Hi!
I tried to train NVIDIA TAO model on my two class object detection model. Here’s a link to the project: mouse vs cup - Dashboard - Edge Impulse
The accuracy after fine tuning any model is pretty low. The best accuracy I could achieve is 49%(precision score mentioned on the training page after training). Am I doing something wrong? Is the model not pre-trained on any dataset? How can I increase the accuracy here?
Thanks
Hi @karkapur,
I chatted with our ML team, and it seems there is an issue with pixel scaling that should be fixed soon.
Here are the results on the TAO YOLOv4 for the mouse vs cup(438 samples in total). Are they not enough?
Hi @karkapur,
With such a low validation score, it probably means you need to train longer. Object detection models can be notorious on how many training cycles they require. Maybe try 500 epochs and see if that helps.
I tried training it for 400 epochs before but I noticed that the accuracy dropped even more. This shouldn’t happen because the platform chooses the best weights based on the best validation set accuracy, right?
Here’s validation score for 500 epochs. Same settings as before:
Perhaps a tutorial would be useful…
Hi @karkapur,
One of our engineers used EON Tuner with your dataset and was able to get 70%+ precision. I recommend trying that to find an architecture that works for you.
For example, here’s MobileNet v2:
I’ll try the EON tuner thanks. I tried to retrain the model with the settings you provided and the accuracy is not what I expected:
I’m training a TAO YOLOv4 btw.
@shawn_edgeimpulse requesting an update. Please let me know if you have any new direction for me to explore.
Hi @karkapur,
Could you try YOLOv3 instead of YOLOv4? As your dataset is not very big, it should work better with v3. The engineer I chatted with used YOLOv3 with a MobileNetV2 backbone to get ~72% accuracy.
