Evaluating Custom Detection Model Performance on Jetson Nano 2GB


I managed to train a custom detection model via jetson-inference, but I wonder, is there any possible way I can get any information or statistics for the performance of the custom model? Such as, accuracy, max fps, average confidence or etc.

I noticed the “eval_ssd.py” program inside jetson-inference, but I do not have any idea what it is for and how to use it. However, I suspect it is for trained model performance evaluation.

Thank you.


Please check below for some details:

After the training jobs, you can find some loss information on the console:

2020-07-10 13:19:26 - Epoch: 0, Validation Loss: 5.6730, Validation Regression Loss 1.7096, Validation Classification Loss: 3.9634
2020-07-10 13:19:26 - Saved model models/fruit/mb1-ssd-Epoch-0-Loss-5.672993580500285.pth

Then you can get some mAP score with your validation dataset by applying the inference script like run_ssd_example.py.
And if you deploy the model with live camera (streaming data), the fps will show on the window automatically.


1 Like


Owh, I did notice those information at the console during training. Thank you for your clarification.