Number of parameters

[ducha-aiki]

Hi,

If I understand source code correctly
https://github.com/forestagostinelli/Learned-Activation-Functions-Source/blob/master/src/caffe/layers/apl_layer.cpp#L71
and
https://github.com/forestagostinelli/Learned-Activation-Functions-Source/blob/master/src/caffe/layers/apl_layer.cpp#L30

then when used in convolution layers, APL learns not C_sums paramers, but H_W_C (== count/num) * sums parameters. So same channel in different locations does not share same APL coefficients. => APL with sums = 2 has h_w more parametrs than PReLU.
Is it intentional or typo?

[forestagostinelli]

Yes, that was the initial design. We are working on a new version that only
takes up C*sums or even just sums number of parameters.

On Sun, Apr 17, 2016 at 5:52 AM, Dmytro Mishkin notifications@github.com
wrote:

Hi,

If I understand source code correctly

https://github.com/forestagostinelli/Learned-Activation-Functions-Source/blob/master/src/caffe/layers/apl_layer.cpp#L71
and

https://github.com/forestagostinelli/Learned-Activation-Functions-Source/blob/master/src/caffe/layers/apl_layer.cpp#L30

then when used in convolution layers, APL learns not C_sums paramers, but
H_W_C (== count/num) * sums parameters. So same channel in different
locations does not share same APL coefficients. => APL with sums = 2 has h_w
more parametrs than PReLU.
Is it intentional or typo?

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub
#4