number of classes confusion afer onnx export

[ksokolov-vaisto]

There have been discussions about how num_classes inferred from dataset is increased by 1 internally when creating the model head. However I don't think this is actually happening in my case.

I have a custom dataset with 33 categories, category_ids ranging 1-33. When launching training, it reports a Namespace(num_classes=33.... in the log. Assuming the increment, I would expect the model to have a total num_classes of 34. After training, I run following code for onnx export

from rfdetr import RFDETRMedium

model = RFDETRMedium(num_classes=33, pretrain_weights='training_output/checkpoint.pth')

model.export()

the export happens with following log

Using a different number of positional encodings than DINOv2, which means we're not loading DINOv2 backbone weights. This is not a problem if finetuning a pretrained RF-DETR model.
Using patch size 16 instead of 14, which means we're not loading DINOv2 backbone weights. This is not a problem if finetuning a pretrained RF-DETR model.
Loading pretrain weights
num_classes mismatch: pretrain weights has 32 classes, but your model has 33 classes
reinitializing detection head with 32 classes

I have read that the reported num_classes in checkpoint is decreased by 1, which means that the checkpoint actually has 33 classes. The resulting onnx model also has shape of labels=[300, 33]. Why not 34?

When running inference, the predicted labels seem to be 1-based (so that class_id 1 is actually corresponding to the category_id=1 from my annotations). To me this implies we have

1. a dummy value at position 0 of the output tensor and
2. no predictions ever for category_id=33 which would be at position 33, but in a 0-based tensor of length 33 it is out of index (0-32).

[isaacrob-roboflow]

Cc @probicheaux there may be an additional model head bug so if you're fixing the other one it might be nice to address this one too. Not sure what's going on here