ceccocats / tkDNN

Deep neural network library and toolkit to do high performace inference on NVIDIA jetson platforms
GNU General Public License v2.0
718 stars 208 forks source link

Yolov4-scaled #180

Closed AleBasso80 closed 2 years ago

AleBasso80 commented 3 years ago

Hi, I tried to export the AlexeyAB yolov4-scaled weights by means of your forked darknet version. I got an assert error at layer 273, I guess that this network is not supported yet. Do you have any plan to do it?

Thanks Alessandro

erikguo commented 3 years ago

There's a pull request to support logistic activation. You can tried. But I found the layer activation just before [yolo] layer is using linear not logistic.

mive93 commented 3 years ago

Hi @AleBasso80 I did not know about this new addition. I will look on that in the next days.

AleBasso80 commented 3 years ago

Hi @mive93, thanks. Its developers claim that this is new state-of-the-art detector.

AlexeyAB commented 3 years ago

@mive93 Previously logistic activation (sigmoid) was used inside [yolo] layer. Currently we use logistic activation (sigmoid) after the previous conv-layer.

LukeAI commented 3 years ago

@mive93 scaled-yolo is SOTA - would be great to see support for it here

AlexeyAB's medium: https://alexeyab84.medium.com/scaled-yolo-v4-is-the-best-neural-network-for-object-detection-on-ms-coco-dataset-39dfa22fa982

Paper: https://arxiv.org/abs/2011.08036

github: https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-large

network diagram https://netron.app/?url=https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4-csp.cfg

mive93 commented 3 years ago

Hi all @AlexeyAB @LukeAI @AleBasso80 @erikguo. Scaled Yolo 4 is now supported

https://github.com/ceccocats/tkDNN/commit/adac8576b0faf515ad3f459b1f50fd16cef6d64d

AleBasso80 commented 3 years ago

Hi @mive93, wonderful... Thanks a lot!!