A wrapper that unrolls the second (time) dimension of a tensor
into the first (batch) dimension, applies some other
and then rolls the time dimension back up.
Given an input shaped like
(batch_size, time_steps, [rest])and a
Modulethat takes inputs like
TimeDistributedreshapes the input to be
(batch_size * time_steps, [rest]), applies the contained
Module, then reshapes it back.
Note that while the above gives shapes with
Modulealso works if
batch_sizeis second - we always just combine the first two dimensions, then split them.
It also reshapes keyword arguments unless they are not tensors or their name is specified in the optional
forward(*inputs, pass_through: typing.List[str] = None, **kwargs)¶
Defines the computation performed at every call.
Should be overridden by all subclasses.
Although the recipe for forward pass needs to be defined within this function, one should call the
Moduleinstance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.