diff --git a/dev/index.html b/dev/index.html index d26b01ebe..96f65d4c4 100644 --- a/dev/index.html +++ b/dev/index.html @@ -1,2 +1,2 @@ -Home · NNlib.jl
Edit on GitHub

NNlib.jl

NNlib provides a library of functions useful for neural networks, such as softmax, sigmoid, batched multiplication, convolutions and pooling. Many of these are used by Flux.jl, which loads this package, but they may be used independently.

For use with automatic differentiation, this package defines gradients using ChainRules.jl. These will be seen by various packages including Zygote.jl.

GPU support is provided as package extensions. In order to load the extensions, use the imports

using NNlib, CUDA, cuDNN

for CUDA support, or

using NNlib, AMDGPU

for AMDGPU support.

+Home · NNlib.jl
Edit on GitHub

NNlib.jl

NNlib provides a library of functions useful for neural networks, such as softmax, sigmoid, batched multiplication, convolutions and pooling. Many of these are used by Flux.jl, which loads this package, but they may be used independently.

For use with automatic differentiation, this package defines gradients using ChainRules.jl. These will be seen by various packages including Zygote.jl.

GPU support is provided as package extensions. In order to load the extensions, use the imports

using NNlib, CUDA, cuDNN

for CUDA support, or

using NNlib, AMDGPU

for AMDGPU support.

diff --git a/dev/reference/index.html b/dev/reference/index.html index 6c22ba166..854efc02d 100644 --- a/dev/reference/index.html +++ b/dev/reference/index.html @@ -13,7 +13,7 @@ ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀ julia> celu(-10f0) --0.9999546f0source
NNlib.elu — Function
elu(x, α=1) = x > 0 ? x : α * (exp(x) - 1)

Exponential Linear Unit activation function. See "Fast and Accurate Deep Network Learning by Exponential Linear Units". You can also specify the coefficient explicitly, e.g. elu(x, 1).

julia> lineplot(elu, -2, 2, height=7)
+-0.9999546f0
source
NNlib.elu — Function
elu(x, α=1) = x > 0 ? x : α * (exp(x) - 1)

Exponential Linear Unit activation function. See "Fast and Accurate Deep Network Learning by Exponential Linear Units". You can also specify the coefficient explicitly, e.g. elu(x, 1).

julia> lineplot(elu, -2, 2, height=7)
            ┌────────────────────────────────────────┐       
          2 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⠤⠒⠉│ elu(x)
            │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⡠⠔⠊⠉⠀⠀⠀⠀│       
@@ -30,7 +30,7 @@
 -0.9999546f0
 
 julia> elu(-10f0, 2)
--1.9999092f0
source
NNlib.gelu — Function
gelu(x) = 0.5x * (1 + tanh(√(2/π) * (x + 0.044715x^3)))

Activation function from "Gaussian Error Linear Units".

julia> lineplot(gelu, -2, 2, height=7)
+-1.9999092f0
source
NNlib.gelu — Function
gelu(x) = 0.5x * (1 + tanh(√(2/π) * (x + 0.044715x^3)))

Activation function from "Gaussian Error Linear Units".

julia> lineplot(gelu, -2, 2, height=7)
            ┌────────────────────────────────────────┐        
          2 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⡠⠔⠊│ gelu(x)
            │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⠔⠊⠁⠀⠀⠀│        
@@ -56,7 +56,7 @@
         -0.2 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠓⣄⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢠⠇⠀⠀⠀│         
              └────────────────────────────────────────┘         
              ⠀-5⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀0⠀         
-             ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.hardsigmoid — Function
hardσ(x) = max(0, min(1, (x + 3) / 6))

Piecewise linear approximation of sigmoid.

julia> lineplot(hardsigmoid, -5, 5, height=7)
+             ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.hardsigmoid — Function
hardσ(x) = max(0, min(1, (x + 3) / 6))

Piecewise linear approximation of sigmoid.

julia> lineplot(hardsigmoid, -5, 5, height=7)
           ┌────────────────────────────────────────┐         
         1 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⢀⡠⠖⠋⠉⠉⠉⠉⠉⠉⠉⠉│ hardσ(x)
           │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⣀⡤⠒⠁⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│         
@@ -80,14 +80,14 @@
         0 │⣀⣀⣀⣀⣀⣀⣀⠤⠤⠤⠒⠊⠉⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│     
           └────────────────────────────────────────┘     
           ⠀-5⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀5⠀     
-          ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.sigmoid_fast — Function
sigmoid_fast(x)

This is a faster, and very slightly less accurate, version of sigmoid. For `x::Float32, perhaps 3 times faster, and maximum errors 2 eps instead of 1.

See also tanh_fast.

julia> sigmoid(0.2f0)
+          ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.sigmoid_fast — Function
sigmoid_fast(x)

This is a faster, and very slightly less accurate, version of sigmoid. For `x::Float32, perhaps 3 times faster, and maximum errors 2 eps instead of 1.

See also tanh_fast.

julia> sigmoid(0.2f0)
 0.54983395f0
 
 julia> sigmoid_fast(0.2f0)
 0.54983395f0
 
 julia> hardσ(0.2f0)
-0.53333336f0
source
NNlib.hardtanh — Function
hardtanh(x) = max(-1, min(1, x))

Segment-wise linear approximation of tanh, much cheaper to compute. See "Large Scale Machine Learning".

See also tanh_fast.

julia> lineplot(hardtanh, -2, 2, height=7)
+0.53333336f0
source
NNlib.hardtanh — Function
hardtanh(x) = max(-1, min(1, x))

Segment-wise linear approximation of tanh, much cheaper to compute. See "Large Scale Machine Learning".

See also tanh_fast.

julia> lineplot(hardtanh, -2, 2, height=7)
            ┌────────────────────────────────────────┐            
          1 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⣀⠔⠋⠉⠉⠉⠉⠉⠉⠉⠉⠉⠉│ hardtanh(x)
            │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⣀⡤⠊⠁⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│            
@@ -111,14 +111,14 @@
         -1 │⣀⣀⣀⡠⠤⠤⠤⠖⠒⠊⠉⠁⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│        
            └────────────────────────────────────────┘        
            ⠀-2⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀2⠀        
-           ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.tanh_fast — Function
tanh_fast(x)

This is a faster but slighly less accurate version of tanh.

Where Julia's tanh function has an error under 2 eps, this may be wrong by 5 eps, a reduction by less than one decimal digit.

For x::Float32 this is usually about 10 times faster, with a smaller speedup for x::Float64. For any other number types, it just calls tanh.

See also sigmoid_fast.

julia> tanh(0.5f0)
+           ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.tanh_fast — Function
tanh_fast(x)

This is a faster but slighly less accurate version of tanh.

Where Julia's tanh function has an error under 2 eps, this may be wrong by 5 eps, a reduction by less than one decimal digit.

For x::Float32 this is usually about 10 times faster, with a smaller speedup for x::Float64. For any other number types, it just calls tanh.

See also sigmoid_fast.

julia> tanh(0.5f0)
 0.46211717f0
 
 julia> tanh_fast(0.5f0)
 0.46211714f0
 
 julia> hard_tanh(0.5f0)
-0.5f0
source
NNlib.leakyrelu — Function
leakyrelu(x, a=0.01) = max(a*x, x)

Leaky Rectified Linear Unit activation function. You can also specify the coefficient explicitly, e.g. leakyrelu(x, 0.01).

julia> lineplot(x -> leakyrelu(x, 0.5), -2, 2, height=7)
+0.5f0
source
NNlib.leakyrelu — Function
leakyrelu(x, a=0.01) = max(a*x, x)

Leaky Rectified Linear Unit activation function. You can also specify the coefficient explicitly, e.g. leakyrelu(x, 0.01).

julia> lineplot(x -> leakyrelu(x, 0.5), -2, 2, height=7)
            ┌────────────────────────────────────────┐       
          2 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⠤⠒⠉│ #42(x)
            │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⡠⠔⠊⠉⠀⠀⠀⠀│       
@@ -135,7 +135,7 @@
 -2.0f0
 
 julia> leakyrelu(-10f0, 0.02)
--0.5f0
source
NNlib.lisht — Function
lisht(x) = x * tanh(x)

Activation function from "LiSHT: Non-Parametric Linearly Scaled Hyperbolic Tangent ..."

julia> lineplot(lisht, -2, 2, height=7)
+-0.5f0
source
NNlib.lisht — Function
lisht(x) = x * tanh(x)

Activation function from "LiSHT: Non-Parametric Linearly Scaled Hyperbolic Tangent ..."

julia> lineplot(lisht, -2, 2, height=7)
           ┌────────────────────────────────────────┐         
         2 │⠢⣄⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⠔│ lisht(x)
           │⠀⠈⠑⢦⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⡤⠊⠁⠀│         
@@ -159,7 +159,7 @@
         0 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠉⠓⠪⠷⣦⣄⣀⣀⣇⣀⣀⣤⠶⠕⠒⠉⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│           
           └────────────────────────────────────────┘           
           ⠀-2⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀2⠀           
-          ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.logcosh — Function
logcosh(x)

Return log(cosh(x)) which is computed in a numerically stable way.

julia> lineplot(logcosh, -5, 5, height=7)
+          ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.logcosh — Function
logcosh(x)

Return log(cosh(x)) which is computed in a numerically stable way.

julia> lineplot(logcosh, -5, 5, height=7)
           ┌────────────────────────────────────────┐           
         5 │⡀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│ logcosh(x)
           │⠉⠢⣄⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⠔⠋│           
@@ -170,7 +170,7 @@
         0 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠑⠢⢄⣀⣀⣇⣀⡠⠔⠊⠁⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│           
           └────────────────────────────────────────┘           
           ⠀-5⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀5⠀           
-          ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.logsigmoid — Function
logσ(x)

Return log(σ(x)) which is computed in a numerically stable way.

julia> lineplot(logsigmoid, -5, 5, height=7)
+          ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.logsigmoid — Function
logσ(x)

