trt-yolo-app's performance on TX2???

Is there anyone who has test the performance of trt-yolo-app on TX2?
For the original yolov3-tiny, I see that tx2 can only process 12 frame per second. How about it for trt-yolo-app??

Hi liuhang,
I want to know how do you test yolov3-tiny?

I have tried trt(/usr/src/tensorrt/bin/trtexec) for yolo, find that crop layer in yoloV3_tiny is not supported.
So if you want to use TRT for yolo, you have to implement crop layer by yourself.

BTW, could you let me know where can I get trt-yolo-app? I did not think it is included in TRT.

Thanks
wayne zhu

Hi Liuhang:

We have modified yolov3-tiny to let it run on TRT direct.
You can have a try on your side.

Crop6 layer:
layer {
name: “crop6”
type: “Crop”
bottom: “pool6”
bottom: “scale6”
top: “crop6”
crop_param {
#axis: 1
#offset: 0
axis: 2
offset: 1
offset: 1
}
}

Output layer:
#layer {

name: “output”

type: “Concat”

bottom: “conv13_flat”

bottom: “conv10_flat”

top: “output”

concat_param {

axis: 1

}

#}

Then you can use command to test it directly:

/usr/src/tensorrt/bin/trtexec --deploy=yolov3-tiny.prototxt --output=conv10_perm --output=conv13_perm --int8 --batch=4

This is just for testing perf.

Thanks
wayne zhu

Hi, Waynezhu:

Thanks a lot for your help.
I just use deepstream-plugin to verify yolov3-tiny:
https://github.com/vat-nvidia/deepstream-plugins

And here on my TX2, I found that trt-yolo-app will spend 88ms to finish the detection job for one image;
But I run the “jetson_clocks.sh” script, and I found that the time will down to about 20ms per image.

It is really wonderful.