FactorVAE

FactorVAE implementation using Pytorch, Pytorch Lightning, Pipenv and Hydra. Currently, this implementation supports only the DSprites dataset. This implementation follows as much as possible the specifications contained in Disentangling by Factorising (Kim & Mnih, 2018) https://arxiv.org/pdf/1802.05983.pdf

Install

1. Clone the repository

git clone https://github.com/Michedev/FactorVAE.git

2. Install anaconda

3. Run a terminal with the anaconda environment

## Train

To train the model, run the following command:

```bash
anaconda-project run python train.py

or alternatively, to train single GPU:

anaconda-project run train-gpu

Project structure

├── data  # Data folder
├── deep_learning_template  # source code
│   ├── config
│   │   ├── dataset  # Dataset config
│   │   ├── model  # Model config
│   │   ├── model_dataset  # model and dataset specific config
│   │   ├── test.yaml   # testing configuration
│   │   └── train.yaml  # training configuration
│   ├── dataset  # Dataset definition
│   ├── model  # Model definition
│   ├── utils
│   │   ├── experiment_tools.py # Iterate over experiments
│   │   └── paths.py  # common paths
│   ├── train.py  # Entrypoint point for training
│   └── test.py  # Entrypoint point for testing
├── pyproject.toml  # Project configuration
├── saved_models  # where models are saved
└── readme.md  # This file

Generate batch of images

Once trained a model, generate the image through the script generate.py via the following command

anaconda-project run python generate.py checkpoint_path=saved_models/{model_folder}

Then, inside {model_folder} there will be the file generated.png containing the batch of generated images.

TODO

• Add disentanglement evaluation
• Add generation procedure