Return log(σ(x)) which is computed in a numerically stable way.

julia> lineplot(logsigmoid, -5, 5, height=7)
            ┌────────────────────────────────────────┐        
          0 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⡧⠤⠔⠒⠒⠒⠉⠉⠉⠉⠉⠉⠉⠉⠉⠉⠉⠉⠉⠉│ logσ(x)
            │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⡤⠖⠊⠉⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│        
@@ -181,7 +181,7 @@
         -6 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│        
            └────────────────────────────────────────┘        
            ⠀-5⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀5⠀        
-           ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.mish — Function
mish(x) = x * tanh(softplus(x))

Activation function from "Mish: A Self Regularized Non-Monotonic Neural Activation Function".

julia> lineplot(mish, -5, 5, height=7)
+           ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.mish — Function
mish(x) = x * tanh(softplus(x))

Activation function from "Mish: A Self Regularized Non-Monotonic Neural Activation Function".

julia> lineplot(mish, -5, 5, height=7)
            ┌────────────────────────────────────────┐        
          5 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⡠⠖⠋│ mish(x)
            │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⡤⠒⠁⠀⠀⠀│        
@@ -192,7 +192,7 @@
         -1 │⠀⠀⠀⠀⠀⠀⠀⠀⠉⠉⠉⠉⠉⠉⠉⠉⠉⠉⠉⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│        
            └────────────────────────────────────────┘        
            ⠀-5⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀5⠀        
-           ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.relu — Function
relu(x) = max(0, x)

Rectified Linear Unit activation function.

julia> lineplot(relu, -2, 2, height=7)
+           ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.relu — Function
relu(x) = max(0, x)

Rectified Linear Unit activation function.

julia> lineplot(relu, -2, 2, height=7)
           ┌────────────────────────────────────────┐        
         2 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⠔⠋│ relu(x)
           │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⡤⠊⠁⠀⠀│        
@@ -203,7 +203,7 @@
         0 │⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣇⠔⠋⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│        
           └────────────────────────────────────────┘        
           ⠀-2⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀2⠀        
-          ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.relu6 — Function
relu6(x) = min(max(0, x), 6)

Rectified Linear Unit activation function capped at 6. See "Convolutional Deep Belief Networks" from CIFAR-10.

julia> lineplot(relu6, -10, 10, height=7)
+          ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.relu6 — Function
relu6(x) = min(max(0, x), 6)

Rectified Linear Unit activation function capped at 6. See "Convolutional Deep Belief Networks" from CIFAR-10.

julia> lineplot(relu6, -10, 10, height=7)
           ┌────────────────────────────────────────┐         
         6 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⠎⠉⠉⠉⠉⠉⠉⠉⠉│ relu6(x)
           │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⢀⡔⠁⠀⠀⠀⠀⠀⠀⠀⠀⠀│         
@@ -214,7 +214,7 @@
         0 │⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⡧⠋⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│         
           └────────────────────────────────────────┘         
           ⠀-10⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀10⠀         
-          ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.rrelu — Function
rrelu(x, lo=1/8, hi=1/3) = max(a*x, x)
+          ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.rrelu — Function
rrelu(x, lo=1/8, hi=1/3) = max(a*x, x)
 # where `a` is randomly sampled from uniform distribution `U(lo, hi)`

Randomized Leaky Rectified Linear Unit activation function. See "Empirical Evaluation of Rectified Activations" You can also specify the bound explicitly, e.g. rrelu(x, 0.0, 1.0).

julia> lineplot(rrelu, -20, 10, height=7)
             ┌────────────────────────────────────────┐         
          10 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢸⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⡤⠖⠋│ rrelu(x)
@@ -229,7 +229,7 @@
             ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀         
 
 julia> extrema(rrelu.(fill(-10f0, 1000)))
-(-3.3316886f0, -1.2548422f0)
source
NNlib.selu — Function
selu(x) = λ * (x ≥ 0 ? x : α * (exp(x) - 1))
+(-3.3316886f0, -1.2548422f0)
source
NNlib.selu — Function
selu(x) = λ * (x ≥ 0 ? x : α * (exp(x) - 1))
 
 λ ≈ 1.05070...
 α ≈ 1.67326...

Scaled exponential linear units. See "Self-Normalizing Neural Networks".

julia> lineplot(selu, -3, 2, height=7)
@@ -246,7 +246,7 @@
            ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀        
 
 julia> selu(-10f0)
--1.7580194f0
source
NNlib.sigmoid — Function
σ(x) = 1 / (1 + exp(-x))

Classic sigmoid activation function. Unicode σ can be entered as \sigma then tab, in many editors. The ascii name sigmoid is also exported.

See also sigmoid_fast.

julia> using UnicodePlots
+-1.7580194f0
source
NNlib.sigmoid — Function
σ(x) = 1 / (1 + exp(-x))

Classic sigmoid activation function. Unicode σ can be entered as \sigma then tab, in many editors. The ascii name sigmoid is also exported.

See also sigmoid_fast.

julia> using UnicodePlots
 
 julia> lineplot(sigmoid, -5, 5, height=7)
           ┌────────────────────────────────────────┐     
@@ -262,7 +262,7 @@
           ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀     
 
 julia> sigmoid === σ
-true
source
NNlib.softplus — Function
softplus(x) = log(exp(x) + 1)

See "Deep Sparse Rectifier Neural Networks", JMLR 2011.

julia> lineplot(softplus, -3, 3, height=7)
+true
source
NNlib.softplus — Function
softplus(x) = log(exp(x) + 1)

See "Deep Sparse Rectifier Neural Networks", JMLR 2011.

julia> lineplot(softplus, -3, 3, height=7)
           ┌────────────────────────────────────────┐            
         4 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│ softplus(x)
           │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⡠│            
@@ -289,7 +289,7 @@
           ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀            
 
 julia> softplus(16f0)
-16.0f0
source
NNlib.softshrink — Function
softshrink(x, λ=0.5) =
+16.0f0
source
NNlib.softshrink — Function
softshrink(x, λ=0.5) =
     (x ≥ λ ? x - λ : (-λ ≥ x ? x + λ : 0))

See "Softshrink Activation Function".

julia> lineplot(softshrink, -2, 2, height=7)
            ┌────────────────────────────────────────┐              
          2 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀│ softshrink(x)
@@ -317,7 +317,7 @@
            ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀
 
 julia> softshrink.((-10f0, 10f0))
-(-9.5f0, 9.5f0)
source
NNlib.softsign — Function
softsign(x) = x / (1 + |x|)

See "Quadratic Polynomials Learn Better Image Features" (2009).

julia> lineplot(softsign, -5, 5, height=7)
+(-9.5f0, 9.5f0)
source
NNlib.softsign — Function
softsign(x) = x / (1 + |x|)

See "Quadratic Polynomials Learn Better Image Features" (2009).

julia> lineplot(softsign, -5, 5, height=7)
            ┌────────────────────────────────────────┐            
          1 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⣀⣀⣀⣀⠤⠤⠤⠤⠤│ softsign(x)
            │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⣀⡤⠖⠒⠋⠉⠉⠁⠀⠀⠀⠀⠀⠀⠀⠀⠀│            
@@ -347,7 +347,7 @@
 0.5f0
 
 julia> softsign(100f0)
-0.990099f0
source
NNlib.swish — Function
swish(x) = x * σ(x)

Self-gated activation function. See "Swish: a Self-Gated Activation Function".

julia> lineplot(swish, -2, 2, height=7)
+0.990099f0
source
NNlib.swish — Function
swish(x) = x * σ(x)

Self-gated activation function. See "Swish: a Self-Gated Activation Function".

julia> lineplot(swish, -2, 2, height=7)
            ┌────────────────────────────────────────┐         
          2 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⡤│ swish(x)
            │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⡤⠖⠋⠁⠀│         
@@ -358,7 +358,7 @@
         -1 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│         
            └────────────────────────────────────────┘         
            ⠀-2⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀2⠀         
-           ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.hardswish — Function
hardswish(x) = x * hardσ(x)

Hard-Swish activation function. See "Searching for MobileNetV3".

julia> lineplot(hardswish, -2, 5, height = 7)
+           ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source
NNlib.hardswish — Function
hardswish(x) = x * hardσ(x)

Hard-Swish activation function. See "Searching for MobileNetV3".

julia> lineplot(hardswish, -2, 5, height = 7)
            ┌────────────────────────────────────────┐             
          5 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⡤⠔⠒⠉│ hardswish(x)
            │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⡠⠔⠒⠉⠁⠀⠀⠀⠀│             
@@ -388,7 +388,7 @@
 
 julia> hardswish.(-5:5)'
 1×11 adjoint(::Vector{Float64}) with eltype Float64:
- -0.0  -0.0  -0.0  -0.333333  -0.333333  0.0  0.666667  1.66667  3.0  4.0  5.0
source
NNlib.tanhshrink — Function
tanhshrink(x) = x - tanh(x)

See "Tanhshrink Activation Function".

julia> lineplot(tanhshrink, -3, 3, height=7)
+ -0.0  -0.0  -0.0  -0.333333  -0.333333  0.0  0.666667  1.66667  3.0  4.0  5.0
source
NNlib.tanhshrink — Function
tanhshrink(x) = x - tanh(x)

See "Tanhshrink Activation Function".

julia> lineplot(tanhshrink, -3, 3, height=7)
            ┌────────────────────────────────────────┐              
          3 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│ tanhshrink(x)
            │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⡠⠤⠖⠊│              
@@ -402,7 +402,7 @@
            ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀              
 
 julia> tanhshrink.((-10f0, 10f0))
-(-9.0f0, 9.0f0)
source
NNlib.trelu — Function
trelu(x, theta=1) = x > theta ? x : 0

Threshold gated rectified linear activation function. See "Zero-bias autoencoders and the benefits of co-adapting features"

julia> lineplot(trelu, -2, 4, height=7)
+(-9.0f0, 9.0f0)
source
NNlib.trelu — Function
trelu(x, theta=1) = x > theta ? x : 0

Threshold gated rectified linear activation function. See "Zero-bias autoencoders and the benefits of co-adapting features"

julia> lineplot(trelu, -2, 4, height=7)
           ┌────────────────────────────────────────┐         
         4 │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⣀⡤⠖⠋│ trelu(x)
           │⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⢀⡠⠖⠋⠁⠀⠀⠀│         
@@ -413,8 +413,8 @@
         0 │⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣀⣇⣀⣀⣀⣀⣀⣀⡇⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀│         
           └────────────────────────────────────────┘         
           ⠀-2⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀4⠀         
-          ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source

Attention

NNlib.dot_product_attention — Function
dot_product_attention(query, key, value, [bias]; [fdrop, mask, nheads])

Multihead dot product attention used in transformer architectures.

