Open pnmartinez opened 3 years ago
@jdb78 I have updated this to pick other activation functions as a parameter like in:
net = NBeats.from_dataset(
training,
activation_fn = 'lrelu' # RELU by default
)
This way is easier to loop on different activation functions. In my case, Leaky_RELU
achieves same performance much faster (image).
I have a fork of my own with that (__init__.py
must be also modified). Let me know and I'll open a pull request.
Hi,
I am sharing some results on the performance of other activations.
In the problem I am attacking, it seems that some variants of RELU achieve same performance, but Leaky-RELU is faster.
@jdb78 I would consider making it the default in case this result is proofed to hold in other problems as well.
Very interesting analysis. Do you want to open a PR?
Maybe also relevant for #471
Hi Jan,
Throughout the week I will try to open it.
Cheers!
De: Jan Beitner @.> Enviado: jueves, 29 de abril de 2021 13:15 Para: jdb78/pytorch-forecasting @.> Cc: Pablo @.>; Author @.> Asunto: Re: [jdb78/pytorch-forecasting] [Code review request] N-Beats implemented with SELU() instead of RELU() (#444)
Very interesting analysis. Do you want to open a PR?
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHubhttps://github.com/jdb78/pytorch-forecasting/issues/444#issuecomment-829143984, or unsubscribehttps://github.com/notifications/unsubscribe-auth/AHEB2L6YQTE7DFSXDQSSOLTTLE5WDANCNFSM43BMTA2A.
Hi all,
If everything is in check, we can open a pull request to include this as a variant of N-beats.
I am interested in the
SELU
activation function for Self-Normalizing-Networks (SNNs, see, e.g., Pytorch docs). I didn't find any N-Beats corrected to use SELU and its requirements (i.e.AlphaDropout
, proper weights init), so I made an implementation myself patching thesub_modules.py
file inpytorch-forecasting
.It would be great if any of you with experience with these concepts -NBeats architecture,
pytorch-forecasting
, or SELU()- could review whether everything is right in my implementation.The implementation below and as Gist (tiny, commented modifications of the
/models/nbeats/sub_modules.py
in the lib): https://gist.github.com/pnmartinez/fef1f488497fa85a2cc1626af2a5b4bd