add: yolov4-csp tensorrt implementation

[makaveli10]

• updated readme with tensorrt implementation

[casper-hansen]

@makaveli10 it looks like you only updated the readme in your GitHub repository, is this correct? Do you have an actual tensorrt implementation?

[makaveli10]

Yeah so I have added the link to the tensorrt implementation scaled-yolov4-tensorrt in the read me.

[casper-hansen]

Yeah so I have added the link to the tensorrt implementation scaled-yolov4-tensorrt in the read me.

Ok, I just checked it out. It looks good!

A couple of questions:

1. Since the implementation is not the same with the YOLO layers, have you checked the accuracy to see if there are differences?
2. What are the reported FPS in your implementation? I'm currently running 8 FPS with YOLO-tiny 3L on a Jetson Nano just using the regular PyTorch version.
3. Can you make your repository extensible to other YOLO versions like the tiny one?

[makaveli10]

Yeah so I have added the link to the tensorrt implementation scaled-yolov4-tensorrt in the read me.

Ok, I just checked it out. It looks good!

A couple of questions:

1. Since the implementation is not the same with the YOLO layers, have you checked the accuracy to see if there are differences?
2. What are the reported FPS in your implementation? I'm currently running 8 FPS with YOLO-tiny 3L on a Jetson Nano just using the regular PyTorch version.
3. Can you make your repository extensible to other YOLO versions like the tiny one?
   @casperbh96

1. What do you mean its not the same with YOLO layers? I think you misunderstood I mentioned there is a slight difference in cfg i.e. change in a couple of parameter values of the yolo layer. Yes i have checked the accuracy there are no differences. If you check the readme of the scaled-yolov4-tensorrt repo you'll find the results down there.
2. I have not tested on a video yet, but i'll be updating FPS that soon.
3. Yes thats the plan to add yolov4-tiny because most of us would want to use smaller models .

[casper-hansen]

In your repo it says "Note: There is a slight difference in yolov4-csp.cfg(yolo layer mainly) for darknet and pytorch. Use the one given for pytorch in the above repo."

Otherwise it is looking good! I'm currently looking to test out the tiny version to report FPS using TensorRT.

[makaveli10]

In your repo it says "Note: There is a slight difference in yolov4-csp.cfg(yolo layer mainly) for darknet and pytorch. Use the one given for pytorch in the above repo."

Otherwise it is looking good! I'm currently looking to test out the tiny version to report FPS using TensorRT.

Just some parameters in the yolo layer.