Dropout

[Riccardo231]

Fixed a conceptual error present in nf_dropout_layer_submodule.f90. In order to get sum(input) == sum(output) the definition of scale should be sum(input) / sum(input * self % mask)

Proof:

sum(input) == sum(output)
sum(input) == sum(input * mask * scale)

Since scale is constant,
sum(input) == scale * sum(input * mask)
But scale == sum(input) / sum(input * self % mask), so:
sum(input) == sum(input) * sum(input * mask) / sum(input * mask)

[milancurcic]

Thanks @ricor07. Can you make a PR against #194 rather than main?

[Riccardo231]

Sorry I'm not very used with GitHub. Could you tell me how?

[milancurcic]

No problem and I appreciate your patience.

Here's how to submit a new PR against this one.

1. Fork my fork of this repository (https://github.com/milancurcic/neural-fortran)
2. Clone your newly forked repo: `git clone [email protected]:ricor07/neural-fortran (since you already have a fork of modern-fortran/neural-fortran, I'm not sure how GitHub will handle the name conflict; it may be that it will be called something else)
3. After cloning it, do

cd neural-fortran
git remote add upstream https://github.com/milancurcic/neural-fortran
git checkout dropout
# make your changes, add, and commit
git push origin dropout # this will push the changes to your fork

4. Navigate to https://github.com/milancurcic/neural-fortran and create a new PR; choose dropout as the target branch and your own fork's dropout branch as the source branch.

Let me know how it goes.

[Riccardo231]

I cloned your repo but it doesn't have the dropout layer file. Since it is only a line of change, could you change your code? Thank you.

In the next days, I wanted to work on #171. Is it possible? Thanks

[milancurcic]

Sure, thanks. I will add you as a co-author to the commit for proper attribution, probably tomorrow.

…

On Tue, Jan 28, 2025, 14:18 Riccardo Orsi ***@***.***> wrote: I cloned your repo but it doesn't have the dropout layer file. Since it is only a line of change, could you change your code? Thank you. In the next days, I wanted to work on #171 <#171>. Is it possible? Thanks — Reply to this email directly, view it on GitHub <#195 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA7RDPV7BCHXRLMNVKUBIMT2M7JWZAVCNFSM6AAAAABV4O2T7KVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMMJZHA3DENJTG4> . You are receiving this because you commented.Message ID: ***@***.***>