The input arrays must have the first two dimensions given by the number of features and the sequence length, then an arbitrary number of batch dimensions or none.

Returns the attention output array of size (v_dim, q_len, batch_size...) and the attention scores of size (kv_len, q_len, nheads, batch_size...).

See also dot_product_attention_scores if you only need the attention scores.

Arguments

  • query: Query array of size (qk_dim, q_len, batch_size...).
  • key: Key array of size (qk_dim, kv_len, batch_size...).
  • value: Value array of size (v_dim, kv_len, batch_size...).
  • bias: Either nothing or an array broadcastable to size (kv_len, q_len, nheads, batch_size). It will be added to the attention scores before applying the softmax. Default nothing.
  • fdrop: A dropout function or layer to be applied on the attention scores right after the softmax. Default identity (no dropout).
  • mask: Either nothing or a boolean array broadcastable to size (kv_len, q_len, nheads, batch_size). The mask is applied to the attention scores just before the softmax. See make_causal_mask fore creating causal masks. Default nothing.
  • nheads: Number of heads to split the input arrays into. Default 1.

Examples

q, k, v = rand(10, 20, 2), rand(10, 30, 2), rand(20, 30, 2)
-y, α = dot_product_attention(q, k, v)
source
NNlib.dot_product_attention_scores — Function
dot_product_attention_scores(query, key, [bias]; [fdrop, mask])

Return the attention scores for the dot_product_attention. Input arrays must have dimensions (num_features ÷ nheads, nheads, sequence_length, batch_size).

See dot_product_attention for more details.

source
NNlib.make_causal_mask — Function
make_causal_mask(x, dims=2)

Return a boolean square matrix m of the same type as x and of side size(x, dims). Its elements are set such that m[i, j] == i ≤ j.

Can be used to mask the attention scores in dot_product_attention.

source

Softmax

Flux's logitcrossentropy uses NNlib.softmax internally.

NNlib.softmax — Function
softmax(x; dims = 1)

Softmax turns input array x into probability distributions that sum to 1 along the dimensions specified by dims. It is semantically equivalent to the following:

softmax(x; dims = 1) = exp.(x) ./ sum(exp.(x), dims = dims)

with additional manipulations enhancing numerical stability.

For a matrix input x it will by default (dims = 1) treat it as a batch of vectors, with each column independent. Keyword dims = 2 will instead treat rows independently, and so on.

See also logsoftmax.

Examples

julia> softmax([1, 2, 3])
+          ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀x⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀
source

Attention

NNlib.dot_product_attention — Function
dot_product_attention(query, key, value, [bias]; [fdrop, mask, nheads])

Multihead dot product attention used in transformer architectures.

The input arrays must have the first two dimensions given by the number of features and the sequence length, then an arbitrary number of batch dimensions or none.

Returns the attention output array of size (v_dim, q_len, batch_size...) and the attention scores of size (kv_len, q_len, nheads, batch_size...).

See also dot_product_attention_scores if you only need the attention scores.

Arguments

  • query: Query array of size (qk_dim, q_len, batch_size...).
  • key: Key array of size (qk_dim, kv_len, batch_size...).
  • value: Value array of size (v_dim, kv_len, batch_size...).
  • bias: Either nothing or an array broadcastable to size (kv_len, q_len, nheads, batch_size). It will be added to the attention scores before applying the softmax. Default nothing.
  • fdrop: A dropout function or layer to be applied on the attention scores right after the softmax. Default identity (no dropout).
  • mask: Either nothing or a boolean array broadcastable to size (kv_len, q_len, nheads, batch_size). The mask is applied to the attention scores just before the softmax. See make_causal_mask fore creating causal masks. Default nothing.
  • nheads: Number of heads to split the input arrays into. Default 1.

Examples

q, k, v = rand(10, 20, 2), rand(10, 30, 2), rand(20, 30, 2)
+y, α = dot_product_attention(q, k, v)
source
NNlib.dot_product_attention_scores — Function
dot_product_attention_scores(query, key, [bias]; [fdrop, mask])

Return the attention scores for the dot_product_attention. Input arrays must have dimensions (num_features ÷ nheads, nheads, sequence_length, batch_size).

See dot_product_attention for more details.

source
NNlib.make_causal_mask — Function
make_causal_mask(x, dims=2)

Return a boolean square matrix m of the same type as x and of side size(x, dims). Its elements are set such that m[i, j] == i ≤ j.

Can be used to mask the attention scores in dot_product_attention.

source

Softmax

Flux's logitcrossentropy uses NNlib.softmax internally.

NNlib.softmax — Function
softmax(x; dims = 1)

Softmax turns input array x into probability distributions that sum to 1 along the dimensions specified by dims. It is semantically equivalent to the following:

softmax(x; dims = 1) = exp.(x) ./ sum(exp.(x), dims = dims)

with additional manipulations enhancing numerical stability.

For a matrix input x it will by default (dims = 1) treat it as a batch of vectors, with each column independent. Keyword dims = 2 will instead treat rows independently, and so on.

See also logsoftmax.

Examples

julia> softmax([1, 2, 3])
 3-element Vector{Float64}:
  0.09003057317038046
  0.24472847105479764
@@ -438,8 +438,8 @@
 (7, 13)
 
 julia> Dense(4 => 7, softmax)(x)
-ERROR: `softmax(x)` called with a number, but it expects an array.
source
NNlib.logsoftmax — Function
logsoftmax(x; dims = 1)

Computes the log of softmax in a more numerically stable way than directly taking log.(softmax(xs)). Commonly used in computing cross entropy loss.

It is semantically equivalent to the following:

logsoftmax(x; dims = 1) = x .- log.(sum(exp.(x), dims = dims))

See also softmax.

source

Pooling

Flux's AdaptiveMaxPool, AdaptiveMeanPool, GlobalMaxPool, GlobalMeanPool, MaxPool, MeanPool and lpnormpool use NNlib.PoolDims, NNlib.maxpool, NNlib.meanpool and NNlib.lpnormpool as their backend.

NNlib.PoolDims — Type
PoolDims(x_size::NTuple{M}, k::Union{NTuple{L, Int}, Int};
-        stride=k, padding=0, dilation=1)  where {M, L}

Dimensions for a "pooling" operation that can have an arbitrary input size, kernel size, stride, dilation, and channel count. Used to dispatch onto efficient implementations at compile-time.

source
NNlib.maxpool — Function
maxpool(x, k::NTuple{N, Integer}; pad=0, stride=k)

Perform max pool operation with window size k on input tensor x.

Arguments:

  • x and k: Expects ndim(x) ∈ 3:5, and always length(k) == ndim(x) - 2
  • pad: See pad_zeros for details.
  • stride: Either a tuple with the same length as k, or one integer for all directions. Default is k.
source
NNlib.meanpool — Function
meanpool(x, k::NTuple{N, Integer}; pad=0, stride=k)

Perform mean pool operation with window size k on input tensor x.

Arguments:

  • x and k: Expects ndim(x) ∈ 3:5, and alwayslength(k) == ndim(x) - 2`
  • pad: See pad_zeros for details.
  • stride: Either a tuple with the same length as k, or one integer for all directions. Default is k.
source
NNlib.lpnormpool — Function
lpnormpool(x, p::Real, k::NTuple{N, Integer}; pad=0, stride=k)

Perform Lp pool operation with value of the Lp norm p and window size k on input tensor x, also known as LPPool in pytorch. This pooling operator from Learned-Norm Pooling for Deep Feedforward and Recurrent Neural Networks.

Arguments:

  • x and k: Expects ndim(x) ∈ 3:5, and alwayslength(k) == ndim(x) - 2`
  • p is restricted to 0 < p < Inf.
  • pad: See pad_zeros for details.
  • stride: Either a tuple with the same length as k, or one integer for all directions. Default is k.

For all elements x in a size k window, lpnormpool computes (∑ᵢ xᵢ^p)^(1 / p) as an element of the output.

Thus lpnormpool(x, 1, k) ./ prod(k) ≈ meanpool(x, k) and lpnormpool(x, 2, k).^2 ./ prod(k) ≈ meanpool(x.^2, k).

source

Padding

NNlib.pad_reflect — Function
pad_reflect(x, pad::Tuple; [dims])
+ERROR: `softmax(x)` called with a number, but it expects an array.
source
NNlib.logsoftmax — Function
logsoftmax(x; dims = 1)

Computes the log of softmax in a more numerically stable way than directly taking log.(softmax(xs)). Commonly used in computing cross entropy loss.

It is semantically equivalent to the following:

logsoftmax(x; dims = 1) = x .- log.(sum(exp.(x), dims = dims))

See also softmax.

source

Pooling

Flux's AdaptiveMaxPool, AdaptiveMeanPool, GlobalMaxPool, GlobalMeanPool, MaxPool, MeanPool and lpnormpool use NNlib.PoolDims, NNlib.maxpool, NNlib.meanpool and NNlib.lpnormpool as their backend.

