ResNet

Observations

We introduced UNet, with learnable downsamples, and heavily reduced our augmentation pipeline. We pruned for hard data samples for training/validation PSNRs to be more representative of each other’s difficulty.

We trained for 100 epochs, but at the current architecture and pipeline, we overfit and begin to regress at epoch 47.

Possible Improvements

We could re-introduce augmentation since we now have a point of reference of overfitting.