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

Jump to bottom

Use faster activation functions #1837

Merged
merged 3 commits into from
Feb 5, 2022
Merged

Use faster activation functions #1837

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
7e4480b
fast activation functions
mcabbott Dec 12, 2021
3b43af0
Update src/layers/basic.jl
mcabbott Feb 5, 2022
0a8fade
Update src/layers/conv.jl
mcabbott Feb 5, 2022
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
3 changes: 2 additions & 1 deletion src/layers/basic.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -154,7 +154,8 @@ end
@functor Dense

function (a::Dense)(x::AbstractVecOrMat)
W, b, σ = a.weight, a.bias, a.σ
W, b= a.weight, a.bias
DhairyaLGandhi marked this conversation as resolved.
Show resolved Hide resolved
mcabbott marked this conversation as resolved.
Show resolved Hide resolved
σ = NNlib.fast_act(a.σ, x) # replaces tanh => tanh_fast, etc
return σ.(W*x .+ b)
end

Expand Down
12 changes: 8 additions & 4 deletions src/layers/conv.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -161,7 +161,8 @@ end
@functor Conv

function (c::Conv)(x::AbstractArray)
σ, b = c.σ, reshape(c.bias, ntuple(_ -> 1, length(c.stride))..., :, 1)
b = reshape(c.bias, ntuple(_ -> 1, length(c.stride))..., :, 1)
mcabbott marked this conversation as resolved.
Show resolved Hide resolved
σ = NNlib.fast_act(c.σ, x)
cdims = DenseConvDims(x, c.weight; stride = c.stride, padding = c.pad, dilation = c.dilation, groups = c.groups)
σ.(conv(x, c.weight, cdims) .+ b)
end
Expand Down Expand Up @@ -278,7 +279,8 @@ end
@nograd conv_transpose_dims

function (c::ConvTranspose)(x::AbstractArray)
σ, b = c.σ, reshape(c.bias, map(_->1, c.stride)..., :, 1)
b = reshape(c.bias, map(_->1, c.stride)..., :, 1)
σ = NNlib.fast_act(c.σ, x)
cdims = conv_transpose_dims(c, x)
σ.(∇conv_data(x, c.weight, cdims) .+ b)
end
Expand Down Expand Up @@ -371,7 +373,8 @@ depthwiseconvfilter(filter::NTuple{N,Integer}, ch::Pair{<:Integer,<:Integer};
init = glorot_uniform) where N = init(filter..., div(ch[2], ch[1]), ch[1])

function (c::DepthwiseConv)(x)
σ, b = c.σ, reshape(c.bias, map(_->1, c.stride)..., :, 1)
b = reshape(c.bias, map(_->1, c.stride)..., :, 1)
σ = NNlib.fast_act(c.σ, x)
cdims = DepthwiseConvDims(x, c.weight; stride=c.stride, padding=c.pad, dilation=c.dilation)
σ.(depthwiseconv(x, c.weight, cdims) .+ b)
end
Expand Down Expand Up @@ -450,7 +453,8 @@ function crosscor(x, w, ddims::DenseConvDims)
end

function (c::CrossCor)(x::AbstractArray)
σ, b = c.σ, reshape(c.bias, map(_->1, c.stride)..., :, 1)
b = reshape(c.bias, map(_->1, c.stride)..., :, 1)
σ = NNlib.fast_act(c.σ, x)
cdims = DenseConvDims(x, c.weight; stride=c.stride, padding=c.pad, dilation=c.dilation)
σ.(crosscor(x, c.weight, cdims) .+ b)
end
Expand Down
11 changes: 6 additions & 5 deletions src/layers/recurrent.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -117,7 +117,8 @@ RNNCell(in::Integer, out::Integer, σ=tanh; init=Flux.glorot_uniform, initb=zero
RNNCell(σ, init(out, in), init(out, out), initb(out), init_state(out,1))

function (m::RNNCell{F,A,V,<:AbstractMatrix{T}})(h, x::Union{AbstractVecOrMat{T},OneHotArray}) where {F,A,V,T}
σ, Wi, Wh, b = m.σ, m.Wi, m.Wh, m.b
Wi, Wh, b = m.Wi, m.Wh, m.b
σ = NNlib.fast_act(m.σ, x)
h = σ.(Wi*x .+ Wh*h .+ b)
return h, reshape_cell_output(h, x)
end
Expand Down Expand Up @@ -224,8 +225,8 @@ function (m::LSTMCell{A,V,<:NTuple{2,AbstractMatrix{T}}})((h, c), x::Union{Abstr
b, o = m.b, size(h, 1)
g = m.Wi*x .+ m.Wh*h .+ b
input, forget, cell, output = multigate(g, o, Val(4))
c′ = @. σ(forget) * c + σ(input) * tanh(cell)
h′ = @. σ(output) * tanh(c′)
c′ = @. sigmoid_fast(forget) * c + sigmoid_fast(input) * tanh_fast(cell)
h′ = @. sigmoid_fast(output) * tanh_fast(c′)
Comment on lines +228 to +229
Copy link
Member

Choose a reason for hiding this comment

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

Ought these to get the fast_act treatment too? I'm ok with revisiting too, RNN cell inflexibility is a bit of a long standing issue.

Copy link
Member Author

Choose a reason for hiding this comment

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

Oh right. If we decide to disable fast_tanh on CuArrays, that will be ignored here. But perhaps revisit if & when... it's a bit more clutter to squeeze that in.

Copy link
Member

Choose a reason for hiding this comment

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

It may be a blessing in disguise, as currently plain tanh.(...) will hit the likely obsolete https://github.com/FluxML/NNlibCUDA.jl/blob/master/src/cudnn/activations.jl.

return (h′, c′), reshape_cell_output(h′, x)
end

Expand Down Expand Up @@ -309,7 +310,7 @@ function (m::GRUCell{A,V,<:AbstractMatrix{T}})(h, x::Union{AbstractVecOrMat{T},O
Wi, Wh, b, o = m.Wi, m.Wh, m.b, size(h, 1)
gxs, ghs, bs = multigate(Wi*x, o, Val(3)), multigate(Wh*h, o, Val(3)), multigate(b, o, Val(3))
r, z = _gru_output(gxs, ghs, bs)
h̃ = @. tanh(gxs[3] + r * ghs[3] + bs[3])
h̃ = @. tanh_fast(gxs[3] + r * ghs[3] + bs[3])
h′ = @. (1 - z) * h̃ + z * h
return h′, reshape_cell_output(h′, x)
end
Expand Down Expand Up @@ -387,7 +388,7 @@ function (m::GRUv3Cell{A,V,<:AbstractMatrix{T}})(h, x::Union{AbstractVecOrMat{T}
Wi, Wh, b, Wh_h̃, o = m.Wi, m.Wh, m.b, m.Wh_h̃, size(h, 1)
gxs, ghs, bs = multigate(Wi*x, o, Val(3)), multigate(Wh*h, o, Val(2)), multigate(b, o, Val(3))
r, z = _gru_output(gxs, ghs, bs)
h̃ = tanh.(gxs[3] .+ (Wh_h̃ * (r .* h)) .+ bs[3])
h̃ = tanh_fast.(gxs[3] .+ (Wh_h̃ * (r .* h)) .+ bs[3])
h′ = @. (1 - z) * h̃ + z * h
return h′, reshape_cell_output(h′, x)
end
Expand Down