NNlib.PoolDims — Type
PoolDims(x_size::NTuple{M}, k::Union{NTuple{L, Int}, Int};
+        stride=k, padding=0, dilation=1)  where {M, L}

Dimensions for a "pooling" operation that can have an arbitrary input size, kernel size, stride, dilation, and channel count. Used to dispatch onto efficient implementations at compile-time.

source
NNlib.maxpool — Function
maxpool(x, k::NTuple{N, Integer}; pad=0, stride=k)

Perform max pool operation with window size k on input tensor x.

Arguments:

  • x and k: Expects ndim(x) ∈ 3:5, and always length(k) == ndim(x) - 2
  • pad: See pad_zeros for details.
  • stride: Either a tuple with the same length as k, or one integer for all directions. Default is k.
source
NNlib.meanpool — Function
meanpool(x, k::NTuple{N, Integer}; pad=0, stride=k)

Perform mean pool operation with window size k on input tensor x.

Arguments:

  • x and k: Expects ndim(x) ∈ 3:5, and alwayslength(k) == ndim(x) - 2`
  • pad: See pad_zeros for details.
  • stride: Either a tuple with the same length as k, or one integer for all directions. Default is k.
source
NNlib.lpnormpool — Function
lpnormpool(x, p::Real, k::NTuple{N, Integer}; pad=0, stride=k)

Perform Lp pool operation with value of the Lp norm p and window size k on input tensor x, also known as LPPool in pytorch. This pooling operator from Learned-Norm Pooling for Deep Feedforward and Recurrent Neural Networks.

Arguments:

  • x and k: Expects ndim(x) ∈ 3:5, and alwayslength(k) == ndim(x) - 2`
  • p is restricted to 0 < p < Inf.
  • pad: See pad_zeros for details.
  • stride: Either a tuple with the same length as k, or one integer for all directions. Default is k.

For all elements x in a size k window, lpnormpool computes (∑ᵢ xᵢ^p)^(1 / p) as an element of the output.

Thus lpnormpool(x, 1, k) ./ prod(k) ≈ meanpool(x, k) and lpnormpool(x, 2, k).^2 ./ prod(k) ≈ meanpool(x.^2, k).

source

Padding

NNlib.pad_reflect — Function
pad_reflect(x, pad::Tuple; [dims])
 pad_reflect(x, pad::Int; [dims])

Pad the array x reflecting its values across the border.

pad can a tuple of integers (l1, r1, ..., ln, rn) of some length 2n that specifies the left and right padding size for each of the dimensions in dims. If dims is not given, it defaults to the first n dimensions.

If pad is an integer, it is applied on both sides on every dimension in dims. In this case, dims defaults to the first ndims(x)-2 dimensions (i.e. excludes the channel and batch dimension).

See also pad_repeat, pad_symmetric, pad_circular, and pad_constant.

julia> r = reshape(1:9, 3, 3)
 3×3 reshape(::UnitRange{Int64}, 3, 3) with eltype Int64:
  1  4  7
@@ -453,7 +453,7 @@
  5  2  5  8  5  2
  6  3  6  9  6  3
  5  2  5  8  5  2
- 4  1  4  7  4  1
source
NNlib.pad_symmetric — Function
pad_symmetric(x, pad::Tuple; [dims])
+ 4  1  4  7  4  1
source
NNlib.pad_symmetric — Function
pad_symmetric(x, pad::Tuple; [dims])
 pad_symmetric(x, pad::Int; [dims])

Pad the array x reflecting its values symmetrically across the border, i.e. the border values of x are present in the padding values, in contrast to pad_reflect.

pad can a tuple of integers (l1, r1, ..., ln, rn) of some length 2n that specifies the left and right padding size for each of the dimensions in dims. If dims is not given, it defaults to the first n dimensions.

If pad is an integer, it is applied on both sides on every dimension in dims. In this case, dims defaults to the first ndims(x)-2 dimensions (i.e. excludes the channel and batch dimension).

See also pad_repeat, pad_reflect, pad_circular, and pad_constant.

julia> r = reshape(1:9, 3, 3)
 3×3 reshape(::UnitRange{Int64}, 3, 3) with eltype Int64:
  1  4  7
@@ -467,7 +467,7 @@
  2  2  5  8  8  5
  3  3  6  9  9  6
  3  3  6  9  9  6
- 2  2  5  8  8  5
source
NNlib.pad_circular — Function
pad_circular(x, pad::Tuple; [dims])
+ 2  2  5  8  8  5
source
NNlib.pad_circular — Function
pad_circular(x, pad::Tuple; [dims])
 pad_circular(x, pad::Int; [dims])

Pad the array x "circularly" across the border by wrapping around values from the opposite side of x.

pad can a tuple of integers (l1, r1, ..., ln, rn) of some length 2n that specifies the left and right padding size for each of the dimensions in dims. If dims is not given, it defaults to the first n dimensions.

If pad is an integer, it is applied on both sides on every dimension in dims. In this case, dims defaults to the first ndims(x)-2 dimensions (i.e. excludes the channel and batch dimension).

The pad length on either side in any dimension must not exceed the size of x in that dimension, i.e. pad_circular is not able to create abitrary sized tilings of x.

See also pad_repeat, pad_reflect, pad_symmetric, and pad_constant.

julia> r = reshape(1:9, 3, 3)
 3×3 reshape(::UnitRange{Int64}, 3, 3) with eltype Int64:
  1  4  7
@@ -481,7 +481,7 @@
  8  2  5  8  2  5
  9  3  6  9  3  6
  7  1  4  7  1  4
- 8  2  5  8  2  5
source
NNlib.pad_repeat — Function
pad_repeat(x, pad::Tuple; [dims])
+ 8  2  5  8  2  5
source
NNlib.pad_repeat — Function
pad_repeat(x, pad::Tuple; [dims])
 pad_repeat(x, pad::Int; [dims])

Pad the array x repeating the values on the border.

pad can a tuple of integers (l1, r1, ..., ln, rn) of some length 2n that specifies the left and right padding size for each of the dimensions in dims. If dims is not given, it defaults to the first n dimensions.

If pad is an integer, it is applied on both sides on every dimension in dims. In this case, dims defaults to the first ndims(x)-2 dimensions (i.e. excludes the channel and batch dimension).

See also pad_reflect, pad_symmetric, pad_circular, and pad_constant.

julia> r = reshape(1:9, 3, 3)
 3×3 reshape(::UnitRange{Int64}, 3, 3) with eltype Int64:
  1  4  7
@@ -495,7 +495,7 @@
  2  2  2  2  5  8  8  8  8  8
  3  3  3  3  6  9  9  9  9  9
  3  3  3  3  6  9  9  9  9  9
- 3  3  3  3  6  9  9  9  9  9
source
NNlib.pad_constant — Function
pad_constant(x, pad::Tuple, val = 0; [dims = :])
+ 3  3  3  3  6  9  9  9  9  9
source
NNlib.pad_constant — Function
pad_constant(x, pad::Tuple, val = 0; [dims = :])
 pad_constant(x, pad::Int, val = 0; [dims = :])

Pad the array x with the constant value val.

pad can be a tuple of integers. If it is of some length 2 * length(dims) that specifies the left and right padding size for each of the dimensions in dims as (l1, r1, ..., ln, rn). If supplied with a tuple of length length(dims) instead, it applies symmetric padding. If dims is not given, it defaults to all dimensions.

For integer pad input, it is applied on both sides on every dimension in dims.

See also pad_zeros, pad_repeat, pad_reflect, pad_symmetric, and pad_circular.

julia> r = reshape(1:4, 2, 2)
 2×2 reshape(::UnitRange{Int64}, 2, 2) with eltype Int64:
  1  3
@@ -562,8 +562,8 @@
 julia> pad_constant(r, (2,1, 3), dims = (1,2)) # padding must always be either the same length as dims, or double it
 ERROR: ArgumentError: Could not parse padding (2, 1, 3) and dims (1, 2)
 Stacktrace:
-[...]
source
NNlib.pad_zeros — Function
pad_zeros(x, pad::Tuple; [dims])
-pad_zeros(x, pad::Int; [dims])

Pad the array x with zeros. Equivalent to pad_constant with the constant equal to 0.

source

Convolution

Flux's Conv and CrossCor layers use NNlib.DenseConvDims and NNlib.conv internally.

NNlib.conv supports complex datatypes on CPU and CUDA devices.

!!! AMDGPU MIOpen supports only cross-correlation (flipkernel=true). Therefore for every regular convolution (flipkernel=false) kernel is flipped before calculation. For better performance, use cross-correlation (flipkernel=true) and manually flip the kernel before NNlib.conv call. Flux handles this automatically, this is only required for direct calls.

NNlib.conv — Function
conv(x, w; stride = 1, pad = 0, dilation = 1, flipped = false, groups = 1)

Apply convolution filter w to input x. x and w are 3d/4d/5d tensors in 1d/2d/3d convolutions respectively. x and w may have real or complex element types.

source
NNlib.ConvDims — Type
ConvDims

Type system-level information about convolution dimensions. Critical for things like im2col!() to generate efficient code, and helpful to reduce the number of kwargs getting passed around.

source
NNlib.depthwiseconv — Function
depthwiseconv(x, w; stride=1, pad=0, dilation=1, flipped=false)

Depthwise convolution operation with filter w on input x. x and w are 3d/4d/5d tensors in 1d/2d/3d convolutions respectively.

source
NNlib.DepthwiseConvDims — Type
DepthwiseConvDims

Concrete subclass of ConvDims for a depthwise convolution. Differs primarily due to characterization by Cin, Cmult, rather than Cin, Cout. Useful to be separate from DenseConvDims primarily for channel calculation differences.

source
NNlib.DenseConvDims — Type
DenseConvDims

Concrete subclass of ConvDims for a normal, dense, conv2d/conv3d.

source
NNlib.unfold — Function
unfold(x, kernel_size; stride = 1, pad = 0, dilation = 0, flipped = true)

Places sliding windows of x into a container tensor of size (num_windows, window_size, batchsize). The window size is determined by the prod(spatial dims of kernel)*input_channels. The number of sliding windows will match those of convolution (conv) with the same kernel_size and arguments. Note that by default conv flips the spatial dimensions of its kernel (default flipped=false), whereas unfold does not (default flipped=true). Uses NNlib.im2col! as backend.

