Denoising Diffusion Probabilistic Model

Objective:

Implemented a Denoising Diffusion Probabilistic Model (DDPM) for denoising MNIST data, based on the paper "Denoising Diffusion Probabilistic Models" by Jonathan Ho, Ajay Jain, and Pieter Abbeel from UC Berkeley (arXiv:2006.11239v2).

Tasks Completed:

1. Model Initialization:

   • Implemented the __init__ function in the MNISTDiffusion class.
   • Initialized the UNet model using the provided _cosine_variance_schedule function.

2. Forward Diffusion:

   • Utilized the reparameterization trick for the forward pass.
   • Implemented the forward and _forward_diffusion functions.
   • Sampled t uniformly from the timesteps for each training sample.

3. Reverse Diffusion:

   • Implemented reverse diffusion using trainable mean and variance parameters from the UNet model.
   • Developed the _reverse_diffusion function.

4. Image Sampling:

   • Implemented the sampling procedure.
   • Developed the sampling function to generate noise and run reverse diffusion for each sample.
   • Returned the entire trajectory of samples.

5. Model Training:

   • Selected hyperparameters such as learning rate, batch size, epochs, and timesteps.
   • Chose optimizer, loss function, and learning rate scheduler.
   • Displayed 36 final samples using provided code.

Additional Tasks:

• Visualized a sample diffusion trajectory demonstrating noise turning into a clean sample.
• Labeled each image with corresponding denoising step.
• Ensured code is well-documented and organized for reproducibility.

Note: Please refer to the provided code and documentation for detailed implementation and explanations.