LSTM future values prediction

Course Q&A TensorFlow Developer Professional Sequences, Time Series and Prediction
1

Hi there, I am dealing with LSTM forecast in Course C4 W4 in Sequences, Time Series and Prediction within DeepLearning.AI TensorFlow Developer Professional Certificate Specialization and I hope someone can help me out. In all the given examples we in fact compare the predicted data from model.predict with validation data by means of below mentioned code - so we see how well the model predicts. However, how can I predict the future, e.g. + 300 (in depends on given dataset - days, months, whatever, …) ? I understand I should use windowed series + something to get prediction.

Thank you,
J.

image
image771×826 37.2 KB

2

Please see this user guide and move your post to the right topic.

4

Hi Jan. Thank you for updating the topic.

In this lab, Model predicts 1 timestep in the future for window_size timesteps of input data. Think of it as a list of lists where each inner list has window_size worth of data. The outer list represents the batch dimension.

Should you choose to make your model predict more than 1 timestep in the future, say, 10 steps, the following changes are required:

  1. The model output layer should now have 10 output units since the Dense layer now outputs 10 timesteps worth of data instead of 1.
  2. y_valid would be a list of values for each prediction. Care needs to be taken when generating data for training as well.
5

Hi Balaji I really appreciate u try to help me, but I simply cannot figure this one out. I havet tried a lot but no success. Windowing is pretty straightforward for me, with window_size = 20 and batch_size = 16 I get:

  • tf.Tensor: shape=(16, 20), dtype=float32, numpy= array for values, and
  • tf.Tensor: shape=(16,), dtype=float32, numpy= array for targets,

The dataset is shifted+1 and shuffled inside the nested list. But the rest no way … could you pls. give me more clues?
Thank you in advance,
Jan

6

Sure.

Here are a few hints to build a model that takes in window_size data points as input and produces 2 steps of prediction as output.

When creating the dataset, the changes are:

dataset = dataset.flat_map(lambda window: window.batch(window_size + 2))
dataset = dataset.map(lambda window: (window[:-2], window[-2:]))

Here’s an example after making the above changes:

for X_batch, y_batch in train_set.take(1):
    print(f'Model input has shape {X_batch.shape}')
    print(f'Model output has shape {y_batch.shape}')
    print(f'Single input looks like this: {X_batch[0].numpy()}')
    print(f'Output for the above input is: {y_batch[0].numpy()}')

output:

image
image1358×231 28.3 KB

As far as the model is concerned, the output layer should have 2 units instead of 1.

7

Hope I got it right (do not mind precision, test re-learning with 10 Epochs only) but predictions copy the peak at the end. Thank’s a lot Balaji!
Jan

image