[Question] Network with different output kind (softmax & scalar)

[MelnikovIG]

Is it possible to create network with different outputs kind? For example, some outputs should be probability outputs (using softmax), some should be scalar values (from 0 to 1). If it possible, it would be good if anyone can provide example, thanks.

[cbovar]

Do you want to output the two output kinds at the same time ? It's already possible to have proba as output using SoftmaxLayer as last layer (e.g. Mnist demo) or scalar as output using RegressionLayer as last layer (e.g. Regression1DDemo).

[MelnikovIG]

Yes, i want two output kinds at the same time. Is it possible?

[cbovar]

Do you want to get the output of the softmax layer And the output of the layer just before softmax ? Or do you want to get the output of two distinct branches of a network (In that case it is not a stacked layer network anymore and you might be able to do it by using a computation graph)

[MelnikovIG]

@cbovar Thanks, yes i need 2 distinct branches, any example with computation graph?

[cbovar]

Here is a quick example:

(to make it simple, I only don't use any dataset: I use one constant input and two constant outputs)

using System;
using System.Collections.Generic;
using System.Linq;
using ConvNetSharp.Flow;
using ConvNetSharp.Flow.Ops;
using ConvNetSharp.Flow.Training;
using ConvNetSharp.Volume;

namespace Demo
{
    internal static class TwoOutputsExample
    {
        public static void Main()
        {
            var cns = new ConvNetSharp<double>();

            // Graph Creation
            var y_class = cns.PlaceHolder("class_ground_truth");
            var y_scalar = cns.PlaceHolder("scalar_ground_truth");

            Op<double> x = cns.PlaceHolder("x");
            x = cns.Conv(x, 3, 3, 16);
            x = cns.Flatten(x);

            var output1 = cns.Softmax(cns.Dense(x, 10)); // 1st branch: proba
            var output2 = cns.Dense(x, 1); // 2nd branch: scalar

            // Loss 
            var classification_loss = cns.CrossEntropyLoss(output1, y_class); // should output y_class
            var scalar_loss = (output2 - y_scalar) * (output2 - y_scalar); // should output a scalar near y_scalar
            var total_loss = classification_loss + scalar_loss;

            var optimizer = new GradientDescentOptimizer<double>(cns, learningRate: 0.0001);

            var input = BuilderInstance<double>.Volume.Random(Shape.From(16, 16, 1, 1));

            using (var session = new Session<double>())
            {
                session.Differentiate(total_loss); // computes dCost/dW at every node of the graph

                // Training
                double currentLoss;
                do
                {
                    // Build dico containing 1 input and  2 x outputs (1 for each output branch).
                    var class_ground_truth = BuilderInstance<double>.Volume.SameAs(Shape.From(1, 1, 10, 1));
                    class_ground_truth.Set(0, 0, 5, 0, 1.0); // target is class 5
                    var scalar_ground_truth = 0.5;

                    var dico = new Dictionary<string, Volume<double>> {
                        { "x", input },
                        { "class_ground_truth", class_ground_truth },
                        { "scalar_ground_truth", scalar_ground_truth },
                    };

                    // Compute loss
                    currentLoss = session.Run(total_loss, dico);
                    Console.WriteLine($"cost: {currentLoss}");

                    // Run optimizer (will update weights in the network)
                    session.Run(optimizer, dico);
                } while (currentLoss > 0.1);

                // Test
                var result1 = session.Run(output1, new Dictionary<string, Volume<double>> { { "x", input } });
                var class_output = result1.ToArray().ToList().IndexOf(result1.ToArray().Max()); // should be 5
                var result2 = (double)session.Run(output2, new Dictionary<string, Volume<double>> { { "x", input } }); // should be ~0.5
            }

            Console.ReadLine();
        }
    }
}