keras-team/keras-preprocessing

TimeseriesGenerator for many-to-many architectures

kopytjuk opened this issue · 3 comments

kopytjuk commented

What do you think about introducing support for "many-to-many" architectures (see here). In keras-team/keras#11293 there is an issue about that functionality.

In the left side we have current implementation, on the right side the desired behaviour.

image

I would also love to contribute!

herzog-ch commented

Yes please, this feature would help a lot in training RNNs.

annaformaniuk commented

Good idea - for me this would also be a helpful utility, replacing a lot of preprocessing code.

nacheteam commented

I have been looking how to do this in a generator fashion, any examples of how to do this with the actual functions? I suppose I need to extend the Sequence class but I'm not pretty sure how to do it... Thanks in advance!