Modified misleading num_classes Warnings

[panagiotamoraiti]

Description

I modified the misleading num_classes mismatch: pretrain weights has x classes, but your model has 90 classes, that was appearing during model loading.
I modified in both files used for training and inference. I just let the warnings report the correct number of classes that the head is reinitializing.

This issue had been addressed in issue #51.

@isaacrob-roboflow @SkalskiP

[panagiotamoraiti]

I have spotted another issue in this correction. In some coco style annotations, when label ids start from 0, this change reports the correct number of classes.
If a roboflow dataset is used, where ids start from 1, the warning reports one more class. So i think the most consistent way is to follow the official annotation format and start from 1.
In this case we have to subtract 1 from num_classes.

Do you have any suggestions about this issue?
Maybe also the num_classes that the head is initializing should be modified.

Finally, i left it as it was because if label ids don't start from 0 i observed that the argument num_classes initialize to a wrong value (num_classes + 1). So now the warning works well, but label ids in coco json should start from 0.