Diffusion Model for generating monochrome digits (from 0-9), trained on MNIST.
Final Loss (goal of < 0.1): 0.069
After extensive hyperparameter optimization (using both RayTune and custom modules), the best (lowest-loss) model is a batch size of 16, a learning rate of 5e-4 and 10 as the number of epochs.
The following is an example of the input, and the model's prediction:
An example of generated digits are the following:
As can be seen, viable digits from 0-9 are shown. The model performs the best for ones (1), eights (8) and nines (9).
This work was done with the help of the HuggingFace Diffusion Models course.