script_train.py --dataset=sector2dunif \
      --lambda_recon=1000.0 --lambda_kld=1000.0 --lambda_jacob_sn=1000.0 --lambda_jacob_sv=1.0 --lambda_capacity=10.0 \
      --desc="sector2dunif/gdrae" 
  

• Train on Image Datasets

  The following ${DATASET} argument can be shapes2d, shapes3d, faces, cars3d, small_norb.

  cd gdrae-images
  python script_train.py --dataset="${DATASET}" --sigma=0.01 \
      --lambda_jacob_sv=0.01 --lambda_norm_logcap=10.0 --lambda_norm_sv=0.01 \
      --desc="${DATASET}/gdrae"
  

By running the above commands, a directory STORAGE/experiments will be created, which stores training logs, evaluations, visualizations, and saved checkpoints. Another directory STORAGE/tensorboard will also be created, which stores tensorboardX visualization of training logs.

Results

Sector2D Synthetic Manifold

Image Datasets