search

cazala / synaptic

architecture-free neural network library for node.js and the browser
http://caza.la/synaptic
Other
6.91k stars 666 forks source link

Multi feature input for LSTM? #259

Open fpopa opened 6 years ago

fpopa commented 6 years ago

Hello, nice work.

I've started using synaptic Architect with the predefined LSTM network. Everything works fine and I am quite pleased with it.

However I didn't find anything regarding suport for multiple features input for LSTM time series.

E.g. day, number of bus trips, weather, accidents.

How can I use multiple inputs for the same timestamp? Should I try to build my own network which combines results from networks that predict each feature?

thanks.

wagenaartje commented 6 years ago

Ehm, what have you tried? Examples on the README.md show examples of a dataset. Each sample in a data set can have basically unlimited length (input and output). Input and output are just arrays. It is your job to normalize the data.

fpopa commented 6 years ago

That is true.

It is unlimited length for input / output.

I can correctly predict number of bus trips / day by using only this metric and training a LSTM network on historical data, i want to add more data features other than number of bus trips / day to this network in order to enhance the predictions.

I did normalize all data but i don't know how i can use multiple sources of data for the same element in the input array.

I could normalize weather from 0 to 1 as in good -> bad. As well as number of bus trips, 0 none, 1 maximum for a day.

But I don't really know how can I combine both these for a single input?

does it makes sense?

thegamecat commented 6 years ago

I have this exact same question.