Datamodule for SpaceNet1

[calebrob6]

Some things to note:

• SpaceNet1 patches are (406, 438) so I zero-pad to 448 x 448
• I add one to the mask layer in preprocessing so the classes are {nodata, background, building}
• I ignore the 0 class (the value that gets added due to padding) in the train config
• I changed SpaceNet's data.py to generate 3 samples by default so that we can test train/val/test (@ashnair1)

I can confirm that this is working with train.py, here's a shot from TensorBoard

[adamjstewart]

SpaceNet1 patches are (406, 438) so I zero-pad to 448 x 448

I'm guessing U-Nets require 1) square patches, and 2) patch sizes that are multiples of 32?

[calebrob6]

I'm guessing U-Nets require 1) square patches, and 2) patch sizes that are multiples of 32?

They just need patch sizes that are multiples of 32. There are 5 downsampling operations by default (going down the left side of the "U") which reduces the spatial resolution of the input by a factor of 32. If the input isn't evenly divisible by 32, then there will be some rounding, and in the upsample part of the "U", the spatial dimensions of the features in the skip connections won't match.

[calebrob6]

This is good to go pending the question in the above comment about larger test images vs. epsilon in the percentile normalization.