Mapping classes for fine-tune

I want to find out how class mapping works for a fine tuning task.
Denote that we have the custom dataset with class such as “motorcycle”.
We also have weight of TrafficCamNet, that was trained for class of “two-wheelers”.

How should I write target_class_mapping to the specs file to train the model on a custom dataset using knowledge of the pre-trained TrafficCamNet?

What option will lead me to the goal?

target_class_mapping {
      key: "motorcycle"
      value: "two-wheelers"
  }

OR

target_class_mapping {
      key: "two-wheelers"
      value: "motorcycle"
  }

OR

target_class_mapping {
      key: "motorcycle"
      value: "motorcycle"
  }

Please use above. Because if your labels do not contain “two-wheelers”, you can not set it in the training spec.

Ok, so this class will be as good as “two-wheelers”?

No, we cannot draw conclusion since you are training with new own dataset for a new class.

And if I redefine all classes in a custom dataset as “two-wheelers”?

For this case, you can firstly run "tao detectnet_v2 evaluation xxx " or “tao detectnet_v2 inference xxx” to get the baseline.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.