Does TAO model really pretrained?

Please provide the following information when requesting support.

• Hardware (T4/V100/Xavier/Nano/etc) : RTX 3060
• Network Type (Detectnet_v2/Faster_rcnn/Yolo_v4/LPRnet/Mask_rcnn/Classification/etc) : Detectnet_v2
• TLT Version (Please run “tlt info --verbose” and share “docker_tag” here): 3.22.05

I would like to know whether TAO models are pretrained or not.
Because TAO document said pretrained models with custom data can produce highly accurate model.
I tried detectnet v2 with 2000 images, which produce 71.82% accuracy. (2 classes: helmet: 82.64%, person: 61.00%)
I’m not sure this result benefits from pretrained model or not.
Another issue is, I use 8000 images to train detectnet v2, but I got worse result: 60.25%(helmet: 72.11%, person: 48.40%)

I would be very grateful if you could give me some hints or feedbacks.

For detectnet_v2 network, if train for a long time, the final mAP’s difference is not obvious between with pretrained model and without pretrained model.

But for detectnet_v2 network, with pretrained model, the mAP will getting a higher mAP more quickly . That means it will

  • Achieve higher accuracy with less data
  • Lower Training cost.

For other network, for example yolov4 network, if train with pretrained model, the final mAP different will obviously higher than the one without pretrained model.

For 8000 images, please check the data distribution versus 2000 images. If they are different, it is normal to get a different mAP result. Also, please make sure use the same validation dataset.

Thank you ver much!

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