More classes in config than trained on (tensorflow object detection API)

Question

I trained a faster_rcnn_inception network continuing from tensorflow's OD API provided Checkpoint on my dataset containing only one class.

I changed the config file to my needs and did a successful training. Testing my trained network gives me good results in accuracy.

But now i is realized that in the config file i forgot to switch num_classes from 90 to 1.

What effects has this on my network? Is it now slower than it would be if the config was correctly set to 1?

Does tensorflow's API automatically assign number/sizes of filters according to the provided num_classes?

Answer 1

I'll answer my own question:

I did a retraining with the correct num_classes set to 1 and there are only small noticeable differences:

• The checkpoint is now 3MB smaller (103 instead of 106MB).

• mAP has a smoother line for num_classes set wrongly to 90

  

But still i don't know what of a difference it makes inside tensorflow, but it does not really seem to matter.