CODEWITHSUNDEEP
pythonkeras
newpyguy
newpyguy

Reputation: 31

lstm autoencoder time series data input shape

I looked into different cases but could not understand what I have to choose. My code is also working but my prediction looks strange only having 2 parallel lines. I have an LSTM Autoencoder for regression of time series. Autoencoder because I need to reduce the dimensions. Mydata looks like: 400 samples with each of them containing 5.000 rows.I concatenated them into an array. (Between the samples in real I have 5 minutes). How do I have to choose time steps for the model ? is it like (400,10,5000) ?can anyone please give me an advise with an example also in regarding to the batchsize?

Upvotes: 0

Views: 611

Answers (1)

Daniel Möller
Daniel Möller

Reputation: 86600

  • Sample: an individual "sequence", not connected or related to any other sequence
  • Timesteps: the length of your sequences, each sequence has a start and an end, between them the sequence has steps
  • Features (the last dimension): different parallel variables measured in a same sequence.

Only you can organize your data based on what you know about it. Number of rows and columns don't say anything about it. You must know what they mean and organize them in a shape (samples, timesteps, features) according to those definitions above.

Example:

You measured vital signs in patients for 10 hours.

You had 5 patients, each one was measured for ten hours, and you have the features: body temperature, heartbeat frequency, breath frequency, every five minutes.

Then you have

  • 5 individual sequences (they're not related to each other)
  • 10 hours in steps for each 5 minutes (120 steps)
  • 3 features

Upvotes: 1

Related Questions