num_valid_file = 399 if b_16k: n_ins = 87 n_outs = 187 else: n_ins = 87 n_outs = 193 dir_lab_norm = 'nn_no_silence_lab_norm_' + str(n_ins) dir_cmp_norm = 'nn_norm_mgc_lf0_vuv_bap_' + str(n_outs) file_names = 'file_id_list_full.scp' # prepare the training data file_list = pap.read_file_list(dir_base + file_names) cmp_norm_file_list = pap.prepare_file_path_list(file_list, dir_base + dir_cmp_norm, '.cmp') lab_norm_file_list = pap.prepare_file_path_list(file_list, dir_base + dir_lab_norm, '.lab') train_x_file_list = lab_norm_file_list[0:num_train_file] train_y_file_list = cmp_norm_file_list[0:num_train_file] valid_x_file_list = lab_norm_file_list[num_train_file:num_train_file + num_valid_file] valid_y_file_list = cmp_norm_file_list[num_train_file:num_train_file + num_valid_file] train_data_reader = dap.ListDataProvider(x_file_list=train_x_file_list, y_file_list=train_y_file_list, n_ins=n_ins,
n_outs = 193 dir_lab_norm = 'nn_no_silence_lab_norm_' + str(n_ins) file_names = 'test_id_list.scp' saved_epoch = 1 saved_ckp_model = dir_base + 'models/mxnet_bigru_sym_' output_dir = dir_base + 'synthesis' if not os.path.exists(output_dir): os.makedirs(output_dir) sym, arg_params, aux_params = mx.model.load_checkpoint(saved_ckp_model, saved_epoch) file_list = pap.read_file_list(dir_base + file_names) lab_norm_file_list = pap.prepare_file_path_list(file_list, dir_base + dir_lab_norm, '.lab') generate_file_list = pap.prepare_file_path_list(file_list, output_dir, '.cmp') # generate cmp file num_file = len(lab_norm_file_list) for i in range(num_file): # read label feature from file features = np.fromfile(lab_norm_file_list[i], dtype=np.float32) # evaluation features = features[:(n_ins * (features.size / n_ins))] input_labels = features.reshape((-1, n_ins)) mod = mx.mod.Module(symbol=sym, context=mx.gpu(), label_names=[]) mod.bind(data_shapes=[('data', input_labels.shape)], for_training=False) mod.set_params(arg_params, aux_params, allow_missing=True) mod.forward(mx.io.DataBatch([nd.array(input_labels)]))