I had a question about using LSTMs for processing data over time. That is, how can I feed data one-by-one into an LSTM, without the LSTM forgetting about my previous inputs?
I had looked through the Keras “stateful” argument a bit, but it had only made me more confused. I’m not sure whether it’s relevant or not for my purposes.
E.g.
Say I am sequentially/chronologically getting data from a soil sensor. The first sample of data is s1. s1 is composed of multiple values (e.g. sunlight, soil temperature, soil humidity). I will get another sample of data, s2, in a few seconds, but I only have s1 right now. So I input s1 into the LSTM as the first timestep. When s2 comes along, I then propagate it through the LSTM as the second timestep. Is this possible to do in Keras?
Advertisement
Answer
RNNs or LSTMs made to remember what happend in past, though, without stateful=True the state is set to 0 for every batch. So, the stateful flag is made for LSTMs to “remeber what happened” in previous batches, with stateful=True LSTM passes its state from previous batch to the next one, so the information doesn’t vanish.
https://fairyonice.github.io/Stateful-LSTM-model-training-in-Keras.html heres a good acticle for you about the statful flag.