Fashion MNIST unconditional Unet Model trained using DDPM
Model Hyperparams:
- Model size: 51,834,625 params
- 3 stages: 128, 256, 512 channels
- Linear Attention in 2nd and 3rd stages, Self Attention in Middle Stage
- Optimizer: Adam
- LR: 3e-4
- Batch Size: 64
- Grad Accumulation: 8 steps
- Total steps: 5,000
