tf.contrib.training.NextQueuedSequenceBatch

NextQueuedSequenceBatch stores deferred SequenceQueueingStateSaver data.

This class is instantiated by SequenceQueueingStateSaver and is accessible via its next_batch property.

Attributes
batch_size The batch_size of the given batch.

Usually, this is the batch_size requested when initializing the SQSS, but if allow_small_batch=True this will become smaller when inputs are exhausted.

context A dict mapping keys of input_context to batched context.
insertion_index The insertion indices of the examples (when they were first added).

These indices start with the value -2**63 and increase with every call to the prefetch op. Each whole example gets its own insertion index, and this is used to prioritize the example so that its truncated segments appear in adjacent iterations, even if new examples are inserted by the prefetch op between iterations.

key The key names of the given truncated unrolled examples.

The format of the key is:

"%05d_of_%05d:%s" % (sequence, sequence_count, original_key)

where original_key is the unique key read in by the prefetcher.

length The lengths of the given truncated unrolled examples.

For initial iterations, for which sequence * num_unroll < length, this number is num_unroll. For the remainder, this number is between 0 and num_unroll.

next_key The key names of the next (in iteration) truncated unrolled examples.

The format of the key is:

"%05d_of_%05d:%s" % (sequence + 1, sequence_count, original_key)

if sequence + 1 < sequence_count, otherwise:

"STOP:%s" % original_key

where original_key is the unique key read in by the prefetcher.

sequence An int32 vector, length batch_size: the sequence index of each entry.

When an input is split up, the sequence values

0, 1, ..., sequence_count - 1

are assigned to each split.

sequence_count An int32 vector, length batch_size: the sequence count of each entry.

When an input is split up, the number of splits is equal to: padded_length / num_unroll. This is the sequence_count.

sequences A dict mapping keys of input_sequences to split and rebatched data.
total_length The lengths of the original (non-truncated) unrolled examples.

Methods

save_state

View source

Returns an op to save the current batch of state state_name.

Args
state_name string, matches a key provided in initial_states.
value A Tensor. Its type must match that of initial_states[state_name].dtype. If we had at input:
initial_states[state_name].get_shape() == [d1, d2, ...]

then the shape of value must match:

tf.shape(value) == [batch_size, d1, d2, ...]
name string (optional). The name scope for newly created ops.
Returns
A control flow op that stores the new state of each entry into the state saver. This op must be run for every iteration that accesses data from the state saver (otherwise the state saver will never progress through its states and run out of capacity).
Raises
KeyError if state_name does not match any of the initial states declared in initial_states.

state

View source

Returns batched state tensors.

Args
state_name string, matches a key provided in initial_states.
Returns
A Tensor: a batched set of states, either initial states (if this is the first run of the given example), or a value as stored during a previous iteration via save_state control flow. Its type is the same as initial_states["state_name"].dtype. If we had at input:
initial_states[state_name].get_shape() == [d1, d2, ...],

then

state(state_name).get_shape() == [batch_size, d1, d2, ...]
Raises
KeyError if state_name does not match any of the initial states declared in initial_states.

© 2020 The TensorFlow Authors. All rights reserved.
Licensed under the Creative Commons Attribution License 3.0.
Code samples licensed under the Apache 2.0 License.
https://www.tensorflow.org/versions/r1.15/api_docs/python/tf/contrib/training/NextQueuedSequenceBatch