See also fold, the adjoint/transpose operator and a potential inverse of unfold.

Example

The below example demonstrates that unfold uses the same sliding windows as conv. In general batched_mul + unfold should not be used to achieve convolution.

julia> x = reshape([100 2 3 40 5 6 700], 7, 1, 1);  # 1D data, 1 channel, batch of 1
+[...]
source
NNlib.pad_zeros — Function
pad_zeros(x, pad::Tuple; [dims])
+pad_zeros(x, pad::Int; [dims])

Pad the array x with zeros. Equivalent to pad_constant with the constant equal to 0.

source

Convolution

Flux's Conv and CrossCor layers use NNlib.DenseConvDims and NNlib.conv internally.

NNlib.conv supports complex datatypes on CPU and CUDA devices.

!!! AMDGPU MIOpen supports only cross-correlation (flipkernel=true). Therefore for every regular convolution (flipkernel=false) kernel is flipped before calculation. For better performance, use cross-correlation (flipkernel=true) and manually flip the kernel before NNlib.conv call. Flux handles this automatically, this is only required for direct calls.

NNlib.conv — Function
conv(x, w; stride = 1, pad = 0, dilation = 1, flipped = false, groups = 1)

Apply convolution filter w to input x. x and w are 3d/4d/5d tensors in 1d/2d/3d convolutions respectively. x and w may have real or complex element types.

source
NNlib.ConvDims — Type
ConvDims

Type system-level information about convolution dimensions. Critical for things like im2col!() to generate efficient code, and helpful to reduce the number of kwargs getting passed around.

source
NNlib.depthwiseconv — Function
depthwiseconv(x, w; stride=1, pad=0, dilation=1, flipped=false)

Depthwise convolution operation with filter w on input x. x and w are 3d/4d/5d tensors in 1d/2d/3d convolutions respectively.

source
NNlib.DepthwiseConvDims — Type
DepthwiseConvDims

Concrete subclass of ConvDims for a depthwise convolution. Differs primarily due to characterization by Cin, Cmult, rather than Cin, Cout. Useful to be separate from DenseConvDims primarily for channel calculation differences.

source
NNlib.DenseConvDims — Type
DenseConvDims

Concrete subclass of ConvDims for a normal, dense, conv2d/conv3d.

source
NNlib.unfold — Function
unfold(x, kernel_size; stride = 1, pad = 0, dilation = 0, flipped = true)

Places sliding windows of x into a container tensor of size (num_windows, window_size, batchsize). The window size is determined by the prod(spatial dims of kernel)*input_channels. The number of sliding windows will match those of convolution (conv) with the same kernel_size and arguments. Note that by default conv flips the spatial dimensions of its kernel (default flipped=false), whereas unfold does not (default flipped=true). Uses NNlib.im2col! as backend.

See also fold, the adjoint/transpose operator and a potential inverse of unfold.

Example

The below example demonstrates that unfold uses the same sliding windows as conv. In general batched_mul + unfold should not be used to achieve convolution.

julia> x = reshape([100 2 3 40 5 6 700], 7, 1, 1);  # 1D data, 1 channel, batch of 1
 
 julia> w = reshape([1 0 -1], 3, 1, 1);  # 1D conv kernel of length 3
 
@@ -591,7 +591,7 @@
   -2
  -38
   34
-   6
source
NNlib.fold — Function
fold(y, output_size, kernel_size; stride = 1, pad = 0, dilation = 0, flipped = true)

The adjoint/transpose operator of unfold. It accumulates sliding windows from the output of unfold into a container tensor of size output_size. An inverse to unfold may be obtained (in some cases) by using fold and accounting for scaling issues with a divisor (see example). Uses NNlib.col2im! as backend.

See also unfold.

Example

julia> x = reshape([100 2 3 40 5 6 700], 7, 1, 1);  # 1D data, 1 channel, batch of 1
+   6
source
NNlib.fold — Function
fold(y, output_size, kernel_size; stride = 1, pad = 0, dilation = 0, flipped = true)

The adjoint/transpose operator of unfold. It accumulates sliding windows from the output of unfold into a container tensor of size output_size. An inverse to unfold may be obtained (in some cases) by using fold and accounting for scaling issues with a divisor (see example). Uses NNlib.col2im! as backend.

See also unfold.

Example

julia> x = reshape([100 2 3 40 5 6 700], 7, 1, 1);  # 1D data, 1 channel, batch of 1
 
 julia> y = NNlib.unfold(x, (3,1,1))  # sliding window of size 3
 5×3×1 Array{Int64, 3}:
@@ -633,7 +633,7 @@
   40.0
    5.0
    6.0
- 700.0

In general, an inverse to unfold does not exist if divisor contains zeros.

source

Upsampling

Flux's Upsample layer uses NNlib.upsample_nearest, NNlib.upsample_bilinear, and NNlib.upsample_trilinear as its backend. Additionally, Flux's PixelShuffle layer uses NNlib.pixel_shuffle as its backend.

NNlib.upsample_nearest — Function
upsample_nearest(x, scale::NTuple{S,Int})
+ 700.0

In general, an inverse to unfold does not exist if divisor contains zeros.

source

Upsampling

Flux's Upsample layer uses NNlib.upsample_nearest, NNlib.upsample_bilinear, and NNlib.upsample_trilinear as its backend. Additionally, Flux's PixelShuffle layer uses NNlib.pixel_shuffle as its backend.

NNlib.upsample_nearest — Function
upsample_nearest(x, scale::NTuple{S,Int})
 upsample_nearest(x; size::NTuple{S,Int})

Upsamples the array x by integer multiples along the first S dimensions. Subsequent dimensions of x are not altered.

Either the scale factors or the final output size can be specified.

See also upsample_bilinear, for two dimensions of an N=4 array.

Example

julia> upsample_nearest([1 2 3; 4 5 6], (2, 3))
 4×9 Matrix{Int64}:
  1  1  1  2  2  2  3  3  3
@@ -652,8 +652,8 @@
  4  5  6
 
 julia> ans == upsample_nearest([1 2 3; 4 5 6], size=(4,))
-true
source
Missing docstring.

Missing docstring for ∇upsample_nearest. Check Documenter's build log for details.

NNlib.upsample_linear — Function
upsample_linear(x::AbstractArray{T,3}, scale::Real; align_corners::Bool = true)
-upsample_linear(x::AbstractArray{T,3}; size::Integer, align_corners::Bool = true)

Upsamples the first dimension of the array x by the upsample provided scale, using linear interpolation. As an alternative to using scale, the resulting array size can be directly specified with a keyword argument.

The size of the output is equal to (scale*S1, S2, S3), where S1, S2, S3 = size(x).

source
NNlib.∇upsample_linear — Function
∇upsample_linear(Δ::AbstractArray{T,3}; size::Integer, align_corners::Bool = true) where T

Arguments

  • Δ: Incoming gradient array, backpropagated from downstream layers
  • size: Size of the image upsampled in the first place

Outputs

  • dx: Downsampled version of Δ
source
NNlib.upsample_bilinear — Function
upsample_bilinear(x::AbstractArray{T,4}, scale::NTuple{2,Real}; align_corners::Bool = true)
+true
source
Missing docstring.

Missing docstring for ∇upsample_nearest. Check Documenter's build log for details.

NNlib.upsample_linear — Function
upsample_linear(x::AbstractArray{T,3}, scale::Real; align_corners::Bool = true)
+upsample_linear(x::AbstractArray{T,3}; size::Integer, align_corners::Bool = true)

Upsamples the first dimension of the array x by the upsample provided scale, using linear interpolation. As an alternative to using scale, the resulting array size can be directly specified with a keyword argument.

The size of the output is equal to (scale*S1, S2, S3), where S1, S2, S3 = size(x).

source
NNlib.∇upsample_linear — Function
∇upsample_linear(Δ::AbstractArray{T,3}; size::Integer, align_corners::Bool = true) where T

Arguments

  • Δ: Incoming gradient array, backpropagated from downstream layers
  • size: Size of the image upsampled in the first place

Outputs

  • dx: Downsampled version of Δ
source
NNlib.upsample_bilinear — Function
upsample_bilinear(x::AbstractArray{T,4}, scale::NTuple{2,Real}; align_corners::Bool = true)
 upsample_bilinear(x::AbstractArray{T,4}; size::NTuple{2,Integer}, align_corners::Bool = true)

Upsamples the first 2 dimensions of the array x by the upsample factors stored in scale, using bilinear interpolation. As an alternative to using scale, the resulting image size can be directly specified with a keyword argument.

The size of the output is equal to (scale[1]*S1, scale[2]*S2, S3, S4), where S1, S2, S3, S4 = size(x).

