roserustowicz
commented
5 years ago - Allows planet input to stay at it's original resolution
- Incorporates gridsize x gridsize inputs at a later stage in the unet
- Uses averages over all timestamps in the decoder
- If larger planet images are used in the input, features are pooled to the label size and concatenated during decoding
- Add gradient clipping to suppress NaNs during training