How to set enable_autoweighting in training and retraing spec

[music1913]

Please provide the following information when requesting support.

• Hardware (T4/V100/Xavier/Nano/etc)
Ubuntu, x86, RTX3090
• Network Type (Detectnet_v2)
• TLT Version (Please run “tlt info --verbose” and share “docker_tag” here)
• Training spec file(If have, please share here)
• How to reproduce the issue ? (This is for errors. Please share the command line and the detailed log here.)

I have a small proprietary dataset which contains 4 classes of images, the annotations statistics when doing dataset convert (in tao jupyter note book) is:

iva.detectnet_v2.dataio.dataset_converter_lib: 
Wrote the following numbers of objects:
b'people': 2275
b'dog': 1194
b'cat': 1102
b'horse': 450

so this is a typical imbalanced dataset.
I noticed the properties of class_weight and enable_autoweighting under section cost_function_config both in train and retrain spec are looked like to help the scenario, I want to understanding how to set them for my case.

1. if set enable_autoweighting to true, then does it mean the class_weight are all meaningless?
2. does class_weight for different class should just simply reflect the linear relation of their annotation count?
   then the class_weight for my case is: 10, 20, 20, 40 respectively for the 4 classes?
3. how much improve I can expect after apply these settings?

[Morganh]

1. It is still used for computing cost.

2. See FAQ in user guide.

• Distribute the dataset class: How do I balance the weight between classes if the dataset has significantly higher samples for one class versus another? To account for imbalance, increase the class_weight for classes with fewer samples. You can also try disabling enable_autoweighting; in this case initial_weight is used to control cov/regression weighting. It is important to keep the number of samples of different classes balanced, which helps improve mAP.