ymtoo / NNUtils.jl

NNUtils

This package provides neural net utilities:

• standard networks
  • mobilenets
  • resnet
• neural network blocks
  • DepthwiseSeparableConv
  • BottleneckResidual
  • ResnetResidualv1
  • ResnetResidualv2
• neural network layers
  • Sinc Conv
• saliency maps
  • Gradient
  • SmoothGradient

based on Flux.jl.

Installation

using Pkg; pkg"add https://github.com/ymtoo/NNUtils.jl.git"

Usage

SincConv

julia> using Flux, NNUtils

julia> fs = 9600f0
9600.0f0

julia> model = SincConv(fs, (200, 1), 1=>8, identity)
SincConv(9600.0, (200, 1), 1=>8)

julia> params(model) |> length
2

julia> params(model)[1] |> size
(1, 8)

julia> params(model)[2] |> size
(1, 8)

julia> x = randn(Float32, 4800, 1, 1, 16)
4800×1×1×16 Array{Float32, 4}:
[:, :, 1, 1] =
 1.3832613
 0.42255098
 ⋮
 0.3134887

[:, :, 1, 2] =
 2.0712132
 0.13467419
 ⋮

 julia> model(x)
 4601×1×8×16 Array{Float32, 4}:
 [:, :, 1, 1] =
   -3498.0605
  -17983.041
   20055.25
   -1296.112
       ⋮
    1844.9044
   10672.3125
  -12256.136
 
 [:, :, 2, 1] =
    521.0923
   8794.186
   4179.6655
  -4084.1067
      ⋮