$OpDocRNNBidirectionalDynamicRNN
$OpDocRNNBidirectionalDynamicRNN
RNN cell to use for the forward direction.
RNN cell to use for the backward direction.
Input to the RNN loop.
Initial state to use for the forward RNN, which is a sequence of tensors with shapes
[batchSize, stateSize(i)]
, where i
corresponds to the index in that sequence.
Defaults to a zero state.
Initial state to use for the backward RNN, which is a sequence of tensors with shapes
[batchSize, stateSize(i)]
, where i
corresponds to the index in that sequence.
Defaults to a zero state.
Boolean value indicating whether the inputs
are provided in time-major format (i.e.,
have shape [time, batch, depth]
) or in batch-major format (i.e., have shape
[batch, time, depth]
).
Number of RNN loop iterations allowed to run in parallel.
If true
, GPU-CPU memory swapping support is enabled for the RNN loop.
Optional INT32
tensor with shape [batchSize]
containing the sequence lengths for
each row in the batch.
Name prefix to use for the created ops.
Tuple containing: (i) the forward RNN cell tuple after the forward dynamic RNN loop is completed, and (ii)
the backward RNN cell tuple after the backward dynamic RNN loop is completed. The output
of these tuples
has a time axis prepended to the shape of each tensor and corresponds to the RNN outputs at each iteration
in the loop. The state
represents the RNN state at the end of the loop.
InvalidShapeException
If the inputs or the provided sequence lengths have invalid or unknown shapes.
$OpDocRNNDynamicRNN
$OpDocRNNDynamicRNN
RNN cell to use.
Input to the RNN loop.
Initial state to use for the RNN, which is a sequence of tensors with shapes
[batchSize, stateSize(i)]
, where i
corresponds to the index in that sequence.
Defaults to a zero state.
Boolean value indicating whether the inputs
are provided in time-major format (i.e.,
have shape [time, batch, depth]
) or in batch-major format (i.e., have shape
[batch, time, depth]
).
Number of RNN loop iterations allowed to run in parallel.
If true
, GPU-CPU memory swapping support is enabled for the RNN loop.
Optional INT32
tensor with shape [batchSize]
containing the sequence lengths for
each row in the batch.
Name prefix to use for the created ops.
RNN cell tuple after the dynamic RNN loop is completed. The output
of that tuple has a time axis
prepended to the shape of each tensor and corresponds to the RNN outputs at each iteration in the loop.
The state
represents the RNN state at the end of the loop.
InvalidArgumentException
If neither initialState
nor zeroState
is provided.
InvalidShapeException
If the inputs or the provided sequence lengths have invalid or unknown shapes.