ae_vae_understanding

Work by Regis Djaha and Iñigo Urtega on understanding AEs and VAEs.

Regis Djaha , Iñigo Urteaga

This work lead to this Master's Thesis project.

Repository structure

doc

Directory for documentation (reports, papers of interest, etc).

src

Directory where our code will be place in.

Installation

1. Clone this repository.

    git clone https://github.com/RegisKonan/ae_vae_understanding.git
    cd ae_vae_understanding

2. Install package virtualenv

   pip install virtualenv

3. Create a environment python

   virtualenv [name], example :  virtualenv env

4. Activate virtualenv

macOS version :   source [name]/bin/activate, example: source env/bin/activate

windows version :  \[name]\Scripts\activate, example: \env\Scripts\activate

5. Install requirements.txt.

pip install -r requirements.txt

By using conda

1. conda create -n [name], example: conda create -n env or conda create -n env python=3.12.2

2. Activate

conda activate [name], example: conda activate env

3. Install requirements.txt.

pip install -r requirements.txt

Running our experiments

1. Make sure you are in the main ae_vae_understanding directory

cd /path-to-ae_vae_understanding

AEs Experiments and VAEs Experiments

You can choose the type of dataset and model you want to use, as well as the number $n_L$ of layers and dimensions $D_z$.

dataset_name= [MNIST, FashionMNIST, FreyFace]

model_name = [AE, VAE]

To run the experiment scripts, call them from your ae_vae_understanding directory:

1. Test1: script/AE_VAE_experiment_nL.py.

python script/AE_VAE_experiment_nL.py -dataset_name FreyFace -model_name AE   -n_L 3 7  -D_z 5

In this script, we have fixed the dimension $D_z$ and varied the number of layers.

2. Test2: script/AE_VAE_experiment_Dz.py.

python script/AE_VAE_experiment_Dz.py -dataset_name FashionMNIST -model_name VAE   -n_L 3   -D_z  2 5

In this script, we have fixed the number of layers and varied the dimensions.

VAE Prior Experiments

We used the same code as in script/AE_VAE_experiment_nL.py and script/AE_VAE_experiment_Dz.py, replacing standard_gaussian with mean_conditional_gaussian and defining gaussian_distribution as gaussian_distribution.

In this experiment, we defined means and standard deviations.

$\beta$-VAE Experiments

1. Test: script/beta_VAE_experiment.py

python script/AE_VAE_experiment.py -dataset_name MNIST -model_name VAE  -beta 0.02 1 1.5 10 -n_L 1 -D_z 5

In this script, we have fixed the value of beta and varied the number of layers and the dimension $D_z$ ​ . You don't need to specify a model, as it is already defined in the code.

Run plot metrics, latent space, original and reconstruction images, and number of parameters.

1. Test1: script/Run_code_AE_VAE.py

python script/Run_code_AE_VAE.py -dataset_name FreyFace -model_name AE  -n_L 1 -D_z 2

2. Test2: script/Run_code_beta_VAE.py

python script/Run_code_beta_VAE.py -dataset_name FreyFace -model_name VAE  -beta 0.02 1 1.5 10 -n_L 1 -D_z 5

Notebooks

Notebooks are included to show the following steps.

1. Start jupyter.

   jupyter notebook

1. Navigate to the notebooks folder.

2. First run all cells.

Questions and Issues

If you find any bugs or have any questions about this code, please contact Regis rdjaha@bcamath.org.

Acknowledgments

We thank La Caixa Junior Leader program’s LCF/BQ/PI22/11910028 award for financial support of this work.