EfficientNet

This repo provides a clean implementation of EfficientNet.

TODO

• fix null training values
• do everything in color as float32
• samples by pixels without a mask shouldn't add to loss
• decide whether a pixel can be marked as multiple masks or not
  • if it can be, the comparison code needs to be & not ==
  • if not, just label each pixel with the appropiate index
  • and either way, the maximum number of masks should be encoded differently
• train ResUNet for pixelwise embedding and use UMAP to color the output
• Stop this from happening (?)
• Test TPU compat