I am having a hard time incorporating multiple timesteps in Keras stateful LSTM fo multivariate timeseries classification. I couldn't find much useful resources for understanding LSTM 'timesteps' in detail.
In keras LSTM, the input needs to be reshaped from [number_of_entries, number_of_features] to [new_number_of_entries, timesteps, number_of_features]. For example if my input is of shape [300, 5] and I need to use 3 timesteps for LSTM, how should I reshape the input?
After referring this article https://machinelearningmastery.com/use-timesteps-lstm-networks-time-series-forecasting/ what I understood is that the author is using lag features for creating timesteps. That is, the resultant shape becomes [300, 3, 5] where that additional three came from lagged features. Is it the 'right' way?
you should be using the lag features/variables, the lags are basically time shifted versions of the signals.
Keras work with Tensors. For example your data should be in the form:
(number_of_observations, number_of_timesteps, number_of_features)
Hello Mr. Renjith Baby ,
Were you able to complete the project? I am also working on similar project. Multivariate timeseries classification.
In case you completed he project, can you provideme guidance like references, etc...?
Regards,
Udaysimha Nerella
Hi Udaysimha Nerella ,
I have completed my project. The major reference I used was https://machinelearningmastery.com/use-timesteps-lstm-networks-time-series-forecasting/ and other blogs on LSTM and time series analysis by Machine Learning Mastery.
For gaining intuitions and formulating the problem, I used https://www.coursera.org/learn/nlp-sequence-models as reference.
Hope you will find it useful.
Hi Khandu Om
,Sorry for the late answer, I am not sure if this is going to help you or not!
If you are using two time steps, the features from past two instances will be given importance in deciding the output. If it is five, past five instances of features will be considered for your output prediction.
In my opinion, a rough estimate of time steps to use can be decided based on the domain expertise and then you can fine tune it as a hyper parameter around that value.
- Badges
- Science topic