search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.71k stars 7.96k forks source link

YOLOv4-Tiny Questions, hoping for an answer. #6302

Open francismontalbo opened 4 years ago

francismontalbo commented 4 years ago

Hello, just wondering why YOLOv4-Tiny uses leaky ReLU instead of Mish or Swish? I can't seem to find an article about this.

is there a way for me to visualize or see what is the structure of the yolov4-tiny.conv.29?

Also, are there papers that revolve around the YOLO tiny versions?

Thank you again for everyone helping me.

AlexeyAB commented 4 years ago

Mish/Swish can work very slow on some embedded devices due to exp-function.

Look at first 29 layers: https://lutzroeder.github.io/netron/?url=https%3A%2F%2Fraw.githubusercontent.com%2FAlexeyAB%2Fdarknet%2Fmaster%2Fcfg%2Fyolov4-tiny.cfg

francismontalbo commented 4 years ago

Mish/Swish can work very slow on some embedded devices due to exp-function.

Look at first 29 layers: https://lutzroeder.github.io/netron/?url=https%3A%2F%2Fraw.githubusercontent.com%2FAlexeyAB%2Fdarknet%2Fmaster%2Fcfg%2Fyolov4-tiny.cfg

Thank you so much @AlexeyAB you are a very helpful person. Be safe.

harsco-jfernandez commented 4 years ago

hi @AlexeyAB,

Would the pre-train weights(yolov4-tiny.conv.29) work if I change the activation from leaky to mish and leave everything else as is?

thanks!