Low accuracy at the beginning of transfer learning

I was exploring TAO toolkit for transfer learning with different models, including YoloV3, YoloV4, and retinanet. I used my custom dataset, and focused on detecting people and vehicles. One thing I observed is that for the first few (~10) epochs, the accuracy is always quite low whichever model I start with.
Here is an example of training with retinanet, and this is from the 5th epoch.


person AP 0.0
vehicle AP 0.02571
mAP 0.01286


Since “person” and “vehicle” are quite commonly seen categories, I do not quite understand why the accuracy is so low with a pre-trained model?

What is the pre-trained model? Could you share the name?

The example I provided here is with a pre-trained Retinanet. I also observed similar results with Yolo3 and Yolo4.

What is the name? Is it an hdf5 file?

Yes. It is resnet_18.hdf5.

It is expected.
This kind of file is trained on a subset of the Google OpenImages dataset.
The pre-trained weight for each backbone is provided on NGC. The pretrained weights can be used as a starting point . These are unpruned models with just the feature extractor weights, and may not be used without re-training in an object detection application.

That makes sense. Thanks for the clarification.

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