Examples

julia> x = reshape(Float32[1 2 3; 4 5 6], (2,3,1,1))
 2×3×1×1 Array{Float32, 4}:
 [:, :, 1, 1] =
@@ -678,10 +678,10 @@
  1.75  1.97222  2.19444  2.41667  2.63889     3.08333  3.30556  3.52778  3.75
  2.5   2.72222  2.94444  3.16667  3.38889     3.83333  4.05556  4.27778  4.5
  3.25  3.47222  3.69444  3.91667  4.13889     4.58333  4.80556  5.02778  5.25
- 4.0   4.22222  4.44444  4.66667  4.88889     5.33333  5.55556  5.77778  6.0
source
NNlib.∇upsample_bilinear — Function
∇upsample_bilinear(Δ::AbstractArray{T,4}; size::NTuple{2,Integer}, align_corners::Bool = true) where T

Arguments

  • Δ: Incoming gradient array, backpropagated from downstream layers
  • size: Lateral (W,H) size of the image upsampled in the first place

Outputs

  • dx: Downsampled version of Δ
source
NNlib.upsample_trilinear — Function
upsample_trilinear(x::AbstractArray{T,5}, scale::NTuple{3,Real}; align_corners::Bool = true)
+ 4.0   4.22222  4.44444  4.66667  4.88889     5.33333  5.55556  5.77778  6.0
source
NNlib.∇upsample_bilinear — Function
∇upsample_bilinear(Δ::AbstractArray{T,4}; size::NTuple{2,Integer}, align_corners::Bool = true) where T

Arguments

  • Δ: Incoming gradient array, backpropagated from downstream layers
  • size: Lateral (W,H) size of the image upsampled in the first place

Outputs

  • dx: Downsampled version of Δ
source
NNlib.upsample_trilinear — Function
upsample_trilinear(x::AbstractArray{T,5}, scale::NTuple{3,Real}; align_corners::Bool = true)
 upsample_trilinear(x::AbstractArray{T,5}; size::NTuple{3,Integer}, align_corners::Bool = true)

Upsamples the first 3 dimensions of the array x by the upsample factors stored in scale, using trilinear interpolation. As an alternative to using scale, the resulting image size can be directly specified with a keyword argument.

The size of the output is equal to (scale[1]*S1, scale[2]*S2, scale[3]*S3, S4, S5), where S1, S2, S3, S4, S5 = size(x).

Examples

upsample_trilinear(x, (2, 3, 4))
 upsample_trilinear(x; size=(4, 9, 11))  # specify ouput size instead
-upsample_trilinear(x, (2.5, 3.5, pi))  # non-integer scaling factors are allowed
source
NNlib.∇upsample_trilinear — Function
∇upsample_trilinear(Δ::AbstractArray{T,5}; size::NTuple{3,Integer}, align_corners::Bool = true) where T

Arguments

  • Δ: Incoming gradient array, backpropagated from downstream layers
  • size: Lateral size & depth (W,H,D) of the image upsampled in the first place

Outputs

  • dx: Downsampled version of Δ
source
NNlib.pixel_shuffle — Function
pixel_shuffle(x, r::Integer)

Pixel shuffling operation, upscaling by a factor r.

For 4-arrays representing N images, the operation converts input size(x) == (W, H, r^2*C, N) to output of size (r*W, r*H, C, N). For D-dimensional data, it expects ndims(x) == D+2 with channel and batch dimensions, and divides the number of channels by r^D.

Used in super-resolution networks to upsample towards high resolution features. Reference: Shi et. al., "Real-Time Single Image and Video Super-Resolution ...", CVPR 2016, https://arxiv.org/abs/1609.05158

Examples

julia> x = [10i + j + channel/10 for i in 1:2, j in 1:3, channel in 1:4, batch in 1:1]
+upsample_trilinear(x, (2.5, 3.5, pi))  # non-integer scaling factors are allowed
source
NNlib.∇upsample_trilinear — Function
∇upsample_trilinear(Δ::AbstractArray{T,5}; size::NTuple{3,Integer}, align_corners::Bool = true) where T

Arguments

  • Δ: Incoming gradient array, backpropagated from downstream layers
  • size: Lateral size & depth (W,H,D) of the image upsampled in the first place

Outputs

  • dx: Downsampled version of Δ
source
NNlib.pixel_shuffle — Function
pixel_shuffle(x, r::Integer)

Pixel shuffling operation, upscaling by a factor r.

For 4-arrays representing N images, the operation converts input size(x) == (W, H, r^2*C, N) to output of size (r*W, r*H, C, N). For D-dimensional data, it expects ndims(x) == D+2 with channel and batch dimensions, and divides the number of channels by r^D.

Used in super-resolution networks to upsample towards high resolution features. Reference: Shi et. al., "Real-Time Single Image and Video Super-Resolution ...", CVPR 2016, https://arxiv.org/abs/1609.05158

Examples

julia> x = [10i + j + channel/10 for i in 1:2, j in 1:3, channel in 1:4, batch in 1:1]
 2×3×4×1 Array{Float64, 4}:
 [:, :, 1, 1] =
  11.1  12.1  13.1
@@ -722,7 +722,7 @@
  2.1  2.3  2.5
  2.2  2.4  2.6
  3.1  3.3  3.5
- 3.2  3.4  3.6
source

Batched Operations

Flux's Bilinear layer uses NNlib.batched_mul internally.

NNlib.batched_mul — Function
batched_mul(A, B) -> C
+ 3.2  3.4  3.6
source

Batched Operations

Flux's Bilinear layer uses NNlib.batched_mul internally.

NNlib.batched_mul — Function
batched_mul(A, B) -> C
 A ⊠ B  # \boxtimes

Batched matrix multiplication. Result has C[:,:,k...] == A[:,:,k...] * B[:,:,k...] where k... represent any indices in the last dimensions.

If ndims(A) == ndims(B) == 3 and size(B,3) == 1 then instead C[:,:,k] == A[:,:,k] * B[:,:,1], and similarly for A.

To transpose each matrix, apply batched_transpose to the array, or batched_adjoint for conjugate-transpose:

julia> A, B = randn(2,5,17), randn(5,9,17);
 
 julia> A ⊠ B |> size
@@ -738,7 +738,7 @@
 (2, 9, 17)
 
 julia> batched_transpose(A) == PermutedDimsArray(A, (2,1,3))
-true

The equivalent PermutedDimsArray may be used in place of batched_transpose. Other permutations are also handled by BLAS, provided that the batch index k is not the first dimension of the underlying array. Thus PermutedDimsArray(::Array, (1,3,2)) and PermutedDimsArray(::Array, (3,1,2)) are fine.

However, A = PermutedDimsArray(::Array, (3,2,1)) is not acceptable to BLAS, since the batch dimension is the contiguous one: stride(A,3) == 1. This will be copied, as doing so is faster than batched_mul_generic!.

Both this copy and batched_mul_generic! produce @debug messages, and setting for instance ENV["JULIA_DEBUG"] = NNlib will display them.

source
batched_mul(A::Array{T,3}, B::Matrix)
+true

The equivalent PermutedDimsArray may be used in place of batched_transpose. Other permutations are also handled by BLAS, provided that the batch index k is not the first dimension of the underlying array. Thus PermutedDimsArray(::Array, (1,3,2)) and PermutedDimsArray(::Array, (3,1,2)) are fine.

However, A = PermutedDimsArray(::Array, (3,2,1)) is not acceptable to BLAS, since the batch dimension is the contiguous one: stride(A,3) == 1. This will be copied, as doing so is faster than batched_mul_generic!.

Both this copy and batched_mul_generic! produce @debug messages, and setting for instance ENV["JULIA_DEBUG"] = NNlib will display them.

source
batched_mul(A::Array{T,3}, B::Matrix)
 batched_mul(A::Matrix, B::Array{T,3})
 A ⊠ B

This is always matrix-matrix multiplication, but either A or B may lack a batch index.

  • When B is a matrix, result has C[:,:,k] == A[:,:,k] * B[:,:] for all k.

  • When A is a matrix, then C[:,:,k] == A[:,:] * B[:,:,k]. This can also be done by reshaping and calling *, for instance A ⊡ B using TensorCore.jl, but is implemented here using batched_gemm instead of gemm.

julia> randn(16,8,32) ⊠ randn(8,4) |> size
 (16, 4, 32)
@@ -747,19 +747,19 @@
 (16, 4, 32)
 
 julia> randn(16,8) ⊠ randn(8,4,32) |> size
-(16, 4, 32)

See also batched_vec to regard B as a batch of vectors, A[:,:,k] * B[:,k].

source
NNlib.batched_mul! — Function
batched_mul!(C, A, B) -> C
-batched_mul!(C, A, B, α=1, β=0)

In-place batched matrix multiplication, equivalent to mul!(C[:,:,k], A[:,:,k], B[:,:,k], α, β) for all k. If size(B,3) == 1 then every batch uses B[:,:,1] instead.

This will call batched_gemm! whenever possible. For real arrays this means that, for X ∈ [A,B,C], either strides(X,1)==1 or strides(X,2)==1, the latter may be caused by batched_transpose or by for instance PermutedDimsArray(::Array, (3,1,2)). Unlike batched_mul this will never make a copy.

For complex arrays, the wrapper made by batched_adjoint must be outermost to be seen. In this case the strided accepted by BLAS are more restricted, if stride(C,1)==1 then only stride(AorB::BatchedAdjoint,2) == 1 is accepted.

source
NNlib.batched_adjoint — Function
batched_transpose(A::AbstractArray{T,3})
+(16, 4, 32)

See also batched_vec to regard B as a batch of vectors, A[:,:,k] * B[:,k].

source
NNlib.batched_mul! — Function
batched_mul!(C, A, B) -> C
+batched_mul!(C, A, B, α=1, β=0)

In-place batched matrix multiplication, equivalent to mul!(C[:,:,k], A[:,:,k], B[:,:,k], α, β) for all k. If size(B,3) == 1 then every batch uses B[:,:,1] instead.

This will call batched_gemm! whenever possible. For real arrays this means that, for X ∈ [A,B,C], either strides(X,1)==1 or strides(X,2)==1, the latter may be caused by batched_transpose or by for instance PermutedDimsArray(::Array, (3,1,2)). Unlike batched_mul this will never make a copy.

For complex arrays, the wrapper made by batched_adjoint must be outermost to be seen. In this case the strided accepted by BLAS are more restricted, if stride(C,1)==1 then only stride(AorB::BatchedAdjoint,2) == 1 is accepted.

source
NNlib.batched_adjoint — Function
batched_transpose(A::AbstractArray{T,3})
 batched_adjoint(A)

Equivalent to applying transpose or adjoint to each matrix A[:,:,k].

