With stateful LSTM the network state is propagated to subsequent sequences and batches. I have multiple data files with data that I present to the network for training (making this multi-series). My question is how to create the validation data.

At the moment I take 20% of the data from either head or tail of each file and use that for the validation data, and the other 80% I submit to the NN.

However, because this is stateful, I wonder if that's not a good way to create my validation data. To elaborate; when training I potentially present several years or a decade of data, so when it comes to the final sequence the network, it has the state built up from previous years of data. If I present some validation data which is maybe just a week of data, then the network doesn't build up any prior state to use to generate the final output. Am I right in thinking this isn't a good way to validate a stateful RNN/LSTM?

What I'm thinking instead is to put aside 20% of the data files I have, and use the entire data in those files to create the validation loss. Would that be better?


1 Answer 1


Using a stateful model, am I correct in assuming that you have time-series data?

If so, it would perhaps make sense (at least for your test accuracy) to always validate against the time-step that immediately follows a test batch. This is what you would actually want to be doing with the model once it is trained.

Have a look at this answer, where I explain the idea in a lot more detail. The main idea is to split data into a kind of rolling forecast pattern, whereby you would have a test batch e.g. with something like:

window_size = 100     #whatever makes sense for your data
val_size = 15

train_batch1 = data[0:window_size]
val_batch1 = data[window_size:window_size + val_size]

train_batch2 = data[1:window_size+1]
val_batch2 = data[window_size + 1:window_size + val_size + 1]


That is just pseudo code to make it as clear as possible. Here is the diagram I created for the post linked above, which visualises the same notion:

rolling window validation

In the image, it just has train/test sets... you could of course add validation into the mix.

If this doesn't seem to make sense for you data, temporal relationships must not matter too much. In that case, perhaps a stateful model is also not entirely useful.

  • 1
    $\begingroup$ Thanks for the answer and sorry for the delayed reply. Yes, it's time-series. I understand your answer, and that's what I do now, but when testing or using in real-life, if I only input 10% of data, then no prior state has been established in the network. I would have thought for stateful you might want to present a more significant amount of data so the network can build up a state in order to form a prediction? But your info re: train/test window answered another of my questions here: datascience.stackexchange.com/questions/34151 (you might want to post as the answer) $\endgroup$
    – BigBadMe
    Jul 17, 2018 at 20:00
  • $\begingroup$ But here you are using the same data for both training and testing. It's a fundamental error to mix training and testing data. Maybe I've misunderstood your answer because it doesn't seem right? $\endgroup$ May 21, 2020 at 0:29

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.