A simple helper to maintain a dict. It is a sub-class of dict with the following extensions/restrictions: - It supports attr access to its memb

Shape represents a tensor's symbolic shape.

A class that holds a variable and a gradient. This does not inherit from namedtuple so that tf.nest operations do not recurse into it.

A dict with a __repr__ that is always sorted by key.

A container for a list of 7-DOF 3D boxes. Each 3D box is represented by the 7-tuple [x, y, z, dx, dy, dz, phi], where x, y, z is the center of th

Function builds a TensorFlow graph function from a callable. In the high level this is similar to tf.Defun and tf.function. In fact this relies o

Represents a sharding annotation for GShard/XLA.

A vertex with (x, y) coordinate. This is an internal implementation detail of RotatedBox2D.

A helper class to construct a recurrent neural net.

Accumulator for the sufficient statistics.

Builder for StarNet Model.

Builder for the Pillars model.

GPipe-compatible Evolved Transformer decoder layer.

Stacked self- multi-head attention and fully connected layers. With optional layer normalization applied to the final output. See 'Attention Is

Represents the input/output signature of a GraphLayer. A signature is of the form: input->output or, optionally just an input: input T

Decoder with state which produces well-defined logits. It determines the best next token according to one of the rules: - increment last token id

Reshapes tensors for functions/layers that expect a single batch dimension. Example: Suppose a sequence `encoder` has signature :: [batch_si

Dataset error exception class.

Class to keep track of error counts.

Defines query-result example pairs to be labeled. `query_examples` and `result_examples` are the input batches of the queries and results being l

GPipe-compatible Evolved Transformer encoder layer.

A collection of named tensors (or NestedMaps of tensors).

Handler for HighlightAlignedHtml.

Computes ground truths for MultiHeadedfAttentionRPE. Written in a non-vectorized way.

A rotated 2D bounding box represented as (cx, cy, w, h, r). cx, cy are the box center coordinates; w, h are the box width and height; heading is the r

Test dataclasses.dataclass.

Non-reentrant context that holds the list of summary tensors.

Plain data object. Much better than namedtuple.

Get graph local value from a defined collection.

Stores data for a single parameter.

A helper to uniqify variables in a NestedMap.

Class to compute the AUC score for binary classification.

Helper class.

Helper class.

Class to compute a weighted (arithmetic) average value metric.

Saveable that loads Variables as bfloat16.

A representation of a 2D rotated bounding box. Box2D is based on conventions for 3D coordinate systems where y-x are switched.

NodeVisitor for collecting datasets for a model.

Encapsulates a TensorFlow graph function and its properties.

Generates encodings of text captions.

EMA related instances which an executor prepares.

Converter utility from car.open_dataset.Frame to tf.Examples.

GPipe compatible transformer layer.

Subclass for LM batch decoding.

Optimizer wrapper providing deferred gradient application. Large hardware configurations are in high-demand, and difficult to come by. This class

Simple tuple for book keeping during beam pruning.

Params which can be instantiated. When using InstantiableParams, callers must provide a class which supports initialization using a Params instan

Describes the `tf.data.Dataset`s that a `DatasetSpec` instance produces.

Record given metric versus global_step history to a file.

Oracle layer used for computing ground truths for MultiHeadedAttention. Written in a non-vectorized way.

Stores data for a set of parameters. Provides attribute-based API, e.g. "params.foo = 5". Uses internal {'name': _Param} dict for storing paramet

Config for tensor partitioning.

Metadata for partitioning.

A summary tensor with the first dim being the batch.

Reads splits of the 'coco_captions' dataset via `tensorflow_datasets`.

Context manager. Rewrites tf.while_loop to propagate summary tensors.

Helper class for managing runners.

Wrapper interface between tf.train.Saver and the custom saver.

Used for testing ToProto.

Adam optimizer that shards the slot variables.

Compute word error rates after the alignment. Attributes: key_phrases: list of important phrases. aligned_htmls: list of diagnois htmls, ea

Collect stats.

Test namedtuple class.

Simple accumulator for counting timesteps.

Ad-hoc hack to recognize all punctuation and symbols.

Un-segments (merges) segmented strings. Used to retain back the original surface form of strings that are encoded using byte-pair-encoding (BPE),

Calculate average precision as function of difficulty.

Wrapper object.

Adafactor optimizer for XLA sharding.

A `Dataset` that caches elements of its input.

Accumulator for a counted min/max. Represented as a tensor of shape [count, min, max]. Every update increases the count and expands the min/max (

An adapter providing `tf.Session`-like interface in eager mode. This adapter is designed to be used with the @DefineAndTrace decorator above, and

Input layers.

Placer which places a variable on the least loaded var device. We use total byte sizes of variables placed on a device to indicate the device's l

A link is a pair of channels.

A thread-local stack of symbol-to-value dicts.

Middle layers.

Output layers.