A Simple Tensorflow Classification Model In Ruby

text

1class Operation
2  attr_accessor  :input_nodes, :output, :input
3
4  def initialize(*input_nodes)
5    @input_nodes = input_nodes
6  end
7
8  def compute
9    raise "Should be implemented"
10  end
11
12end

The Operation class forms the basis of all the computation in tensorflow. The compute method carries out the computation required. Let us create few classes which inherit the Operation class and we will use sigmoid as the activation function for the model

text

1require Matrix
2class Add < Operation
3  def initialize(*nodes)
4    @input_nodes=nodes
5  end
6
7  def compute(x,y)
8    return x.element(0,0) + y
9  end
10end
11
12class Matmul < Operation
13  def initialize(*nodes)
14    @input_nodes=nodes
15  end
16
17  def compute(x,y)
18    return Matrix[x] * Matrix[y].transpose
19  end
20end
21
22class Sigmoid < Operation
23  def initialize(*nodes)
24    @input_nodes=nodes
25  end
26
27  def compute(z)
28    1/(1+Math.exp(-z))
29  end
30end

Load the Matrix library at the begining for matrix computation. Now that we have created basic operation classes we will proceed to other components.

text

1class Placeholder
2  attr_accessor  :value, :output
3end
4
5class Variable
6  attr_accessor  :value, :output
7
8  def initialize(value)
9    @value = value
10  end
11end

The Placeholder class holds objects which are input or output of the model. These object act as constant and their values dont change during the session. The Variable holds objects whose values are either weights or bias given to the model. In the actual tensorflow model these values are updated using the optimizer(eg AdamOptimizer, Stocastic Gradient Descent etc). We wont be updating any value in our approach. We are fixing the value for weight and bias as [1,1] and -5 respectively. Now we will define the session class.

text

1class Session
2
3  def run(operation,feed_dict={})
4    nodes = traverse_postorder(operation)
5    nodes.each do |node|
6      if node.is_a? Placeholder
7        node.output = feed_dict[node.object_id]
8      elsif node.is_a? Variable
9        node.output = node.value
10      else
11        node.input = node.input_nodes.map { |input| input.output }
12        node.output = node.compute(*node.input)
13      end
14    end
15    operation.output
16  end
17  
18  def traverse_postorder(operation)
19    node_operators = []
20    recurse(operation,node_operators)
21  end
22
23  def recurse(node,node_operators)
24    if node.is_a?(Operation)
25      node.input_nodes.each do |input_node|
26        recurse(input_node,node_operators)
27      end
28    end
29    node_operators << node
30  end
31  
32end

The Session class takes care of all the executions. It first converts the set of operations to postfix order. The operations are then executed one by one. Since we have created all the components required, lets jump into action.

text

1W = Variable.new([1,1]) #Weight
2b = Variable.new(-5) #bias
3x = Placeholder.new() #input
4y = Matmul.new(W,x)
5z = Add.new(y,b) 
6a = Sigmoid.new(z) #activation function
7sess = Session.new()

Lets see the prediction for two extreme points

text

result = sess.run(a,feed_dict={x.object_id=>[0,-10]})
puts(result)%

text

3.059022269256247e-07

The point (0,-10) lies towards the lower left corner of the graph hence the sigmoid value is closer to 0

text

result = sess.run(a,feed_dict={x.object_id=>[8,10]})
puts(result)%

text

0.999997739675702

For the point (8,10) which lies near the upper right corner of the graph the sigmoid value is closer to 1.

Hope this example gives a good understanding of basic components of tensorflow in ruby. Happy learning !!!

Article

A Simple Tensorflow Classification Model In Ruby

Continue Reading