damian-666 / latent_distances_finsler

Identifying latent distances with Finslerian geometry

Requirements

You will need the stochman and pyro packages to run this code.

You can install all the requirements in two ways:

• make requirements, or:
• pip install -r requirements.txt and pip install -e .

Training the model

In order to train the GPLVM on the starfish data, we use wandb. You may either want to modify the code or to login to your wandb account.

• To train the model: make train

Experiments

The GPLVM models for the starfish, qPCR and font data have been saved. The figures can be obtained with:

• make figure2. The indicatrices will be plotted in the background and along one geodesic.
• make figure3. The heatmaps will be computed, this code is time consuming.
• make figure4. Illustration of theoretical results are plotted.
• make figure5. The latent spaces of the synthetic data are plotted.
• make figure6. The latent spaces of the qPCR data is plotted.
• make figure7. The latent spaces of the MNIST and fashion MNIST are plotted.

Note that the figure for the fontdata might not be exactly similar to the one from the paper, but it doesn't change the conclusion and the main findings.