TLT 3.0 Yolov4 tiny

Is it possible to train Yolov4 tiny with TLT 3.0? I see yolov4 is supported wondered if I just needed to change the config file

Sorry, the Yolov4 tiny is not supported.

Even though the official tiny version is not supported, you can use their CSPDarknet19 backbone instead of CSPDarknet53. The end result is yolov4-tiny but with 3 yolo heads instead of 2. Still much faster than Yolov4 for inference. Plus the 3 heads means better detection of smaller objects.

1 Like

Thanks, will give it a try

I’ve attached my training and retraining (after pruning) file in case you want to see how I set it up to use with cspdarknet19.
yolo_v4_retrain_cspdarknet19_kitti.txt (2.0 KB)

yolo_v4_train_cspdarknet19_kitti.txt (2.0 KB)

Here are some results I got on Jetson Nano 4GB using Deepstream 5.1:
29.2 FPS - using yolov4/darknet19 backbone with 416x416 input
33.2 FPS - using yolov4/cspcarknet19 backbone with 416x416 input
43.6 FPS - using ssd/mobilenetv2 backbone with 300x300 input
26.1 FPS - using ssd/mobilenetv2 backbone with 416x416 input
0.75 FPS - using frcnn/darknet19 backbone with 416x416 input
0.76 FPS - using frcnn/darknet19 backbone with 300x300 input

1 Like