Adding ETCI2021 datamodule and trainer

[calebrob6]

Relevant links:

• https://competitions.codalab.org/competitions/30440#results
• https://arxiv.org/abs/2107.08369
• https://github.com/sidgan/ETCI-2021-Competition-on-Flood-Detection

[adamjstewart]

Needs to be added to docs

[adamjstewart]

Why not call it "ignore_index" and make it an int that defaults to zero? That gives the most flexibility.

[calebrob6]

Because nn.CrossEntropyLoss uses -1000 as a None value so there is not an easy way to not ignore.

[adamjstewart]

I'm not sure I understand. I'm saying to rename the ignore_zeros parameter to ignore_index and have it accept an int, not to remove the parameter entirely.

[calebrob6]

I know, but say I don't want to ignore anything -- I can't pass None because nn.CrossEntropyLoss behaves differently than everything else that accepts ignore_index

[calebrob6]

@adamjstewart how do you want to test the plot methods?

[adamjstewart]

@adamjstewart how do you want to test the plot methods?

I would personally be fine with a simple test that just runs plot and makes sure it doesn't crash, but if you want to write a more detailed test you can check the figure object for all of the expected attributes. I once wrote a bunch of matplotlib auto-grading stuff for a class I was TAing.

[calebrob6]

Relevant bit for later

        x = batch["image"]
        y = batch["mask"]
        with torch.no_grad():
            # Kornia expects masks to be [B, C, H, W] format with float values while
            # torch losses like cross entropy expect masks to be in [B, H, W] format
            # with long values. We'll just assume that samples are in a format that
            # torch expects and adjust accordingly if we need to do augmentations.
            x, y = self.train_augmentations(x, y.unsqueeze(1).float())
        y = y.long().squeeze()

[calebrob6]

I removed the trainer and reset the test data as I can't be bothered to jump through all these hoops.

Note, this dataset does need augmentations as training will overfit to the train split (the test set is sampled from a different geographic location and those inputs are likely out-of-distribution to the training data).

The above comment is the relevant part to add to train_step for augmentation that explains what kornia augmentations are expecting in terms of dimensions.