# tr_features and te_features contain features, Y_tr contain train labels

K_tr = CombinedKernel()
K_te = CombinedKernel()
ksub = GaussianKernel(tr_features, tr_features, width)
ksub_tr = ksub.get_kernel_matrix()
ksub.init(tr_features, te_features)
ksub_te = ksub.get_kernel_matrix()
K_tr.append_kernel(CustomKernel(ksub_tr))
K_te.append_kernel(CustomKernel(ksub_te))
K_tr.append_kernel(CustomKernel(ksub_tr))  # if I add these lines results
K_te.append_kernel(CustomKernel(ksub_te))  # are bad, otherwise they are fine

mkl = MKLRegression()
mkl.set_kernel(K_tr)
mkl.set_mkl_norm(2.0)  # I tried 1.0 too, results in same problem
mkl.set_labels(Labels(Y_tr))
mkl.set_C_mkl(10.0)
mkl.set_epsilon(1e-3)
mkl.io.enable_progress()
mkl.train()
#K_tr.set_subkernel_weights((1.0, 0.0))
out_tr = mkl.apply().get_labels()
mkl.set_kernel(K_te)
out_te = mkl.apply().get_labels()