TLT custom train DETECTNETV2 resnet 10 NOT giving good Map as expected than given through ngc model what should i do

Please provide the following information when requesting support.

• Hardware (T4/V100/Xavier/Nano/etc)
• Network Type (Detectnet_v2 Resnet 10 )
• Training spec file
detectnet_v2_train_resnet18_kitti_custom.txt (17.3 KB)
i am getting 10 percent map for car , 1 percent for auto and 2 percent for motorcycle , and for other classes like truck , lcv etc i am getting 0 percentage map what should i do

Please use a larger backbone. Also, please try to train with less classes, for example, car and motorcycle.

Should i use Detectnet v2 with Resnet 18 or should use yolov4 tiny since i need to use all classes mentioned above in object detection and deploy my application to the jetson nano, which one of them would be more efficient and i cant compromise with inference speed since i have to run it realtime for one camera with deepstream.

For your case, you can use YOLOv4_tiny network to train.

During training of my pruned yolov4 tiny model my training stops automatically, with fails message kindly check the training logs.txt (34.9 KB)
and my retraining spec file from here
yolo_v4_tiny_retrain_kitti.txt (2.4 KB)
what should i need to do.

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

I am afraid it is due to OOM. You can set a lower batch-size and retry.

1 Like

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