Giters

tjuskyzhang / Scaled-YOLOv4-TensorRT

Got 100fps on TX2. Got 500fps on GeForce GTX 1660 Ti. If the project is useful to you, please Star it.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

No detection visible

cosama opened this issue · comments

commented

Thanks for working on moving yolov4-tiny to tensorrt.

I tried the repo, it all runs but there are no bounding boxes displayed on the resulting images, it looks like it doesn't detect anything.

The only difference to the setup you mentioned in your readme is that I used the official weights from https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.weights as I don't have access to baidu, when creating the wts file. Also https://github.com/WongKinYiu/PyTorch_YOLOv4 seems to not work with the gen_wts.py file anymore, I had to modify it slightly by manually importing darknet.

Any idea what is going on here? Have you had some successful detection in the past?

commented

Thanks for working on moving yolov4-tiny to tensorrt.

I tried the repo, it all runs but there are no bounding boxes displayed on the resulting images, it looks like it doesn't detect anything.

The only difference to the setup you mentioned in your readme is that I used the official weights from https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.weights as I don't have access to baidu, when creating the wts file. Also https://github.com/WongKinYiu/PyTorch_YOLOv4 seems to not work with the gen_wts.py file anymore, I had to modify it slightly by manually importing darknet.

Any idea what is going on here? Have you had some successful detection in the past?

You can directly download the weights from https://github.com/WongKinYiu/PyTorch_YOLOv4/tree/master and I think you will get the right detection results.

There is a little different between the AlexeyAB's version and WongKinYiu's version. At the route layer, AlexeyAB uses "groups=2 group_id=1" , and WongKinYiu uses "route_lhalf". The dimensions of the tensors are same, so "it all runs". But the weights are different, so "it doesn't detect anything".

commented

Thanks so much for pointing this out. I tried converting the pt file and it works indeed now. It runs at 30ms/frame on a TX2 pretty impressive.

Also thanks for the explanation I clearly was missing something there.

commented

Thanks so much for pointing this out. I tried converting the pt file and it works indeed now. It runs at 30ms/frame on a TX2 pretty impressive.

Also thanks for the explanation I clearly was missing something there.

I also test on TX2. When I using the MAX-N mode and jetson_clocks, it can run about 10~11ms/frame.

commented

Thanks so much for pointing this out. I tried converting the pt file and it works indeed now. It runs at 30ms/frame on a TX2 pretty impressive.

Also thanks for the explanation I clearly was missing something there.

@cosama can you give some more details over this i take latest pull from PyTorch_YOLOv4 repo, also downloaded same pytorch weights provided by the author but still facing similar issue as here.

commented

@swapgit Sorry I have never had an issue like that, also I don't really have a setup anymore to work on this sorry.

Given that the original issue was resolved I will close this.