Nils Lehmann

Computed Image statistics on torchgeo train dataset split: ``` min: array([0., 0., 0., 0.], dtype=float32) max: array([1., 1., 1., 1.], dtype=float32) mean: array([0.4101762, 0.4342503, 0.3484594, 0.5473533], dtype=float32) std: array([0.17361328, 0.14048962,...

> * [More normalization](https://github.com/Global-Policy-Lab/mosaiks-paper/blob/v1.0/code/mosaiks/featurization.py#L120) > > * [Whitening](https://github.com/Global-Policy-Lab/mosaiks-paper/blob/v1.0/code/mosaiks/featurization.py#L270) I think those normalizations are unique to the MOSAIK model they use. But [these](https://github.com/Global-Policy-Lab/mosaiks-paper/blob/14deccea3b7f67b34855870ac0807d83450225ba/code/mosaiks/solve/cnn.py#L184) are the augmentations for CNN based approach.

yeah, I want to try and reproduce results first and will update the PR here then.

@calebrob6 do the train/val/test splits that come with the torchgeo dataset version, correspond to any of the checkerboard style splits as seen in Figure 3 of the Mosaik paper or...

With a resnet18 baseline I get 0.95 R-Squared score for treecover (paper 0.91) when doing proper normalization. Since we cannot replicate their results directly anyway as Caleb pointed out, I...

This is how I have hacked it so far.

Also not sure to what extent other current datasets might have the same issue because it's a silent bug.

I am reopening this because I had two other people tell me they had the same thing happening with their custom datasets. Is there maybe a way we could implement...

For rioxarray support I encountered `.nc` files where the latitude and longitude coordinates were sometimes from -90 -> 90 or vice versa etc, but I think I handled that by...

In order to do `trainer.validate(model, dataloaders=datamodule.val_dataloader())` but not having to implement my own normalization scheme as a collate fn for every dataloader from a datamodule I want to use. So...