Add a python demo

[ioir123ju]

[thancaocuong]

@ioir123ju could you update Readme.md to make your work more clear?

[ioir123ju]

@thancaocuong I have updated readme.md.

[thancaocuong]

@ioir123ju currently is this correct if I say that we only can run with batchsize 1 with python binding?

[ioir123ju]

@thancaocuong Yes, I did not complete the batch processing,but It's not difficult.You can try.

[charlielito]

@ioir123ju did you get same speed with this python wrapper? I know python can add overhead, I don't know how much this could be. With python I'm getting 36 FPS with yolov4 FP16 and input size of 416x416 on a XAvier AGX, whilst with the ./demo implementation I get 41 FPS. Did you get similar downgrade of performance?

[ioir123ju]

@ ioir123ju使用此python包装器获得了相同的速度吗？我知道python可能会增加开销，但我不知道这可能会增加多少。使用python，我在yolov4 FP16上获得36 FPS，在XAvier AGX上输入大小为416x416，而在./demo实现上，我获得41 FPS。您是否得到了类似的性能降级？

At almost the same speed

[bobbilichandu]

@ioir123ju can you please add yolov4-csp plugins too

[eric-eric-eric]

@ioir123ju am I right in thinking that to avoid memory leaks, a similar function to https://github.com/AlexeyAB/darknet/blob/master/src/network.c#L971 would be needed?

[yotamraz]

Hi! can you please add support for tensorrt8?

[perseusdg]

there is support for tensorrt8 ,checkout the tensorrt8 branch https://github.com/ceccocats/tkDNN/tree/tensorrt8..if you are asking for python support for tensorrt8 in tkDNNi am not sure about it

[yotamraz]

@perseusdg Thank you, I have already checked out to the 'tensorrt8' branch and now facing some compatability issues with the python binding. I have modified load_network function in darknetTR.cpp according to new tk::dnn::Yolo3Detection interface:

tk::dnn::Yolo3Detection* load_network(char* net_cfg, char* cfg_path, char* name_path, int n_classes, int n_batch, float conf_thresh)
{
    std::string net;
    net = net_cfg;
    tk::dnn::Yolo3Detection *detNN = new tk::dnn::Yolo3Detection;
    detNN->init(net, cfg_path, name_path, n_classes, n_batch, conf_thresh);

    return detNN;
}

also modified header file and input argtypes on python side accordingly: load_network.argtypes = [c_char_p, c_char_p, c_char_p, c_int, c_int, c_float]

In the meantime I'm reaching segfault while running the demo python script:

$ python darknetTR.py build/yolo4_fp16.rt --video=./demo/stabilized.mp4
build/yolo4_fp16.rt
New NetworkRT (TensorRT v8.01)
Float16 support: 1
Int8 support: 1
DLAs: 2
Segmentation fault (core dumped)

Any suggestions how to go on from here?

[gewenpulan]

any update here? I face the same problem. on jetpack4.5.1, ioir123jupython interface can not run correctly