- Creation of a working model which uses a Unet architecture on PyTorch with PyTorch Lightning and plot results on Tensorboard.
- The model is a simple Unet architecture, with encoder / decoder. The decoder uses upsampling and convolutional layers to avoid the checkerboard effect.
- Data augmentation is used (random flip, random crop)
- Tested different models parameters
- Initial channel is 64, tested with 32 and 128
- Added a discriminator
- Added data normalization (-1, 1)
All models except the larger model were trained for 100 epochs.
| Iteration | Train loss | Val Loss | Generated Results |
|---|---|---|---|
| Ground Truth |  |
||
| Initial Model |  |
 |
 |
| Smaller model (initial channel at 32) |  |
 |
 |
| Larger model (initial channel at 128) |  |
 |
 |
| Add Discriminator (initial channel at 64) |  |
 |
 |
| Normalize data between -1 and 1 |  |
 |
 |