i am trying to validate a tensorflow code for simple timeseries prediction in where: X = np.arange(0, 2000, 0.5) y = 2 * np.sin(X) + 0.8 * np.random.rand(X.shape[0]) and the following parameters for the timeseries: LOOK_BACK = 100 FORECAST_HORIZON = 100 any model I use or try gets very bad performance, LSTM, MLP, CNN... loss is not improved in any epoch. def create_dataset(X_data, y_data, shuffle=False, repeat=False): # Ensure input and output are tensors X_data = tf.convert_to_tensor(X_data, dtype=tf.float32) y_data = tf.convert_to_tensor(y_data, dtype=tf.float32) # Create windowed dataset dataset = tf.keras.preprocessing.timeseries_dataset_from_array( data=tf.concat([X_data, y_data], axis=-1), # combine to keep alignment targets=None, sequence_length=look_back + forecast, sequence_stride=1, shuffle=shuffle, batch_size=1, ) def split_input_target(sequence): input_seq = sequence[:, :look_back, : X_data.shape[-1]] target_seq = sequence[:, look_back:, X_data.shape[-1] :] return input_seq, target_seq dataset = dataset.map(split_input_target) return dataset train_dataset = create_dataset(X_train, y_train, shuffle=True, repeat=True) val_dataset = create_dataset(X_val, y_val) test_dataset = create_dataset(X_test, y_test) return train_dataset, val_dataset, test_dataset Continue reading...