These exist to control how batched_mul behaves, as it operates on such matrix slices of an array with ndims(A)==3.

PermutedDimsArray(A, (2,1,3)) is equivalent to batched_transpose(A), and is also understood by batched_mul (and more widely supported elsewhere).

BatchedTranspose{T, S} <: AbstractBatchedMatrix{T, 3}
-BatchedAdjoint{T, S}

Lazy wrappers analogous to Transpose and Adjoint, returned by batched_transpose etc.

source
NNlib.batched_transpose — Function
batched_transpose(A::AbstractArray{T,3})
+BatchedAdjoint{T, S}

Lazy wrappers analogous to Transpose and Adjoint, returned by batched_transpose etc.

source
NNlib.batched_transpose — Function
batched_transpose(A::AbstractArray{T,3})
 batched_adjoint(A)

Equivalent to applying transpose or adjoint to each matrix A[:,:,k].

These exist to control how batched_mul behaves, as it operates on such matrix slices of an array with ndims(A)==3.

PermutedDimsArray(A, (2,1,3)) is equivalent to batched_transpose(A), and is also understood by batched_mul (and more widely supported elsewhere).

BatchedTranspose{T, S} <: AbstractBatchedMatrix{T, 3}
-BatchedAdjoint{T, S}

Lazy wrappers analogous to Transpose and Adjoint, returned by batched_transpose etc.

source
NNlib.batched_vec — Function
batched_vec(A::Array{T,3}, B::Matrix)
+BatchedAdjoint{T, S}

Lazy wrappers analogous to Transpose and Adjoint, returned by batched_transpose etc.

source
NNlib.batched_vec — Function
batched_vec(A::Array{T,3}, B::Matrix)
 batched_vec(A::Array{T,3}, b::Vector)

Batched matrix-vector multiplication: the result has C[:,:,k] == A[:,:,k] * B[:,k] for all k, or else C[:,:,k] == A[:,:,k] * b for b::Vector.

With the same argument types, batched_mul(A, B) would regard B as a fixed matrix, not a batch of vectors. Both reshape and then call batched_mul(::Array{T,3}, ::Array{T,3}).

julia> A, B, b = randn(16,8,32), randn(8,32), randn(8);
 
 julia> batched_vec(A,B) |> size
 (16, 32)
 
 julia> batched_vec(A,b) |> size
-(16, 32)
source

Gather and Scatter

Flux's Embedding layer uses NNlib.gather as its backend.

NNlib.gather — Function
NNlib.gather(src, idx) -> dst

Reverse operation of scatter. Gathers data from source src and writes it in a destination dst according to the index array idx. For each k in CartesianIndices(idx), assign values to dst according to

dst[:, ... , k] .= src[:, ... , idx[k]...]

Notice that if idx is a vector containing integers and src is a matrix, previous expression simplifies to

dst[:, k] .= src[:, idx[k]]

and k will run over 1:length(idx).

The elements of idx can be integers or integer tuples and may be repeated. A single src column can end up being copied into zero, one, or multiple dst columns.

See gather! for an in-place version.

Examples

julia> NNlib.gather([1,20,300,4000], [2,4,2])
+(16, 32)
source

Gather and Scatter

Flux's Embedding layer uses NNlib.gather as its backend.

NNlib.gather — Function
NNlib.gather(src, idx) -> dst

Reverse operation of scatter. Gathers data from source src and writes it in a destination dst according to the index array idx. For each k in CartesianIndices(idx), assign values to dst according to

dst[:, ... , k] .= src[:, ... , idx[k]...]

Notice that if idx is a vector containing integers and src is a matrix, previous expression simplifies to

dst[:, k] .= src[:, idx[k]]

and k will run over 1:length(idx).

The elements of idx can be integers or integer tuples and may be repeated. A single src column can end up being copied into zero, one, or multiple dst columns.

See gather! for an in-place version.

Examples

julia> NNlib.gather([1,20,300,4000], [2,4,2])
 3-element Vector{Int64}:
    20
  4000
@@ -768,7 +768,7 @@
 julia> NNlib.gather([1 2 3; 4 5 6], [1,3,1,3,1])
 2×5 Matrix{Int64}:
  1  3  1  3  1
- 4  6  4  6  4
source
gather(src, IJK...)

Convert the tuple of integer vectors IJK to a tuple of CartesianIndex and call gather on it: gather(src, CartesianIndex.(IJK...)).

Examples

julia> src = reshape([1:15;], 3, 5)
+ 4  6  4  6  4
source
gather(src, IJK...)

Convert the tuple of integer vectors IJK to a tuple of CartesianIndex and call gather on it: gather(src, CartesianIndex.(IJK...)).

Examples

julia> src = reshape([1:15;], 3, 5)
 3×5 Matrix{Int64}:
  1  4  7  10  13
  2  5  8  11  14
@@ -777,7 +777,7 @@
 julia> NNlib.gather(src, [1, 2], [2, 4])
 2-element Vector{Int64}:
   4
- 11
source
NNlib.gather! — Function
NNlib.gather!(dst, src, idx)

Reverse operation of scatter!. Gathers data from source src and writes it in destination dst according to the index array idx. For each k in CartesianIndices(idx), assign values to dst according to

dst[:, ... , k] .= src[:, ... , idx[k]...]

Notice that if idx is a vector containing integers, and both dst and src are matrices, previous expression simplifies to

dst[:, k] .= src[:, idx[k]]

and k will run over 1:length(idx).

The elements of idx can be integers or integer tuples and may be repeated. A single src column can end up being copied into zero, one, or multiple dst columns.

See gather for an allocating version.

source
NNlib.scatter — Function
NNlib.scatter(op, src, idx; [init, dstsize])

Scatter operation allocating a destination array dst and calling scatter!(op, dst, src, idx) on it.

  • If keyword init is provided, it is used to initialize the content of dst. Otherwise, the init values is inferred from the reduction operator op for some common operators (e.g. init = 0 for op = +).

  • If dstsize is provided, it will be used to define the size of destination array, otherwise it will be inferred by src and idx.

See scatter! for full details on how idx works.

Examples

julia> NNlib.scatter(+, [10,100,1000], [3,1,2])
+ 11
source
NNlib.gather! — Function
NNlib.gather!(dst, src, idx)

Reverse operation of scatter!. Gathers data from source src and writes it in destination dst according to the index array idx. For each k in CartesianIndices(idx), assign values to dst according to

dst[:, ... , k] .= src[:, ... , idx[k]...]

Notice that if idx is a vector containing integers, and both dst and src are matrices, previous expression simplifies to

dst[:, k] .= src[:, idx[k]]

and k will run over 1:length(idx).

The elements of idx can be integers or integer tuples and may be repeated. A single src column can end up being copied into zero, one, or multiple dst columns.

See gather for an allocating version.

source
NNlib.scatter — Function
NNlib.scatter(op, src, idx; [init, dstsize])

Scatter operation allocating a destination array dst and calling scatter!(op, dst, src, idx) on it.

  • If keyword init is provided, it is used to initialize the content of dst. Otherwise, the init values is inferred from the reduction operator op for some common operators (e.g. init = 0 for op = +).

  • If dstsize is provided, it will be used to define the size of destination array, otherwise it will be inferred by src and idx.

See scatter! for full details on how idx works.

Examples

julia> NNlib.scatter(+, [10,100,1000], [3,1,2])
 3-element Vector{Int64}:
   100
  1000
@@ -795,7 +795,7 @@
     10
   2000
     10
-    10
source
NNlib.scatter! — Function
NNlib.scatter!(op, dst, src, idx)

Scatter operation, which writes data in src into dst at locations idx. A binary reduction operator op is applied during the scatter. For each index k in idx, accumulates values in dst according to

dst[:, ..., idx[k]...] = (op).(dst[:, ..., idx[k]...], src[:, ..., k...])

See also scatter, gather.

Arguments

  • op: Operations to be applied on dst and src, e.g. +, -, *, /, max, min and mean.
  • dst: The destination for src to aggregate to. This argument will be mutated.
  • src: The source data for aggregating.
  • idx: The mapping for aggregation from source (index) to destination (value). The idx array can contain either integers or tuples.

Examples

julia> NNlib.scatter!(+, ones(3), [10,100], [1,3])
+    10
source
NNlib.scatter! — Function
NNlib.scatter!(op, dst, src, idx)

Scatter operation, which writes data in src into dst at locations idx. A binary reduction operator op is applied during the scatter. For each index k in idx, accumulates values in dst according to

dst[:, ..., idx[k]...] = (op).(dst[:, ..., idx[k]...], src[:, ..., k...])

See also scatter, gather.

Arguments

  • op: Operations to be applied on dst and src, e.g. +, -, *, /, max, min and mean.
  • dst: The destination for src to aggregate to. This argument will be mutated.
  • src: The source data for aggregating.
  • idx: The mapping for aggregation from source (index) to destination (value). The idx array can contain either integers or tuples.

Examples

julia> NNlib.scatter!(+, ones(3), [10,100], [1,3])
 3-element Vector{Float64}:
   11.0
    1.0
@@ -804,7 +804,7 @@
 julia> NNlib.scatter!(*, fill(0.5, 2, 4), [1 10; 100 1000], [3,2])
 2×4 Matrix{Float64}:
  0.5    5.0   0.5  0.5
- 0.5  500.0  50.0  0.5
source

Sampling

NNlib.grid_sample — Function
grid_sample(input::AbstractArray{T, 4}, grid::AbstractArray{T, 4}; padding_mode = :zeros)

Given input, compute output by sampling input values at pixel locations from grid. Uses bilinear interpolation to calculate output values.

This implementation assumes the extrema (-1 and 1) are considered as referring to the center points of the input’s corner pixels (i.e. align corners is true).

Arguments

  • input: Input array in (W_in, H_in, C, N) shape.

