| 40 | } |
| 41 | |
| 42 | void cost(array &J, array &dJ, const array &Weights, const array &X, |
| 43 | const array &Y, double lambda = 1.0) { |
| 44 | // Number of samples |
| 45 | int m = Y.dims(0); |
| 46 | |
| 47 | // Make the lambda corresponding to Weights(0) == 0 |
| 48 | array lambdat = constant(lambda, Weights.dims()); |
| 49 | |
| 50 | // No regularization for bias weights |
| 51 | lambdat(0, span) = 0; |
| 52 | |
| 53 | // Get the prediction |
| 54 | array H = predict(X, Weights); |
| 55 | |
| 56 | // Cost of misprediction |
| 57 | array Jerr = -sum(Y * log(H)); |
| 58 | |
| 59 | // Regularization cost |
| 60 | array Jreg = 0.5 * sum(lambdat * Weights * Weights); |
| 61 | |
| 62 | // Total cost |
| 63 | J = (Jerr + Jreg) / m; |
| 64 | |
| 65 | // Find the gradient of cost |
| 66 | array D = (H - Y); |
| 67 | dJ = (matmulTN(X, D) + lambdat * Weights) / m; |
| 68 | } |
| 69 | |
| 70 | array train(const array &X, const array &Y, double alpha = 0.1, |
| 71 | double lambda = 1.0, double maxerr = 0.01, int maxiter = 1000, |