Fixed bug in trainer notebook

[calebrob6]

No description provided.

[adamjstewart]

Continuing discussion from #101 with @robmarkcole and @caleborb6:

There is an error in trainers.ipynb. The csv is created with upper case column names (e.g. train_RMSE).

What file actually creates metrics.csv? I wonder when it changed from lowercase to uppercase...

This is one of the few things that isn't tested...

It should be tested on our release branch, right? If not, we should fix that.

[calebrob6]

What file actually creates metrics.csv? I wonder when it changed from lowercase to uppercase...

It is created by the CSV logger in the notebook. The CSV logger uses the name of the metrics in the trainer. We changed Cyclone to use the regression trainer and added a more comprehensive set of metrics to log at some point. This involved renaming "rmse" --> "RMSE".

It should be tested on our release branch, right? If not, we should fix that.

It is uniquely not! Because that notebook involves actually training models, which is not feasible on CPU, we skip it during tests.

[adamjstewart]

It is uniquely not! Because that notebook involves actually training models, which is not feasible on CPU, we skip it during tests.

Can we train it for a single mini-batch?

[calebrob6]

Whether the trainer works with the dataset is covered by our other tests. The problem here was the logging assumptions (and the specific problem would only show up at the end of an epoch of training). I don't know how to make a good test for this within the notebook.

[adamjstewart]

Yes, the point of the tutorial tests is not to test TorchGeo, it's to test the tutorials to ensure that they stay up-to-date.