Glossary of TensorFlow Terms _ai

Broadcast operation (broadcasting operation)

An operation that uses Numpy-style broadcasting to ensure the morphological compatibility of tensor parameters. Devices

A piece of hardware that can be used to compute and have its own address space, such as the GPU and CPU. Eval

Tensor a method that returns the value of Tensor. This value is calculated to trigger any graph calculation. Can only be in a session that has been started
To invoke the Tensor value in the diagram. Feed

A concept of TensorFlow: Connect a Tensor directly to any node in a session diagram. The feed is not in the build graph (graph)
Time to create, but to apply when the action of the trigger diagram is performed. A feed temporarily replaces a node with a Tensor value. The Feed data as
Initializes the operation for the parameters of the run () method and the Eval () method. When the method runs, the replacement feed disappears, and the original node definition
Still there. You can create them by Tf.placeholder () to designate specific nodes as feed nodes. See [Basic Usagehref] for details. Fetch

A concept in TensorFlow: To retrieve the output of an operational operation. The retrieved application occurs when a diagram operation is triggered, not a
Born at the time of building the picture. If you want to retrieve the Tensor value of one or more nodes (node), you can call run on the Session object (
Method and executes the chart (graph) as a parameter to the list of nodes to be retrieved. See [Basic Usagehref] for details. Graph (Figure)

Describing an operational task as a direct, DAG Graph (node) represents some of the operations that must be implemented. The Edge generation in the figure
Table data or controllable dependencies. Grathedef is a protocol (API) in the system that describes a chart, which consists of a nodedefs set. A GR
Aphdef can be transformed into a more easily manipulated chart object. Indexedslices (indexed slices)

In the Python API, TensorFlow only embodies the Tensor on the first dimension. If a Tensor has k-dimensional, then a Inde
The Xedslices instance logically represents a collection of (k-1) dimension slices along this Tensor first dimension. The index of the slice is stored continuously in a
In a single one-dimensional vector, the corresponding slices are spliced into a separate K-dimensional Tensor. If sparsity is not limited to the first dimensional space
Room, please use sparsetensor. Node (nodes)

An element in the diagram. A node that initiates a particular operation, including any property used to configure the operation.
The value. For those multiple-form operations, these properties include sufficient information to fully determine the node's signature. See Graph.proto for details. Operation (Op/operation)

In the run-time of TensorFlow, it is an operation similar to add or Matmul or concat. You can use [how to add an Ophre
f) To add a new action to the runtime.

In the Python API, it is a node in the diagram. In the [TF. OPERATIONHREF) class, these operations are enumerated. An action (operatio
N) determines the operation type of this node (nodes), such as Add and Matmul.
Run
The behavior of performing an action in a running diagram. Requires the diagram to run in the session.
In the Python API, it is a method of the session class [TF. SESSION.RUNHREF). You can subscribe to or get run by tensors (
Operation

In the C + + API, it is a method of the [Tensorflow::sessionhre] class. Session (Sessions)

The first step in the start diagram is to create a session object. Session provides some methods for performing operations in diagrams.
In the Python API, use the [TF]. SESSIONHREF).
In the C + + API, [Tensorflow::sessionhref] is the class used to create a diagram and run the operation: Shape

The dimensions of the Tensor and their size.
In a diagram that has been started, it represents the Tensor property that flows between nodes (node). Some operations have a stronger need for shape
Please, if there is no Shape property, an error is reported.
In the Python API, you use the API to create diagrams to illustrate the Shape properties of Tensor. The Shape property of the Tensor is either only partially
Know, or all unknown. See [TF] for details. TENSROSHAPEHREF)
In C + +, the Shape class is used to represent the Tensor dimension. [Tensorflow::tensorshapehref). Sparsetensor

In the Python API, it is used to denote a Tensor that is sparse and scattered in any place in the TensorFlow. Sparsetensor with dictionary-value lattice
Type to store those non-null values along the index. In other words, M non-null values contain a value vector of length m and an index of M columns (indice
s) that comprise the matrix. In order to improve efficiency, Sparsetensor needs to store Indice (indexes) in order by increasing the dimensions, such as the row master
Order. If the sparse value is only along the first dimension, use Indexedslices. Tensor

Tensor is a specific multidimensional array. For example, a four-dimensional array of floating-point types represents a small batch of [Batch,height,width,channel] groups
into the picture.

The

is in a running graph (graph), which is a flow of data between nodes. In Python, the Tensor class represents input and output in operations added to diagram
, see TF. TENSORHREF), such classes do not hold data.
in C + +, the Tensor is the return value of the method [Session::run ()], see tensorflow::tensor, such Tensor holds the data.
Original: [Glossaryhref)