Skip to content

karitra/jannet

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

34 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Jannet ANN's playground

One can watch three things forever: running water, burning fire and neural network learning (C)

Basic (yet artificial, not ready for natural one) neural network implementation in Julia language. Written for studying purposes.

Usage

In order to use module directly from sources root, include path to Jannet module/src folder into LOAD_PATH list, e.g.:

julia> push!(LOAD_PATH, "./src" );

Sorry for following declaration, I'm militant orthodox feminist at a heart, but thats Julia convention:

julia> using Jannet
...
julia> # in ann would be your brand new network
julia> ann = Jannet.WeaveNetwork(Float32, [3 10 2], learningRate=pi, momentum=0.2 );

First argument is a layout of network in form of

[ <input size> <layer 1 size>... <layer K size> <output size> ]

where size corresponds to a number of nodes on the layer, layout organized from left to right, from input layer to the output. Type parameter of template could be any Real type.

Train one sample

julia> x = Float32[ 0.5, 0.1, 0 ]
julia> y = Float32[ 0, 1 ]
julia> learnOnePattern!( ann, x, y )

First vector is input (w/o first bias activation element), second is a desired output pattern vector

Get the response

julia> p = sampleOnce(ann, x)
2-element Array{Float32,1}:
 0.15596 
 0.842187

In p would be the result for x pattern. Jannet.sampleOnce! version exist.

RPROP

TODO

Parallel RPROP

Tests

Function approximation

Note: path to Jannet module should be in LOAD_PATH list

Sample training for following function approximation:

train function

julia> include("tests/tests.jl")
...
julia> @time nn = Tests.t3(Float64, iters=1000000, lr=5, layout=[1 5 7 1], m = 0.05, epsilon=1e-5);
...
iter(604) 0.022774 sec   tr_err 0.00001930
iter(605) 0.022790 sec   tr_err 0.00000953
break out earlier on 605 iteration
train_error = 9.52945782156486e-6
test_error = 9.971634658467265e-6
 14.698823 seconds (52.59 M allocations: 2.060 GB, 4.00% gc time)

iters - count of iterations, can break out of the loop earlier on tr_err <= epsilon, where tr_err is average squared error for training set.

Learning results of trained network can be visualized (checked) as follow:

julia> using Gadfly
...
julia> y = [ first( Jannet.sampleOnce(nn, Float32[x]) ) for x in 0:0.02:1 ];
julia> ysample= Jannet.ftest(0:0.124:1* 2pi);
...
julia> draw( PNG("assets/sample.png", 22cm,12cm), plot( layer(y=ysample, Geom.line), layer(y=y, Geom.point, Theme(default_color=colorant"green")), layer(y=(y-ysample).^2*100, Geom.bar, Theme(default_color=colorant"dark red") ) ) )

Squared error rate for sample is shown in red color bars (scaled by 100), sample results are in green dots, and blue line as function itself:

sample plot

Kaggle tutorial

TODO

Link

Sample ImageMagick script to prepare tutorial images (planning to do it with help of Julia in upcoming revision):

for fn in {1..6283}.Bmp; do 
   convert $fn -colorspace gray -bias 20% -define convolve:scale='7!' -morphology Convolve Laplacian:5 \
    -auto-level -threshold 40% "${fn%.*}".sample.Bmp
done

References

  • Krose B., Smagt. P. An introduction to neural networks. 1996
  • ReidMiller M. RPROP - Description and Implementation Details. Technical Report. 1994

About

Simple, yet playful, artificial neural network implementation for Julia language

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published