This repository provides the Python implementation of the paper "Leveraging Relational Information for Learning Weakly Disentangled Representations", introducing Weak Disentanglement, a new way for learning disentangled representations that relaxes some of the limitations of the previous methods. The project is written in Pytorch. The paper has been accepted at WCCI 2022.
Weakly disentangled representations are able to structure themselves in order to separate the different regions encoding specific combinations of factor of variations. Knowing the location of these regions in the latent space, it is then possible to manipulate the representation in order to obtain controlled changes in the corresponding decoded data samples.
If you want to know more about weak disentanglement, you can read the paper on ArXiv, which also contains a link to some examples of learned realations by the trained models.
The project is separated into three subfolders, corresponding to the different datasets on which the model is trained. Each subfolder contain similar files, and should allow for a complete training of the model on the specific dataset, without the need of files from other folders.
All the folder are structured as follows:
- File train.py is the main training script, implementing the training/eval loop of the model.
- File config.py contains the model's configuration settings, structured as a Python dictionary.
- File prior.py contains all the prior of the adversarial autoencoder taht have been considered during the developement of the model. Some of the prior are not actually used in the final model.
- File model.py contains the implementation of the layers of the AbsAE and ReL models.
- File loss.py contains the implementations of all the losses used during training.
- File log.py contains a collections of logging functions used to produce the final results and to check the progress of training.
- File dataset.py contains the implementation of the dataset object and all data-related utilities.
The model is implemented in Pytorch. You should be able to install all the required dependencies using the ap2.yml file that will be (soon) added to the repository.
First, you need to download the corresponding dataset. Place the raw data into the "./data" subfolder, in the same folder as the one of the scripts.
After that, you are ready to go! You can start a new training by running the train.py script.
In config.py you can freely change many model parameters, from the number of latent variable to the number of hidden layers of the many architectural components. The parameters name should be quite auto-explicatives.
For any other additional information, you can email me at andrea.valenti@phd.unipi.it.
You're free (and encouraged) to make your own contributions to this project.
Just be nice, and respecful of each other!
All source code from this project is released under the MIT license.