def save_preproccesing_parameters(dataset_dir): y_dir = os.path.join(dataset_dir, 'train', 'y') mean_path = os.path.join(dataset_dir, 'transform_0_mean.txt') std_path = os.path.join(dataset_dir, 'transform_0_std.txt') train_set = Datafolder_soundfiles(y_paths=walk_dir(y_dir)) training_data_loader = DataLoader(train_set, batch_size=1, num_workers=2) scaler0 = StandardScaler() t_0 = transform_tf(feature_transform_0()) for index, x, y in training_data_loader: y = y[0].numpy() data, Y_m, Y_a, length = t_0(y=y) scaler0.partial_fit(data.numpy().T) # each time step is treated equally np.savetxt(mean_path, scaler0.mean_) np.savetxt(std_path, scaler0.scale_) print(scaler0.mean_) print(scaler0.scale_)
# training using a phase sesitive masks, ReLU activation on the output p = { 'experiment_name': 'BLSTM_A11', 'model_class': BLSTM_A, 'model_kwargs': { 'input_size': 100, 'output_size': 129, 'hidden_size': 384, 'LSTM_layers': 2, 'output_act_f': 'ReLU' }, 'input_transform': transform_tf( feature_transform_0(os.path.join('data', 'processed', 'dataset_5')), ideal_phase_sensitive_target), 'output_transform': apply_mask, 'training_set': load_dataset_5_train, 'validation_set': load_dataset_4_val, 'batch_size': 10, 'epochs_max': 57, 'samples_per_epoch': 21384, # length of dataset_3 and 4 'criterion': nn.MSELoss(),
# batch_size is douple, but not faster or slower p = { 'experiment_name': 'BLSTM_A14', 'model_class': BLSTM_A, 'model_kwargs': { 'input_size': 100, 'output_size': 129, 'hidden_size': 384, 'LSTM_layers': 2 }, 'input_transform': transform_tf( feature_transform_0(os.path.join('data', 'processed', 'dataset_5')), ideal_amplitude_target), 'output_transform': apply_mask, 'training_set': load_dataset_5_train, 'validation_set': load_dataset_4_val, 'batch_size': 10 * 2, 'epochs_max': 57, 'samples_per_epoch': 21384, # length of dataset_3 and 4 'criterion': nn.MSELoss(),