Copyright © 2017-2018 Nolan Eakins
THIS IS NOT PRODUCTION QUALITY. USE AT YOUR OWN RISK.
A neural network library implemented in Ruby with a CUDA backend.
- NVIDIA's CUDA Toolkit
- Any one of the following C/C++ environments:
- Ruby: install via your distribution package manager or MSYS2's
pacman -S ruby - Bundler
- Cairo
$ gem install CooCoo
First the required dependencies need to be installed. The dependencies include the CUDA compiler and RubyGems.
Once the CUDA toolkit is installed, make sure it and Visual Studio's C++ compiler are in your path. Under MSYS2, use something like:
$ export PATH=/c/Program\ Files\ \(x86\)/Microsoft\ Visual\ Studio/2017/Community/VC/Tools/MSVC/14.10.25017/bin/HostX64/x64:$PATH
$ export PATH=/c/Program\ Files/NVIDIA\ GPU\ Computing\ Toolkit/CUDA/v9.0/bin:$PATH
RubyGems are installed with Bundler: bundle install
And the extension is built with: rake compile
Then to run an example: bundle exec ruby -Ilib -Iexamples examples/seeds.rb
Or IRB: bundle exec irb -Ilib
Or IRB: bundle exec ruby -Ilib
require 'coo-coo'
network = CooCoo::Network.new()
# create the layers
network.layer(CooCoo::Layer.new(28 * 28, 100))
network.layer(CooCoo::Layer.new(100, 10))
# learn
training_data = [ [expected_output, input_data_array ], ...]
trainer = CooCoo::Trainer::Stochastic.new
trainer.train(network: network,
data: training_data,
learning_rate: learning_rate,
batch_size: batch_size) do |stats|
# called every batch_size
puts("Batch #{batch} took #{stats.total_time} seconds with an average loss of #{stats.average_loss}.")
end
# store to disk
network.save("my_first_network.coo-coo_model")
# load from disk
loaded_net = CooCoo::Network.load!("my_first_network.coo-coo_model")
# predict
output = loaded_network.predict([ 0, 0, 0, ... ])
# => [ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ]All examples use OptParse, so refer to the output of running the example with --help. With no arguments, typically a network will be generated, trained, and tested without saving.
To run an example: bundle exec ruby -Ilib -Iexamples examples/EXAMPLE.rb --help
A recursive network that learns byte sequences.
Inspired by the IBAFSIP, this uses the UCI Machine Learning Repository's wheat seed dataset to predict the type of seed given seven parameters.
The MNIST Database is used to train a 10 digit classifier.
Uses a network trained with the MNIST Classifier to generate valid inputs via backpropagation.
Loosely based on Implement Backpropagation Algorithm From Scratch in Python
And more than a few of Siraj Raval's videos
Cross entropy & Softmax sorted out with DeepNotes.io