search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.65k stars 7.96k forks source link

Tiny Mish #6435

Open harsco-jfernandez opened 4 years ago

harsco-jfernandez commented 4 years ago

To make a mish tiny cfg which of the leaky layers would I need to set to mish?

Thanks

harsco-jfernandez commented 4 years ago

Hi @AlexeyAB,

Advice, please?

Thank you!

mrezq651 commented 4 years ago

hey, there I am trying to do the same. I've been researching and as far as I understand: mish requires significantly more processing power than the other activation functions. that is why in the Yolo 4 tiny weights mish is not used. since the author assumes that the user if he had a GPU or he would not need to use tiny weights. some proposed that there is a CPU optimized version of mish called hard mish. I checked that hard mish is implemented too in the original work of @AlexeyAB . I didn't test or implement the suggestion yet. I hope that this information helped you. please share your knowledge or the result after implementation with me when you solve the problem since I don't expect soon that I could test the code since I am working on a different part of the project

harsco-jfernandez commented 4 years ago

I thought it would be useful for something like a Jetson that is slightly better than tiny at the cost of fps. I made a mish model based on the 3 yolo layer model and it feels like the gains are marginal.

mrezq651 commented 4 years ago

I had in mind the same idea. firstly I would like to thank you for mentioning jetson. I noticed that its better than yolo3 on microprocessors. can you please provide any help (code, some references for Yolo 3 with jetson source code, and the model you made). any help is appreciated.

mrezq651 commented 4 years ago

this is my email: mrezq651@gmail.com

mrezq651 commented 4 years ago

still waiting your code

On Thu, Aug 27, 2020 at 9:59 PM harsco-jfernandez notifications@github.com wrote:

I thought it would be useful for something like a Jetson that is slightly better than tiny at the cost of fps. I made a mish model based on the 3 yolo layer model and it feels like the gains are marginal.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/AlexeyAB/darknet/issues/6435#issuecomment-682133596, or unsubscribe https://github.com/notifications/unsubscribe-auth/AQUPRI2Z553RKA467GZHUBTSC2UJNANCNFSM4PXX4JLA .

harsco-jfernandez commented 4 years ago

I don't have any code you don't already have