Support constants in StackedRNNCells

[yuyang-huang]

This PR is probably useful for implementing (some variants of) attention RNNs.

1. Enable the constants argument in StackedRNNCells.call. The constants will be routed to an underlying cell if it also supports constants.
2. Handle the shape of the constants properly in StackedRNNCells.build so that an RNN cell with constants can be stacked on top of another cell. Before this PR,

• If the bottom cell does not support constants:

constants = Input((32,))
x = Input((None, 5))
x = RNN([GRUCell(16), RNNCellWithConstants(32)])(x, constants=constants)

The GRUCell fails to build because input_shape, which is now a list [step_input_shape] + constants_shape, is passed to GRUCell.build directly. We should only pass step_input_shape to GRUCell.build in this case.

• If the bottom cell also supports constants:

constants = Input((32,))
x = Input((None, 5))
x = RNN([RNNCellWithConstants(16), RNNCellWithConstants(32)])(x, constants=constants)

The top cell fails to build because the input_shape passed to it is incorrect.

Also modified the test to cover the two cases above.

[fchollet]

Re-running tests

[fchollet]

That's a reasonable behavior change. Code LGTM.

[yuyang-huang]

Thanks for reviewing! Seems like the tests are passing now.

[LeZhengThu]

Hi, can anyone help explain the argument "constants"? I'm a bit lost about the argument. Thanks.

[fchollet]

They're constant tensors passed to the underlying cell at every call.

[LeZhengThu]

@fchollet Does these tensors have timesteps? Are they something like the static features fed in the network?

[fchollet]

No, they don't have timesteps. They're passed to each cell call. They have no concept of time, that's why they're called "constants".