CODEWITHSUNDEEP
machine-learningtime-seriesdeep-learninglstmrecurrent-neural-network
msa
msa

Reputation: 31

How do you train LSTMs on time series data with too many time steps?

I am trainig LSTM/GRUs on time series data.

My input data's shape is (5000,12) i.e. 5000 time steps and at each time step has 12 dimensions. When I am trying to run LSTM on this data taking each time step (dim: 12) at a time; this means I am unrolling LSTM 5000 times and I am unable to train the network on my laptop as it is exceeding 16gb ram.

One solution I can think of is can I combine, say, 10 time steps data and feed that data (dim: 10*12) to LSTM network (potentially unrolling only 500 times)...

Can this give me good results?

Upvotes: 0

Views: 1450

Answers (1)

KETAN PATIL
KETAN PATIL

Reputation: 2614

there are various ways to tackle this problem depending on the dataset you have, have a look at this: http://machinelearningmastery.com/handle-long-sequences-long-short-term-memory-recurrent-neural-networks/

Upvotes: 1

Related Questions