Skip time-steps in a trained lstm

1

Hi,
I have trained one LSTM network with sequential data. For example, the data is like this:

d0, d1, d2, d3, d4, d5, d6, ...

In test time I want to use the trained model and feed data like this to it:

d0, d2, d4, d6, ...

Is there any way to do this? I tested to feed zero or d0, d0, d2, d2, d4, d4, d6, d6,... sequence but it doesn’t seem to work.

2

If you want to skip every second sample, you could just slice the input tensor in the time dimension via [::2], no?

3

I want to use all data in training, but in test I just want to use every second samples only. But when I do this, outputs of lstm are not good. It seems the LSTM hidden state needs to be updated using both data not just second one. I’m looking for a trick to make it work, in training or in test phase. I’m thinking of input dropout or sth like that.