No warning when forgetting to set num_classes in train.py?

[alexander-soare]

Wondering if I should promote this to an issue (and also know you have timm bits in progress).

I noticed I was able to train on a 10 class dataset without setting num_classes, meaning my model was using the default num_classes. I only realised what was happening after a little while, but there were no warnings or errors. LabelSmoothingCrossEntropy uses a gather according to target labels so it doesn't care if the output has more classes than the target.

[rwightman]

@alexander-soare yes, that's a potential issue, having a warning wouldn't be a bad idea, but there is no formal dataset api to compare dataset classes to model...
So would need to getattr num_classes for the internal datasets with a default fallback... hmmm

[alexander-soare]

Makes sense. Also a consideration: LabelSmoothingCrossEntropy and its cousins should raise errors when target/output dims don't match. Unless you think there exist use cases where a user would actually want it to work that way.

[rwightman]

@alexander-soare I'm not quite sure how LabelSmoothingXEnt makes a significant difference... PyTorch XEnt will work just fine if number of dense target labels is less than your model output, with dense integer labels there is no way of knowing if you have fewer target classes than model output, any given sampled batch can have < num_classes - 1 is so how do you warn? Without baking info of num_classes into your datasets, you'd have to iterate over all samples in your dataset and say oh, you only have 10 classes but model has 1000. You can only detect if your dense labels exceed the dim of your model output, which should crash anyways....

[alexander-soare]

Ahh yeah... didn't think that through