Tensor
【免费下载链接】geGE(Graph Engine)是面向昇腾的图编译器和执行器,提供了计算图优化、多流并行、内存复用和模型下沉等技术手段,加速模型执行效率,减少模型内存占用。 GE 提供对 PyTorch、TensorFlow 前端的友好接入能力,并同时支持 onnx、pb 等主流模型格式的解析与编译。
项目地址: https://gitcode.com/cann/ge
Product Support
| Product | Support |
|---|
| Atlas A3 Training Products/Atlas A3 Inference Products | √ |
| Atlas A2 Training Products/Atlas A2 Inference Products | √ |
Module Import
from ge.graph import Tensor
Functionality Description
Tensor class is tensor data class, supports creating tensors via memory data or files. Supports setting and getting tensor's Format, DataType, Shape, Data (TensorLike) and Placement. Tensors support migration between Host and Device.
Function Prototype
Constructor
Tensor(data=None, file_path=None, data_type=DataType.DT_FLOAT, format=Format.FORMAT_ND, shape=None, placement=Placement.PLACEMENT_HOST)
Properties
@property format -> Format
@property data_type -> DataType
@property shape -> Shape
@property data -> TensorLike
@property placement -> Placement
Methods
set_format(format: Format) -> Tensor
get_format() -> Format
set_data_type(data_type: DataType) -> Tensor
get_data_type() -> DataType
get_shape() -> Shape
get_data() -> TensorLike
get_tensor_desc() -> TensorDesc
get_placement() -> Placement
to_host() -> Tensor
to_device() -> Tensor
Parameter Description
Constructor Parameters
| Parameter | Type | Required | Description |
|---|
| data | Union[List[int], List[float], List[bool], None] | No | Memory data, pass tensor data via list. Choose one between data and file_path, cannot specify both. |
| file_path | Union[str, None] | No | File path, read tensor data from file. Choose one between file_path and data, cannot specify both. |
| data_type | DataType | No | Tensor's data type, use DataType enum value, default is DataType.DT_FLOAT. |
| format | Format | No | Tensor's data format, use Format enum value, default is Format.FORMAT_ND. |
| shape | Union[List[int], None] | No | Tensor's shape, represented as integer list for each dimension size. If None, indicates scalar. |
| placement | Placement | No | Tensor's placement location, use Placement enum value, default is Placement.PLACEMENT_HOST. |
set_format Parameters
| Parameter | Type | Required | Description |
|---|
| format | Format | Yes | Target data format, use Format enum value. |
set_data_type Parameters
| Parameter | Type | Required | Description |
|---|
| data_type | DataType | Yes | Target data type, use DataType enum value. |
Return Value Description
Constructor
Returns Tensor object instance.
Property Return Values
| Property | Return Type | Description |
|---|
| format | Format | Tensor's data format. |
| data_type | DataType | Tensor's data type. |
| shape | Shape | Tensor's shape information. |
| data | TensorLike | Tensor's data content. Scalar returns single value, non-scalar returns nested list structure. |
| placement | Placement | Tensor's placement location. |
Method Return Values
| Method | Return Type | Description |
|---|
| set_format | Tensor | Returns self, supports chained calls. |
| get_format | Format | Returns tensor's data format. |
| set_data_type | Tensor | Returns self, supports chained calls. |
| get_data_type | DataType | Returns tensor's data type. |
| get_shape | Shape | Returns tensor's shape information. |
| get_data | TensorLike | Returns tensor's data content. Scalar returns single value, non-scalar returns nested list structure. |
| get_tensor_desc | TensorDesc | Returns tensor's description info (TensorDesc object). |
| get_placement | Placement | Returns tensor's placement location. |
| to_host | Tensor | Returns self, migrates tensor from Device to Host. |
| to_device | Tensor | Returns self, migrates tensor from Host to Device. |
Constraints
- When constructing tensor, only one of data and file_path can be specified, cannot specify both, and cannot specify neither (when neither specified creates empty tensor).
- Supported data types include: DT_FLOAT, DT_FLOAT16, DT_INT8, DT_INT32, DT_UINT8, DT_INT16, DT_UINT16, DT_UINT32, DT_INT64, DT_UINT64, DT_BOOL. Does not support DT_DOUBLE.
- shape parameter must be integer list (list of int), if None indicates scalar.
- placement parameter must be Placement enum value.
- Tensor does not support copy and deepcopy.
- to_host() only applies to tensors currently on Device; to_device() only applies to tensors currently on Host.
【免费下载链接】geGE(Graph Engine)是面向昇腾的图编译器和执行器,提供了计算图优化、多流并行、内存复用和模型下沉等技术手段,加速模型执行效率,减少模型内存占用。 GE 提供对 PyTorch、TensorFlow 前端的友好接入能力,并同时支持 onnx、pb 等主流模型格式的解析与编译。
项目地址: https://gitcode.com/cann/ge