  • grid: Input grid in (2, W_out, H_out, N) shape. Where for each (W_out, H_out, N) grid contains (x, y) coordinates that specify sampling locations normalized by the input shape.

    Therefore, x and y should have values in [-1, 1] range. For example, (x = -1, y = -1) is the left-top pixel of input, and (x = 1, y = 1) is the right-bottom pixel of input.

    Out-of-bound values are handled according to the padding_mode.

  • padding_mode: Out-of-bound padding. :zeros to use 0 for out-of-bound grid locations. :border to use border values for out-of-bound grid locations. Default is :zeros.

Returns

(W_out, H_out, C, N) sampled grid from input.

Examples

In the example below, grid contains two out-of-bound sampling locations, which are handled differently, depending on the padding_mode.

julia> x = reshape(collect(1.0:4.0), (2, 2, 1, 1))
+ 0.5  500.0  50.0  0.5
source

Sampling

NNlib.grid_sample — Function
grid_sample(input::AbstractArray{T, 4}, grid::AbstractArray{T, 4}; padding_mode = :zeros)

Given input, compute output by sampling input values at pixel locations from grid. Uses bilinear interpolation to calculate output values.

This implementation assumes the extrema (-1 and 1) are considered as referring to the center points of the input’s corner pixels (i.e. align corners is true).

Arguments

  • input: Input array in (W_in, H_in, C, N) shape.

  • grid: Input grid in (2, W_out, H_out, N) shape. Where for each (W_out, H_out, N) grid contains (x, y) coordinates that specify sampling locations normalized by the input shape.

    Therefore, x and y should have values in [-1, 1] range. For example, (x = -1, y = -1) is the left-top pixel of input, and (x = 1, y = 1) is the right-bottom pixel of input.

    Out-of-bound values are handled according to the padding_mode.

  • padding_mode: Out-of-bound padding. :zeros to use 0 for out-of-bound grid locations. :border to use border values for out-of-bound grid locations. Default is :zeros.

Returns

(W_out, H_out, C, N) sampled grid from input.

Examples

In the example below, grid contains two out-of-bound sampling locations, which are handled differently, depending on the padding_mode.

julia> x = reshape(collect(1.0:4.0), (2, 2, 1, 1))
 2×2×1×1 Array{Float64, 4}:
 [:, :, 1, 1] =
  1.0  3.0
@@ -836,7 +836,7 @@
 [:, :, 1, 1] =
  1.0  3.0
  1.5  3.5
- 2.0  4.0
source
NNlib.∇grid_sample — Function
∇grid_sample(Δ::AbstractArray{T, 4}, input::AbstractArray{T, 4}, grid::AbstractArray{T, 4}; padding_mode = :zeros) where T

Arguments

  • Δ: Input gradient in (W_out, H_out, C, N) shape (same as output of the primal computation).
  • input: Input from primal computation in (W_in, H_in, C, N) shape.
  • grid: Grid from primal computation in (2, W_out, H_out, N) shape.
  • padding_mode: Out-of-bound padding. :zeros to use 0 for out-of-bound grid locations. :border to use border values for out-of-bound grid locations. Should be the same as in primal computation. Default is :zeros.

Returns

dinput (same shape as input) and dgrid (same shape as grid) gradients.

source

Losses

NNlib.ctc_loss — Function
ctc_loss(ŷ, y)

Computes the connectionist temporal classification loss between ŷ and y. ŷ must be a classes-by-time matrices, i.e., each row represents a class and each column represents a time step. Additionally, the logsoftmax function will be applied to ŷ, so ŷ must be the raw activation values from the neural network and not, for example, the activations after being passed through a softmax activation function. y must be a 1D array of the labels associated with ŷ. The blank label is assumed to be the last label category in ŷ, so it is equivalent to size(ŷ, 1). Used for sequence-to-sequence classification problems such as speech recognition and handwriting recognition where the exact time-alignment of the output (e.g., letters) is not needed to solve the problem. See Graves et al. (2006) or Graves (2012) for mathematical details.

source

Miscellaneous

NNlib.logsumexp — Function
logsumexp(x; dims = :)

Computes log.(sum(exp.(x); dims)) in a numerically stable way. Without dims keyword this returns a scalar.

See also logsoftmax.

source
NNlib.glu — Function
glu(x, dim = 1)

The gated linear unit from the "Language Modeling with Gated Convolutional Networks" paper.

Calculates a .* sigmoid(b), where x is split in half along given dimension dim to form a and b.

source
NNlib.within_gradient — Function
within_gradient(x) --> Bool

Returns false except when used inside a gradient call, when it returns true. Useful for Flux regularisation layers which behave differently during training and inference.

This should work with any ChainRules-based differentiation package, in which case x is ignored. But Tracker.jl overloads with_gradient(x::TrackedArray), thus for widest use you should pass it an array whose gradient is of interest. There is also an overload for ForwardDiff.jl's Dual types (and arrays of them).

Examples

julia> using ForwardDiff, Zygote, NNlib
+ 2.0  4.0
source
NNlib.∇grid_sample — Function
∇grid_sample(Δ::AbstractArray{T, 4}, input::AbstractArray{T, 4}, grid::AbstractArray{T, 4}; padding_mode = :zeros) where T

Arguments

  • Δ: Input gradient in (W_out, H_out, C, N) shape (same as output of the primal computation).
  • input: Input from primal computation in (W_in, H_in, C, N) shape.
  • grid: Grid from primal computation in (2, W_out, H_out, N) shape.
  • padding_mode: Out-of-bound padding. :zeros to use 0 for out-of-bound grid locations. :border to use border values for out-of-bound grid locations. Should be the same as in primal computation. Default is :zeros.

Returns

dinput (same shape as input) and dgrid (same shape as grid) gradients.

source

Losses

NNlib.ctc_loss — Function
ctc_loss(ŷ, y)

Computes the connectionist temporal classification loss between ŷ and y. ŷ must be a classes-by-time matrices, i.e., each row represents a class and each column represents a time step. Additionally, the logsoftmax function will be applied to ŷ, so ŷ must be the raw activation values from the neural network and not, for example, the activations after being passed through a softmax activation function. y must be a 1D array of the labels associated with ŷ. The blank label is assumed to be the last label category in ŷ, so it is equivalent to size(ŷ, 1). Used for sequence-to-sequence classification problems such as speech recognition and handwriting recognition where the exact time-alignment of the output (e.g., letters) is not needed to solve the problem. See Graves et al. (2006) or Graves (2012) for mathematical details.

source

Miscellaneous

NNlib.logsumexp — Function
logsumexp(x; dims = :)

Computes log.(sum(exp.(x); dims)) in a numerically stable way. Without dims keyword this returns a scalar.

See also logsoftmax.

source
NNlib.glu — Function
glu(x, dim = 1)

The gated linear unit from the "Language Modeling with Gated Convolutional Networks" paper.

Calculates a .* sigmoid(b), where x is split in half along given dimension dim to form a and b.

source
NNlib.within_gradient — Function
within_gradient(x) --> Bool

Returns false except when used inside a gradient call, when it returns true. Useful for Flux regularisation layers which behave differently during training and inference.

This should work with any ChainRules-based differentiation package, in which case x is ignored. But Tracker.jl overloads with_gradient(x::TrackedArray), thus for widest use you should pass it an array whose gradient is of interest. There is also an overload for ForwardDiff.jl's Dual types (and arrays of them).

Examples

julia> using ForwardDiff, Zygote, NNlib
 
 julia> f_good(x) = if NNlib.within_gradient(x)
                      @show 10x
@@ -863,4 +863,4 @@
 
 julia> ForwardDiff.derivative(x -> f_bad(x, 3.0), 2.0)
 x * y = Dual{ForwardDiff.Tag{var"#9#10", Float64}}(6.0,3.0)
-3.0

What goes wrong in f_bad is that Zygote knows any to be non-differentiable, and thus completely ignores its contents. This is not a perfect mechanism, and the only style recommended is precisely that of f_good above.

source
NNlib.bias_act! — Function
bias_act!(σ, x, b)

This is equivalent to x .= σ.(x .+ b), also replacing sigmoid & tanh with sigmoid_fast & tanh_fast. It will only overwrite x when x isa StridedArray{<:AbstractFloat}.

When used within a gradient, it will overwrite only when σ has a method of derivatives_given_output which does not need the input at all. Such methods are defined by e.g. @scalar_rule relu(x) Ω > 0 where the derivative contains only Ω (the output) not x.

Warning

This is not safe to use if x is still needed for the gradient of some other function. Incorrect use will give silently wrong answers. It is intended mainly for Flux layers, in which the previous operation is known to be safe, e.g. bias_act!(σ, weight * input, bias) for a Dense layer.

source
+3.0

What goes wrong in f_bad is that Zygote knows any to be non-differentiable, and thus completely ignores its contents. This is not a perfect mechanism, and the only style recommended is precisely that of f_good above.

source
NNlib.bias_act! — Function
bias_act!(σ, x, b)

This is equivalent to x .= σ.(x .+ b), also replacing sigmoid & tanh with sigmoid_fast & tanh_fast. It will only overwrite x when x isa StridedArray{<:AbstractFloat}.

When used within a gradient, it will overwrite only when σ has a method of derivatives_given_output which does not need the input at all. Such methods are defined by e.g. @scalar_rule relu(x) Ω > 0 where the derivative contains only Ω (the output) not x.

Warning

This is not safe to use if x is still needed for the gradient of some other function. Incorrect use will give silently wrong answers. It is intended mainly for Flux layers, in which the previous operation is known to be safe, e.g. bias_act!(σ, weight * input, bias) for a Dense layer.

source
diff --git a/dev/search/index.html b/dev/search/index.html index fd1ec33da..dcf8716cf 100644 --- a/dev/search/index.html +++ b/dev/search/index.html @@ -1,2 +1,2 @@ -Search · NNlib.jl

Loading search...

    +Search · NNlib.jl

    Loading search...