A codespace for training diffusion models with variational continual learning, to see if it can mitigate catastrophic forgetting.
Standard training. |
VCL training. |
| Images generated during continual learning of MNIST digits, with no data replay. Each row i shows generations from when the model had been trained on the first i tasks (digits), each column j shows generations for digit j. | |
Note that the variational inference for VCL is very expensive, so this code will take several hours to run on a powerful GPU.
This diffusion model in this project is based on original work by Tim Pearce, released under the MIT License in 2022. The original code can be found at Conditional Diffusion MNIST. The modifications made by me in 2024 are also released under the MIT License.