コード例 #1
0
        /// <summary>
        /// Trains the network with inputs and targets.
        /// </summary>
        /// <param name="input_lst"></param>
        /// <param name="targets_lst"></param>
        public void Train(List <Double> input_lst, List <Double> targets_lst)
        {
            // Turn input and target arrays into matrices
            Matrix inputs  = Matrix.FromList(input_lst);
            Matrix targets = Matrix.FromList(targets_lst);

            // The input to the hidden layer is the weights (wih) multiplied by inputs
            var hidden_inputs = Matrix.MultiplyMatrices(Wih, inputs);

            // The outputs of the hidden layer pass through sigmoid activation function
            var hidden_outputs = Matrix.Map(hidden_inputs, Activation);

            // The input to the output layer is the weights (who) multiplied by hidden layer
            var output_inputs = Matrix.MultiplyMatrices(Who, hidden_outputs);

            // The output of the network passes through sigmoid activation function
            var outputs = Matrix.Map(output_inputs, Activation);

            // Error is TARGET - OUTPUT
            var output_errors = Matrix.SubtractMatrices(targets, outputs);

            // ~ Back prop.

            // Transpose hidden <-> output weights
            var whoT = Who.Transpose();

            // Hidden errors is output error multiplied by weights (who)
            var hidden_errors = Matrix.MultiplyMatrices(whoT, output_errors);

            // Calculate the gradient, this is much nicer in python!
            var gradient_output = Matrix.Map(outputs, Derivative);

            // Weight by errors and learing rate
            gradient_output.Multiply(output_errors);
            gradient_output.Multiply(LearningRate);

            // Gradients for next layer, more back propogation!
            var gradient_hidden = Matrix.Map(hidden_outputs, Derivative);

            // Weight by errors and learning rate
            gradient_hidden.Multiply(hidden_errors);
            gradient_hidden.Multiply(LearningRate);

            // Change in weights from HIDDEN --> OUTPUT
            var trans_hidden_outputs = hidden_outputs.Transpose();
            var deltaw_output        = Matrix.MultiplyMatrices(gradient_output, trans_hidden_outputs);

            Who.Add(deltaw_output);

            // Change in weights from INPUT --> HIDDEN
            var trans_inputs  = inputs.Transpose();
            var deltaw_hidden = Matrix.MultiplyMatrices(gradient_hidden, trans_inputs);

            Wih.Add(deltaw_hidden);
        }