min13s
commented
2 years ago is there something wrong with the hessian matrix， since grad is weighted, so should the hessian? it seems that grad and hess is the ensampled gradient and hessian(the algorithm 2 ), why still need grad2 and hess2 ?
for the MT-GBM def mymse2(preds, train_data, ep = 0): labels = train_data.get_label() labels2 = labels.reshape((num_labels,-1)).transpose() preds2 = preds.reshape((num_labels,-1)).transpose() grad2 = (preds2 - labels2) *grad = grad2 np.array([20,0.001])** grad = np.sum(grad,axis = 1) grad2 = grad2.transpose().reshape((-1))
hess = grad 0. + 1 hess2 = grad2 0. + 1 return grad, hess, grad2, hess2
it seems that parameter will change the best mape, rmse, compared to common lightgbm, it has one more parameter ,