[Feature Request opencv-dnn] Swish Activation, Mish Activation

[LukeAI]

Google's famous Efficientnet uses the Swish Activation function
It is currently not possible to use efficientnet in opencv due to the absence of swish.

The newer Mish activation function reportedly outperforms swish and would be great to have as an option.

[dkurt]

It's possible for now to enable this network with custom layers mechanism. Please read a tutorial https://docs.opencv.org/master/dc/db1/tutorial_dnn_custom_layers.html.

[suab321321]

@dkurt @LukeAI I would like to help with this one

[dkurt]

@suab321321, thanks! Feel free to open a PR

[grassknoted]

Hi @dkurt seeing no progress, I was wondering if I could take this issue up? I had a look at the link:

Please read a tutorial https://docs.opencv.org/master/dc/db1/tutorial_dnn_custom_layers.html.

I'm still unsure of where to add the Mish activation function. Could you just help me get started, please?

[alalek]

Usage questions should go to Users OpenCV Q/A forum: http://answers.opencv.org
Custom layers are part of User code. Check "samples".

[grassknoted]

Thanks for the reply @alalek !
Just to clarify, I would need to create an Mish function example in the samples/ directory, right? Something like: samples/python/mish.py?

[alalek]

Well, if you want to add this implementation OpenCV then check this patch: #10889 + add tests on data from original framework (1-2 layers, to ensure that it is parsed properly).

Not sure if this is really "good first issue".

[dkurt]

As far as I know Swish is an activation layer so you need to implement it here: https://github.com/opencv/opencv/blob/master/modules/dnn/src/layers/elementwise_layers.cpp.

Also, you need to register a new name at https://github.com/opencv/opencv/blob/master/modules/dnn/src/init.cpp

and write some tests depends on origin framework.

[grassknoted]

Well, if you want to add this implementation OpenCV then check this patch: #10889 + add tests on data from original framework (1-2 layers, to ensure that it is parsed properly).

Not sure if this is really "good first issue".

Looking at the required changes, I'm not very sure I could take this on. I will instead look for more beginner focussed issues. Thank you @dkurt and @alalek for the help!