A simple Autograd engine written in Crystal.

Usage

Add Gradnite to your shard.yml and run shards install:

dependencies:
  gradnite:
    github: lucianbuzzo/gradnite
    version: 0.1.0

You can now require the module and use it's classes:

# require the gradnite module
require "gradnite"

# create a new Node
node = Gradnite::Node.new(1.0)

# add two nodes together
node = Gradnite::Node.new(1.0) + Gradnite::Node.new(2.0)

# multiply two nodes together
node = Gradnite::Node.new(1.0) * Gradnite::Node.new(2.0)

# divide two nodes together
node = Gradnite::Node.new(1.0) / Gradnite::Node.new(2.0)

# create a Neuron with 4 inputes
neuron = Gradnite::Neuron.new(4)

# create an MLP with 2 inputs, 2 hidden layers of 3 neurons each and 1 output
mlp = Gradnite::MLP.new(2, [3, 3, 1])

# run a forward pass on the mlp
mlp.forward([1.0, 2.0])

# run back propagation on the mlp
mlp.backward

# update the weights of the mlp
mlp.parameters.each { |p|
    p.value += -0.1 * p.grad
}

Gradnite also includes a utility for visualizing the computation graph of a neural network using graphviz. To use it, make sure you have graphviz installed locally and call the draw_dot function with a node:

a = Gradnite::Node.new value: 1.0, label: "a"
b = Gradnite::Node.new value: 2.0, label: "b"
c = a * b
c.label = "c"

Gradnite::draw_dot(c, "examples/multiply.dot")

You can now generate a png image of the computation graph using the dot command:

dot -Tpng examples/multiply.dot -o examples/multiply.png

For more examples of usage see the examples directory.

Development

Install prerequisites:

brew install crystal watchexec graphviz

Run in watch mode:

./watch.sh examples/binary_classifier.cr