Module: FastForward

Defined in:
lib/fast_forward.rb,
lib/fast_forward/nn.rb,
lib/fast_forward/version.rb

Defined Under Namespace

Classes: NN

Constant Summary collapse

DEFAULT_TOL =

Default tolerance for integrity check

1e-8
SUPPORTED_ACTIVATIONS =

Supported activation functions

["identity", "relu", "sigmoid", "tanh", "softmax"].freeze
VERSION = 
"0.1.0"

Class Method Summary collapse

Class Method Details

.load(data, tol: DEFAULT_TOL, exception_if_fail: true) ⇒ Object

Note:

data is supposed to have the following structure: data = { weights: [weights_array_0, weights_array_1 , …], biases: [biases_array_0, biases_array_1 , …], activations: [activation_fct_0, activation_fct_1, …], samples: (optional) { X: sample_inputs_array, y: sample_outputs_array } }

Loads neural network data into a NN object

Parameters:

  • data (Hash) —

    A hash containing all weights, biases and activation functions

  • tol (Numeric) —

    (Default: 1e-8) The tolerance to check integrity of the imported model from the provided samples

  • exception_if_fail (true, false) —

    (Default: true) If true, raises an exception if integrity check doesn't meet the required tolerance

Returns:

  • The loaded neural network



35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
 
# File 'lib/fast_forward.rb', line 35

def self.load(data, tol: DEFAULT_TOL, exception_if_fail: true)

	# check inputs
	if data[:weights].nil? || data[:weights].count == 0
		raise ArgumentError, "Missing weights data"
	elsif data[:biases].nil? || data[:biases].count == 0
		raise ArgumentError, "Missing biases data"
	elsif data[:activations].nil? || data[:activations].count == 0
		raise ArgumentError, "Missing activation functions"
	elsif data[:weights].size != data[:biases].size
		raise ArgumentError, "Array of weights and biases must have the same size"
	elsif data[:weights].size != data[:activations].size
		raise ArgumentError, "Array of weights and activations must have the same size"
	elsif !data[:activations].map{ |act| SUPPORTED_ACTIVATIONS.include?(act) }.all?
		raise ArgumentError, "Some activation functions are not supported"
	end


	layer_sizes = data[:weights].map(&:count) + [data[:weights].last.first.count]

	nn = NN.new(layer_sizes, data[:activations], data[:weights], data[:biases])

	if !data[:samples].nil? && tol >= 0
		FastForward.check_model_integrity(nn, data[:samples], tol: tol, exception_if_fail: exception_if_fail, verbose: true)
	end

	return nn
end

.load_json(data, tol: DEFAULT_TOL, exception_if_fail: true) ⇒ Object

Note:

data is supposed to have the following structure: data = { “weights”: { “0”: weights_array_0, “1”: weights_array_1, … }, “biases”: “0”: biases_array_0, “1”: biases_array_1, … }, “activations”: “0”: activation_fct_0, “1”: activation_fct_1, … }, “samples”: (optional) { “X”: sample_inputs_array, “y”: sample_outputs_array } }

Loads neural network from JSON data/string into a NN object

Parameters:

  • data (Hash, String) —

    A JSON string/object containing all weights, biases and activation functions

  • tol (Numeric) —

    (Default: 1e-8) The tolerance to check integrity of the imported model from the provided samples

  • exception_if_fail (true, false) —

    (Default: true) If true, raises an exception if integrity check doesn't meet the required tolerance

Returns:

  • The loaded neural network



94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
 
# File 'lib/fast_forward.rb', line 94

def self.load_json(data, tol: DEFAULT_TOL, exception_if_fail: true)

	data = JSON.parse(data) if data.is_a?(String)

	if data["weights"].keys != data["biases"].keys
		raise ArgumentError, "Incoherent weights and biases data"
	end
	n_layers = data["weights"].count

	# data["weights"] & data["biases"] are dictionaries of `"idx": array`
	d = {
		weights: n_layers.times.map{ |i| data["weights"][i.to_s] },
		biases: n_layers.times.map{ |i| data["biases"][i.to_s] },
		activations: n_layers.times.map{ |i| FastForward.rename_activation(data["activations"][i.to_s]) },
		samples: FastForward.load_sample_data(data)
	}

	return FastForward.load(d, tol: tol, exception_if_fail: exception_if_fail)
end

.load_pmml(data, sample_data: nil, tol: DEFAULT_TOL, exception_if_fail: true) ⇒ Object

Note:

sample_data is supposed to have the following structure: sample_data = { “samples”: { “X”: sample_inputs_array, “y”: sample_outputs_array } }

Loads neural network from PMML file into a NN object

Parameters:

  • data (String) —

    PMML file content

  • sample_data (Hash, String) —

    (Default: nil) A JSON string/object containing sample data to check the integrity of the loaded model

  • tol (Numeric) —

    (Default: 1e-8) The tolerance to check integrity of the imported model from the provided samples

  • exception_if_fail (true, false) —

    (Default: true) If true, raises an exception if integrity check doesn't meet the required tolerance

Returns:

  • The loaded neural network



130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
 
# File 'lib/fast_forward.rb', line 130

def self.load_pmml(data, sample_data: nil, tol: DEFAULT_TOL, exception_if_fail: true)
	# Load from PMML format: http://dmg.org/pmml/v4-3/NeuralNetwork.html
	# Note: does not support input transforms like NormContinuous
	# TODO: use element ids to check order
	data = Nokogiri::XML(data) if data.is_a?(String)
	
	all_weights = []
	all_biases = []
	all_activations = []
	
	default_act = data.at_css("NeuralNetwork").attribute("activationFunction").value
	
	layers = data.css("NeuralLayer")
	layers.each do |layer|
		w = []
		b = []
		layer.css("Neuron").each do |neuron|
			b << neuron.attribute("bias").value.to_f
			w << neuron.css("Con").map{ |wi| wi.attribute("weight").value.to_f }
		end
		all_weights << w.transpose
		all_biases << b

		# in PMML format, softmax is a normalization and not an activation
		normalization = layer.attribute("normalizationMethod") && layer.attribute("normalizationMethod").value
		if normalization == "softmax"
			activation = "softmax" 
		else
			layer_activation = layer.attribute("activationFunction")
			activation = layer_activation.nil? ? default_act : layer_activation.value
		end

		all_activations << FastForward.rename_activation(activation)
	end

	d = {
		weights: all_weights,
		biases: all_biases,
		activations: all_activations
	}
	d[:samples] = FastForward.load_sample_data(sample_data) unless sample_data.nil?

	return FastForward.load(d, tol: tol, exception_if_fail: exception_if_fail)

end

.load_sample_data(data) ⇒ Object

Note:

data is supposed to have the following structure: data = { “samples”: { “X”: sample_inputs_array, “y”: sample_outputs_array } }

Loads sample data

Parameters:

  • data (Hash, String) —

    A JSON string/object containing sample data to check the integrity of the loaded model

Returns:

  • The loaded data in a ruby hash



189
190
191
192
193
194
195
196
 
# File 'lib/fast_forward.rb', line 189

def self.load_sample_data(data)
	data = JSON.parse(data) if data.is_a?(String)
	d = {
		X: data["samples"]["X"],
		y: data["samples"]["y"]
	}
	return d
end