def train(self, arr, target):
# INPUT => HIDDEN
input = Matrix.array_to_matrix(arr)
hidden = Matrix.multiply(self.ih_weights, input)
hidden = Matrix.add(hidden, self.ih_biases)
hidden.map(sigmoid)
# HIDDEN => OUTPUT
output = Matrix.multiply(self.ho_weights, hidden)
output = Matrix.add(output, self.ho_biases)
output.map(sigmoid)
# BACKPROPAGATION
# OUTPUT => HIDDEN
expected = Matrix.array_to_matrix(target)
output_error = Matrix.subtract(expected, output)
d_output = Matrix.map(output, dsigmoid)
hidden_T = Matrix.transpose(hidden)
gradient = Matrix.hadamard(d_output, output_error)
gradient = Matrix.scalar_multiply(gradient, self.learning_rate)
# Apply to ho_biases
self.ho_biases = Matrix.add(self.ho_biases, gradient)
# Apply to ho_weights
weights_ho_deltas = Matrix.multiply(gradient, hidden_T)
self.ho_weights = Matrix.add(self.ho_weights, weights_ho_deltas)
# HIDDEN => INPUT
weights_ho_T = Matrix.transpose(self.ho_weights)
hidden_error = Matrix.multiply(weights_ho_T, output_error)
d_hidden = Matrix.map(hidden, dsigmoid)
input_T = Matrix.transpose(input)
gradient_H = Matrix.hadamard(d_hidden, hidden_error)
gradient_H = Matrix.scalar_multiply(gradient_H, self.learning_rate)
# Apply to ih_biases
self.ih_biases = Matrix.add(self.ih_biases, gradient_H)
# Apply to ih_weights
weights_ih_deltas = Matrix.multiply(gradient_H, input_T)
self.ih_weights = Matrix.add(self.ih_weights, weights_ih_deltas)
def predict(self, arr):
# INPUT => HIDDEN
input = Matrix.array_to_matrix(arr)
hidden = Matrix.multiply(self.ih_weights, input)
hidden = Matrix.add(hidden, self.ih_biases)
hidden.map(sigmoid)
# HIDDEN => OUTPUT
output = Matrix.multiply(self.ho_weights, hidden)
output = Matrix.add(output, self.ho_biases)
output.map(sigmoid)
output = Matrix.matrix_to_array(output)
return output
