Yolov4 Tao freeze affect sertain classes with augmentation or another way to transfer learning only for several calsses

[chijco]

Hey we train YOLOv4 in TAO3 the question is like this:
I have already trained a model with 6 classes where first 3 have medium AP and another 3 have a great AP.
1)I noticed that setting augmentation hue =0 revrses the results if I train from the beginning with hue=0 then a get relatively same mAP, but this time first 3 classes are ok, while second 3 classes get poor.

2)Is there a way to make transfer lerning based on original model where ( first 3 have medium AP and another 3 have a great AP) but setting this hue=0 to affect somhow only the needed classes?

3. Maybe there is another way to kinda freeze transfer learning not for layers, but for classes that I want to save their carrent AP - and not affecting those classes during training?

Thank you very much

[Morganh]

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

@chijco

make transfer lerning based on original model

Can you set the existing model(first 3 have medium AP and another 3 have a great AP) as the pretrained model and then set hue=0 to train the needed classes?
More, please use latest TAO docker(TAO 4.0.1) since you are using an old version(TAO3).

You can also refer to the " Class Weighting Config" section in YOLOv4 - NVIDIA Docs