TensorDesc

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 TensorDesc

Functionality Description

TensorDesc class describes tensor metadata, including Shape, Format and DataType. Supports distinguishing between original shape/format and current shape/format, suitable for complete description of tensor attributes during graph building. All set_* methods return self, supporting chained calls. Does not support copy and deepcopy.

Class Definition

class TensorDesc:
    def __init__(
        self,
        shape: Optional[List[int]] = None,
        format: Optional[Format] = Format.FORMAT_ND,
        data_type: Optional[DataType] = DataType.DT_FLOAT,
    ) -> None

Function List

Function	Description
__init__(shape=None, format=Format.FORMAT_ND, data_type=DataType.DT_FLOAT)	Constructor, creates TensorDesc object. shape is dimension list, None indicates scalar; format is data format, default FORMAT_ND; data_type is data type, default DT_FLOAT
shape (property)	Get current shape as property, returns Shape object
origin_shape (property)	Get original shape as property, returns Shape object
format (property)	Get current data format as property, returns Format enum value
origin_format (property)	Get original data format as property, returns Format enum value
data_type (property)	Get data type as property, returns DataType enum value
get_shape()	Get current shape
set_shape(shape)	Set current shape, supports chained calls
get_origin_shape()	Get original shape
set_origin_shape(shape)	Set original shape, supports chained calls
get_format()	Get current data format
set_format(format)	Set current data format, supports chained calls
get_origin_format()	Get original data format
set_origin_format(format)	Set original data format, supports chained calls
get_data_type()	Get data type
set_data_type(data_type)	Set data type, supports chained calls

Parameter Description

__init__ Parameters

Parameter	Type	Required	Description
shape	List[int]	No	Tensor's dimension list, e.g. [1, 3, 224, 224]. None indicates scalar (empty list). Default is None
format	Format	No	Tensor's data format, takes Format enum value. Default is Format.FORMAT_ND
data_type	DataType	No	Tensor's element data type, takes DataType enum value. Default is DataType.DT_FLOAT

set_shape / set_origin_shape Parameters

Parameter	Type	Required	Description
shape	List[int]	Yes	Target dimension list, must be integer list

set_format / set_origin_format Parameters

Parameter	Type	Required	Description
format	Format	Yes	Target data format, must be Format enum value

set_data_type Parameters

Parameter	Type	Required	Description
data_type	DataType	Yes	Target data type, must be DataType enum value

Return Value Description

Function	Return Type	Description
shape (property)	Shape	Current shape object
origin_shape (property)	Shape	Original shape object
format (property)	Format	Current data format enum value
origin_format (property)	Format	Original data format enum value
data_type (property)	DataType	Data type enum value
get_shape()	Shape	Current shape object
set_shape(shape)	TensorDesc	Returns self, supports chained calls
get_origin_shape()	Shape	Original shape object
set_origin_shape(shape)	TensorDesc	Returns self, supports chained calls
get_format()	Format	Current data format enum value
set_format(format)	TensorDesc	Returns self, supports chained calls
get_origin_format()	Format	Original data format enum value
set_origin_format(format)	TensorDesc	Returns self, supports chained calls
get_data_type()	DataType	Data type enum value
set_data_type(data_type)	TensorDesc	Returns self, supports chained calls

Constraints

• Does not support copy: calling copy.copy() will throw RuntimeError.
• Does not support deepcopy: calling copy.deepcopy() will throw RuntimeError.
• In constructor, format parameter must be Format enum type, otherwise throws TypeError.
• In constructor, data_type parameter must be DataType enum type, otherwise throws TypeError.
• For set_shape, set_origin_shape, shape parameter must be integer list (list of int), otherwise throws TypeError.
• For set_format, set_origin_format, format parameter must be Format enum type, otherwise throws TypeError.
• For set_data_type, data_type parameter must be DataType enum type, otherwise throws TypeError.
• If underlying C API call fails, methods will throw RuntimeError.

Usage Example

from ge.graph import TensorDesc, Shape, Format, DataType

# Create TensorDesc object
desc = TensorDesc(shape=[1, 3, 224, 224], format=Format.FORMAT_NCHW, data_type=DataType.DT_FLOAT)

# Get tensor info via properties
print(desc.shape)       # [1, 3, 224, 224]
print(desc.format)      # Format.FORMAT_NCHW
print(desc.data_type)   # DataType.DT_FLOAT

# Get tensor info via methods
shape = desc.get_shape()
fmt = desc.get_format()
dtype = desc.get_data_type()

# Chain call to set properties
desc.set_shape([2, 3]).set_data_type(DataType.DT_INT32).set_format(Format.FORMAT_ND)