Assertion t >= 0 && t < n_classes in Chesapeake13

[patriksabol]

Description

When training on Chesapeake13 with num_classes=14 and ignore_index=0, an assertion error ‘t >= 0 && t < n_classes’ occurs. This indicates that there are more class indices in the Chesapeake label imagery than the specified num_classes. However, when running gdalinfo Baywide_13Class_20132014.tif Min=1.000 Max=12.000 Minimum=1.000, Maximum=12.000,

Could there be a problem with computing indices from Chesapeake13? I am using RandomBatchGeoSampler, so this error could randomly appear when there is a patch with a false class ID.”

Steps to reproduce

Training on Chesaspeake13

Version

0.4.1

[adamjstewart]

I also noticed this here: #1368 (comment)

@calebrob6 can you look into this?

[calebrob6]

Sure

[adamjstewart]

Any luck @calebrob6?

[adamjstewart]

Note that Chesapeake13 no longer exists after #2214, but the question still applies for the remaining Chesapeake state-specific products.

[calebrob6]

t >= 0 && t < n_classes means that CrossEntropy is seeing class indices in your mask that are less than 0 or greater than number of classes - 1.

To actually use the new Chesapeake state-specific products in training the user will need to re-map the masks to the range [0, num_classes - 1]. @adamjstewart -- this probably should be a new issue.