Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Added AlphaDropout which is used in SNNs. #656

Merged
merged 10 commits into from
Mar 7, 2019
Merged

Added AlphaDropout which is used in SNNs. #656

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
97f874a
Added AlphaDropout which is used in SNNs.
thebhatman Mar 3, 2019
b5533ee
Exported AlphaDropout
thebhatman Mar 3, 2019
922e9c9
Updated docs with AlphaDropout
thebhatman Mar 3, 2019
29b853e
Made sure Gradients are not lost.
thebhatman Mar 4, 2019
d660868
Suggested changes made
thebhatman Mar 5, 2019
8e5965a
Indentation fixed
thebhatman Mar 5, 2019
47c1324
Merge branch 'master' into patch-3
thebhatman Mar 7, 2019
c6e51f5
Made lambda and alpha of eltype(x)
thebhatman Mar 7, 2019
f4543b7
Value of alpha updated and dot operations changed
thebhatman Mar 7, 2019
1d310d4
Removed {typeof(p)}
thebhatman Mar 7, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions docs/src/models/layers.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,5 +50,6 @@ These layers don't affect the structure of the network but may improve training
Flux.testmode!
BatchNorm
Dropout
AlphaDropout
LayerNorm
```
2 changes: 1 addition & 1 deletion src/Flux.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@ using MacroTools, Juno, Requires, Reexport, Statistics, Random
using MacroTools: @forward

export Chain, Dense, RNN, LSTM, GRU, Conv, ConvTranspose, MaxPool, MeanPool,
DepthwiseConv, Dropout, LayerNorm, BatchNorm,
DepthwiseConv, Dropout, AlphaDropout, LayerNorm, BatchNorm,
params, mapleaves, cpu, gpu, f32, f64

@reexport using NNlib
Expand Down
29 changes: 29 additions & 0 deletions src/layers/normalise.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,6 +43,35 @@ end

_testmode!(a::Dropout, test) = (a.active = !test)

"""
AlphaDropout(p)
A dropout layer. It is used in Self-Normalizing Neural Networks.
(https://papers.nips.cc/paper/6698-self-normalizing-neural-networks.pdf)
The AlphaDropout layer ensures that mean and variance of activations remains the same as before.
"""
mutable struct AlphaDropout{F}
p::F
active::Bool
staticfloat marked this conversation as resolved.
Show resolved Hide resolved
end

function AlphaDropout(p)
@assert 0 ≤ p ≤ 1
AlphaDropout{typeof(p)}(p,true)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You shouldn't need {typeof(p)} here; it should be able to determine this from the p parameter:

julia> struct Foo{F}
       p::F
       active::Bool
       end

julia> Foo(1.0, true)
Foo{Float64}(1.0, true)

end

function (a::AlphaDropout)(x)
a.active || return x
α = eltype(x)(-1.75813631)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You should use the same α parameter as given by the selu() function in NNlib, see the definition here: https://github.com/FluxML/NNlib.jl/blob/d07ac0bfd3c71c3a29bc9c22becbba19227bbeb5/src/activation.jl#L100-L104

noise = randn(eltype(x), size(x))
x = @. x*(noise .> (1 - a.p)) + α .* (noise .<= (1 - a.p))
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think if you use @. you don't need the .> and .* and .<= operators; either use @. to get them all, or manually write them all as dotted operators.

A = (a.p + a.p * (1 - a.p) * α ^ 2)^0.5
B = -A * α * (1 - a.p)
x = @. A .* x .+ B
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Same thing here; this @. is not doing anything, because you have already dotted your operators.

return x
end

_testmode!(a::AlphaDropout, test) = (a.active = !test)

"""
LayerNorm(h::Integer)

Expand Down