TF 2.0 is out! Get hands-on practice at TF World, Oct 28-31. Use code TF20 for 20% off select passes. Register now

tfa.seq2seq.ScheduledOutputTrainingSampler

View source on GitHub

Class ScheduledOutputTrainingSampler

A training sampler that adds scheduled sampling directly to outputs.

Inherits From: TrainingSampler

Aliases:

Returns False for sample_ids where no sampling took place; True elsewhere.

__init__

View source

__init__(
    sampling_probability,
    time_major=False,
    seed=None,
    next_inputs_fn=None
)

Initializer.

Args:

  • sampling_probability: A float32 scalar tensor: the probability of sampling from the outputs instead of reading directly from the inputs.
  • time_major: Python bool. Whether the tensors in inputs are time major. If False (default), they are assumed to be batch major.
  • seed: The sampling seed.
  • next_inputs_fn: (Optional) callable to apply to the RNN outputs to create the next input when sampling. If None (default), the RNN outputs will be used as the next inputs.

Raises:

  • ValueError: if sampling_probability is not a scalar or vector.

Properties

batch_size

Batch size of tensor returned by sample.

Returns a scalar int32 tensor. The return value might not available before the invocation of initialize(), in this case, ValueError is raised.

sample_ids_dtype

DType of tensor returned by sample.

Returns a DType. The return value might not available before the invocation of initialize().

sample_ids_shape

Shape of tensor returned by sample, excluding the batch dimension.

Returns a TensorShape. The return value might not available before the invocation of initialize().

Methods

initialize

View source

initialize(
    inputs,
    sequence_length=None,
    auxiliary_inputs=None
)

Initialize the TrainSampler.

Args:

  • inputs: A (structure of) input tensors.
  • sequence_length: An int32 vector tensor.

Returns:

(finished, next_inputs), a tuple of two items. The first item is a boolean vector to indicate whether the item in the batch has finished. The second item is the first slide of input data based on the timestep dimension (usually the second dim of the input).

next_inputs

View source

next_inputs(
    time,
    outputs,
    state,
    sample_ids
)

Returns (finished, next_inputs, next_state).

sample

View source

sample(
    time,
    outputs,
    state
)

Returns sample_ids.