Exemplo n.º 1
0
        public void ModelGradients3()
        {
            Matrix <double> in_hid = Matrix <double> .Build.DenseOfArray(
                new double[, ] {
                { 1 }, { 2 }
            }
                );

            Matrix <double> hid_class = Matrix <double> .Build.DenseOfArray(
                new double[, ] {
                { 1, 0 }, { 0, 1 }
            }
                );

            Matrix <double> input = Matrix <double> .Build.DenseOfArray(
                new double[, ] {
                { 1 }
            }
                );

            Matrix <double> target = Matrix <double> .Build.DenseOfArray(
                new double[, ] {
                { 1 }, { 0 }
            }
                );

            MutableTypes.Model        networkModel = new MutableTypes.Model(in_hid, hid_class);
            MutableTypes.NetworkState networkState = NetRun.RunNetwork(networkModel, input);

            MutableTypes.Model gradients = Training._getModelGradients(networkModel, networkState, target);

            double l1   = NetworkFunctions.Logistic(1);
            double l2   = NetworkFunctions.Logistic(2);
            double el1  = Math.Exp(l1);
            double el2  = Math.Exp(l2);
            double N    = el1 + el2;
            double el1N = el1 / N;
            double el2N = el2 / N;
            double ld1  = NetworkFunctions.LogisticDerivative_f_output(l1);
            double ld2  = NetworkFunctions.LogisticDerivative_f_output(l2);



            Matrix <double> hid_class_gradient_check = Matrix <double> .Build.DenseOfArray(
                new double[, ] {
                { l1 *(el1N - 1), l2 *(el1N - 1) }, { l1 *el2N, l2 *el2N }
            }
                );

            Matrix <double> in_hid_gradient_check = Matrix <double> .Build.DenseOfArray(
                new double[, ] {
                { ld1 *(el1N - 1) }, { ld2 *el2N }
            }
                );

            Assert.AreEqual(hid_class_gradient_check, gradients.HiddenToClassifier);
            Assert.AreEqual(in_hid_gradient_check, gradients.InputToHidden);
        }
Exemplo n.º 2
0
        public void ModelGradients2()
        {
            Matrix <double> in_hid = Matrix <double> .Build.DenseOfArray(
                new double[, ] {
                { 1 }
            }
                );

            Matrix <double> hid_class = Matrix <double> .Build.DenseOfArray(
                new double[, ] {
                { 1 }, { 1 }
            }
                );

            Matrix <double> input = Matrix <double> .Build.DenseOfArray(
                new double[, ] {
                { 1 }
            }
                );

            Matrix <double> target = Matrix <double> .Build.DenseOfArray(
                new double[, ] {
                { 1 }, { 0 }
            }
                );

            MutableTypes.Model        networkModel = new MutableTypes.Model(in_hid, hid_class);
            MutableTypes.NetworkState networkState = NetRun.RunNetwork(networkModel, input);

            MutableTypes.Model gradients = Training._getModelGradients(networkModel, networkState, target);


            Matrix <double> hid_class_gradient_check = Matrix <double> .Build.DenseOfArray(
                new double[, ] {
                { -NetworkFunctions.Logistic(1) / 2 }, { NetworkFunctions.Logistic(1) / 2 }
            }
                );

            Matrix <double> in_hid_gradient_check = Matrix <double> .Build.DenseOfArray(
                new double[, ] {
                { 0 }
            }
                );

            Assert.AreEqual(hid_class_gradient_check, gradients.HiddenToClassifier);
            Assert.AreEqual(in_hid_gradient_check, gradients.InputToHidden);
        }