TensorHolder

Product Support

Product	Support
Atlas A3 Training Products/Atlas A3 Inference Products	√
Atlas A2 Training Products/Atlas A2 Inference Products	√

Module Import

from ge.es import TensorHolder

Functionality Description

TensorHolder is tensor holder class in Eager-Style graph building, created by GraphBuilder's create_* methods (like create_input(), create_const_float() etc.). This class supports operator overloading (+, -, *, /), can set data type, format and shape via chained calls.

TensorHolder automatically maintains strong reference to its owning GraphBuilder, ensuring underlying C++ resources are not released during TensorHolder lifetime.

Constraints

• Does not support direct instantiation: TensorHolder objects can only be created via GraphBuilder's create_* methods, directly calling constructor will throw RuntimeError.
• Cannot call setter methods after build_and_reset(): After GraphBuilder.build_and_reset() executes, calling set_data_type(), set_format(), set_shape() etc. setter methods will cause errors.
• Operations require same GraphBuilder: When performing tensor operations (add, sub, mul, div), both participating TensorHolder must belong to same GraphBuilder, otherwise will throw ValueError.

---

name Property

Gets producer node name.

Function Prototype

@property
def name(self) -> str:
    ...

Parameter Description

No parameters.

Return Value Description

Type	Description
str	Returns producer node's name.

---

set_data_type Method

Sets tensor data type.

Function Prototype

def set_data_type(self, data_type: DataType) -> TensorHolder:
    ...

Parameter Description

Parameter	Input/Output	Description
data_type	Input	Data type, type is ge.graph.types.DataType enum.

Return Value Description

Type	Description
TensorHolder	Returns current TensorHolder object reference, supports chained calls.

Constraints

• data_type must be DataType enum type, otherwise will throw TypeError.

---

set_format Method

Sets tensor data format.

Function Prototype

def set_format(self, format: Format) -> TensorHolder:
    ...

Parameter Description

Parameter	Input/Output	Description
format	Input	Data format, type is ge.graph.types.Format enum.

Return Value Description

Type	Description
TensorHolder	Returns current TensorHolder object reference, supports chained calls.

Constraints

• format must be Format enum type, otherwise will throw TypeError.

---

set_shape Method

Sets tensor shape.

Function Prototype

def set_shape(self, shape: List[int]) -> TensorHolder:
    ...

Parameter Description

Parameter	Input/Output	Description
shape	Input	Shape dimension list, type is integer list List[int].

Return Value Description

Type	Description
TensorHolder	Returns current TensorHolder object reference, supports chained calls.

Constraints

• shape must be integer list, and all elements must be int type, otherwise will throw TypeError.

---

add Method

Tensor addition operation.

Function Prototype

def add(self, other: Union[TensorHolder, TensorLike]) -> TensorHolder:
    ...

Parameter Description

Parameter	Input/Output	Description
other	Input	Another tensor, type is TensorHolder or TensorLike (scalar/array etc. convertible types).

Return Value Description

Type	Description
TensorHolder	Returns new TensorHolder object representing addition operation result.

Constraints

• If other is TensorHolder, must belong to same GraphBuilder as current tensor.
• Operation library (libes_math.so or default generated library) must be loadable, otherwise will throw RuntimeError.

---

sub Method

Tensor subtraction operation.

Function Prototype

def sub(self, other: Union[TensorHolder, TensorLike]) -> TensorHolder:
    ...

Parameter Description

Parameter	Input/Output	Description
other	Input	Another tensor, type is TensorHolder or TensorLike (scalar/array etc. convertible types).

Return Value Description

Type	Description
TensorHolder	Returns new TensorHolder object representing subtraction operation result.

Constraints

• If other is TensorHolder, must belong to same GraphBuilder as current tensor.
• Operation library (libes_math.so or default generated library) must be loadable, otherwise will throw RuntimeError.

---

mul Method

Tensor multiplication operation.

Function Prototype

def mul(self, other: Union[TensorHolder, TensorLike]) -> TensorHolder:
    ...

Parameter Description

Parameter	Input/Output	Description
other	Input	Another tensor, type is TensorHolder or TensorLike (scalar/array etc. convertible types).

Return Value Description

Type	Description
TensorHolder	Returns new TensorHolder object representing multiplication operation result.

Constraints

• If other is TensorHolder, must belong to same GraphBuilder as current tensor.
• Operation library (libes_math.so or default generated library) must be loadable, otherwise will throw RuntimeError.

---

div Method

Tensor division operation.

Function Prototype

def div(self, other: Union[TensorHolder, TensorLike]) -> TensorHolder:
    ...

Parameter Description

Parameter	Input/Output	Description
other	Input	Another tensor, type is TensorHolder or TensorLike (scalar/array etc. convertible types).

Return Value Description

Type	Description
TensorHolder	Returns new TensorHolder object representing division operation result.

Constraints

• If other is TensorHolder, must belong to same GraphBuilder as current tensor.
• Operation library (libes_math.so or default generated library) must be loadable, otherwise will throw RuntimeError.

---

Operator Overloading

TensorHolder supports the following Python operator overloading with corresponding relationships:

Operator	Corresponding Method	Description
a + b	a.add(b)	Tensor addition
a - b	a.sub(b)	Tensor subtraction
a * b	a.mul(b)	Tensor multiplication
a / b	a.div(b)	Tensor division

Also supports right operand operations (__radd__, __rsub__, __rmul__, __rtruediv__), for handling operations with non-TensorHolder type on left side.

---

get_owner_builder Method

Gets owning GraphBuilder.

Function Prototype

def get_owner_builder(self) -> GraphBuilder:
    ...

Parameter Description

No parameters.

Return Value Description

Type	Description
GraphBuilder	Returns GraphBuilder object that created this TensorHolder.