added ability to use pretrained darknet .weights

[LukeAI]

simple addition to allow training using pretrained darknet weights. have tested and can confirm it works.

[jaqub-manuel]

Hey, @LukeAI
Hello, I used your github project, the results seem a little better but still low compared to the U5 version (yolov4l-mish). I just set filter = 255 -> (18) and classes = 80-> (1) in the cfg file. Is there anything else to do? I've tried a lot but the results are still very different, for example where I got 82 with yolov4l-mish, I could not get more than 75 mAP with yolov4-csp (same as yolov4l-mish). It would be nice if you have something to recommend.
Thank you so much.

[LukeAI]

@jaqub-manuel I'm afraid I don't know - I gave up with yolov4-csp - I now use yolov4-p5-large at 640x640 - using tensorrt for inference it's fast enough and quite a bit more accurate than I could get from yolov4-csp

[jaqub-manuel]

@LukeAI, thanks for reply. I also have same roblem..do you find any yolov4.yaml for large branch.
I have only 500 images, so small model plays better.
U5 and scale brach are same..
btw, could you share tensortt repo.
thanks