Packer

class xitorch.Packer(obj: Any)

Packer is an object that could extract the tensors in a structure and rebuild the structure from the given tensors. This object preserves the structure of the object by performing the deepcopy of the object, except for the tensor.

Parameters

obj (Any) – Any structure object that contains tensors.

Example

>>> a = torch.tensor(1.0)
>>> obj = {
...     "a": a,
...     "b": a * 3,
...     "c": a,
... }
>>> packer = xitorch.Packer(obj)
>>> tensors = packer.get_param_tensor_list()
>>> print(tensors)
[tensor(1.), tensor(3.)]
>>> new_tensors = [torch.tensor(2.0), torch.tensor(4.0)]
>>> new_obj = packer.construct_from_tensor_list(new_tensors)
>>> print(new_obj)
{'a': tensor(2.), 'b': tensor(4.), 'c': tensor(2.)}

get_param_tensor_list(unique: bool = True) → List[torch.Tensor]

Returns the list of tensors contained in the object. It will traverse down the object via elements for list, values for dictionary, or __dict__ for object that has __dict__ attribute.

Parameters

unique (bool) – If True, then only returns the unique tensors. Otherwise, duplicates can also be returned.

Returns

List of tensors contained in the object.

Return type

list of torch.Tensor

get_param_tensor(unique: bool = True) → Optional[torch.Tensor]

Returns the tensor parameters as a single tensor. This can be used, for example, if there are multiple parameters to be optimized using xitorch.optimize.minimize.

Parameters

unique (bool) – If True, then only returns the tensor from unique tensors list. Otherwise, duplicates can also be returned.

Returns

The parameters of the object in a single tensor or None if there is no tensor contained in the object.

Return type

torch.Tensor or None

construct_from_tensor_list(tensors: List[torch.Tensor], unique: bool = True) → Any

Construct the object from the tensor list and returns the object structure with the new tensors. Executing this does not change the state of the Packer object.

Parameters

• tensors (list of torch.Tensor) – The tensor parameters to be filled into the object.

• unique (bool) – Indicating if the tensor list tensors is from the unique parameters of the object.

Returns

A new object with the same structure as the input to __init__ object except the tensor is changed according to tensors.

Return type

Any

construct_from_tensor(a: torch.Tensor, unique: bool = True) → Any

Construct the object from the single tensor (i.e. it is the parameters tensor merged into a single tensor) and returns the object structure with the new tensor. Executing this does not change the state of the Packer object.

Parameters

• a (torch.Tensor) – The single tensor parameter to be filled.

• unique (bool) – Indicating if the tensor a is from the unique parameters of the object.

Returns

A new object with the same structure as the input to __init__ object except the tensor is changed according to a.

Return type

Any