How to handle a timestep length of 0?

[jdevoldere]

It can handle variable timestep size just fine as long as it's not zero, in which case I get this error:

tensorflow.python.framework.errors_impl.InvalidArgumentError: 2 root error(s) found.
  (0) Invalid argument: slice index -1 of dimension 1 out of bounds.
	 [[{{node tcn/lambda/strided_slice}}]]
	 [[loss/mul/_257]]
  (1) Invalid argument: slice index -1 of dimension 1 out of bounds.
	 [[{{node tcn/lambda/strided_slice}}]]

[philipperemy]

@jdevoldere in which case would you have a timestep size of 0? I'm not sure Keras/tensorflow can even handle a tensor of 0.

[jdevoldere]

@philipperemy when predicting the very first step of a sequence. for now I work around it by passing a start state to the beginning of each sequence in preprocessing.

[philipperemy]

Well in deep learning, you need an input to generate an output. Otherwise you first output is just purely random. So in your case, you should leave out a few steps before predicting.

[philipperemy]

I'll close this issue if you don't have any other questions. Feel free to re-open it.