Project for Probabilistic Machine Learning

We write a paper describing image generation models from a probabilistic perspective, which can be found in report.pdf. We cover the math behind variational autoencoders and diffusion models and explain the connection between the two.

pip3 install torch torchvision matplotlib scikit-learn

To re-train the VAE, run the following: python3 main_vae.py --latent_dim 2 --save_name sample_1_latents --latents_to_sample 1 Then, add the flags --viz_latent to visualize the latent space and --case_study to visualize the reconstructions.

To re-train the DDPM, run the following: python3 main_ddpm.py --n_steps 1000 Then, add the flags --viz_trajectory to save images of the trajectory of a single sample and --case_study to visualize the reconstructions.

Once you have the trajectory images, you can create a video with the following command: ffmpeg -framerate 120 -pattern_type glob -i '*.png' -r 15 -vf scale=512:-1 ddpm_trajectory.gif;

Figure 2 in the report.

Degenerate	Miss	Success

Figure 3 in the report.

As mentioned in the code, we thank the authors of https://github.com/cloneofsimo/minDiffusion for their implementation of DDPM.