yanlai00 / continual-ddpm

Continual Learning for Diffusion Models

Continual Learning for Diffusion Models

This is a PyTorch implementation of different continual learning strategies for class-incremental image generation with Denoising Diffusion Probabilistic Models (DDPM). Part of the repository is adopted from the labml.ai DDPM implementation and the official SupSup code repository.

Requirements

python=3.9  
pytorch=1.10  
torchvision=0.11  
tqdm
scipy

Running Continual Learning Experiments

Train DDPM Models

The experiment config files can be found in the experiments/ folder. Here are some examples:

Unconditional image generation on MNIST with knowledge distillation:

python -m experiments.uncond.mnist.kd

Conditional image generation on MNIST with sequential fine-tuning:

python -m experiments.cond.mnist.sft

Unconditional image generation on CIFAR with joint training:

python -m experiments.uncond.cifar.joint

Generate Images from Checkpoint

Change the save_dir variable and the checkpoint loading path in metrics/save_images.py. Then run

python -m metrics.save_images

Compute the FID Score

python -m metrics.save_images