"""Module for LabelTensor"""
from copy import copy, deepcopy
import torch
from torch import Tensor
[docs]
class LabelTensor(torch.Tensor):
"""
Extension of the :class:`torch.Tensor` class that includes labels for
each dimension.
"""
@staticmethod
def __new__(cls, x, labels, *args, **kwargs):
"""
Create a new instance of the :class:`~pina.label_tensor.LabelTensor`
class.
:param torch.Tensor x: :class:`torch.tensor` instance to be casted as a
:class:`~pina.label_tensor.LabelTensor`.
:param labels: Labels to assign to the tensor.
:type labels: str | list[str] | dict
:return: The instance of the :class:`~pina.label_tensor.LabelTensor`
class.
:rtype: LabelTensor
"""
if isinstance(x, LabelTensor):
return x
return super().__new__(cls, x, *args, **kwargs)
@property
def tensor(self):
"""
Returns the tensor part of the :class:`~pina.label_tensor.LabelTensor`
object.
:return: Tensor part of the :class:`~pina.label_tensor.LabelTensor`.
:rtype: torch.Tensor
"""
return self.as_subclass(Tensor)
def __init__(self, x, labels):
"""
Initialize the :class:`~pina.label_tensor.LabelTensor` instance, by
checking the consistency of the labels and the tensor. Specifically, the
labels must match the following conditions:
- At each dimension, the number of labels must match the size of the \
dimension.
- At each dimension, the labels must be unique.
The labels can be passed in the following formats:
:Example:
>>> from pina import LabelTensor
>>> tensor = LabelTensor(
>>> torch.rand((2000, 3)),
... {1: {"name": "space", "dof": ['a', 'b', 'c']}})
>>> tensor = LabelTensor(
>>> torch.rand((2000, 3)),
... ["a", "b", "c"])
The keys of the dictionary are the dimension indices, and the values are
dictionaries containing the labels and the name of the dimension. If
the labels are passed as a list, these are assigned to the last
dimension.
:param torch.Tensor x: The tensor to be casted as a
:class:`~pina.label_tensor.LabelTensor`.
:param labels: Labels to assign to the tensor.
:type labels: str | list[str] | dict
:raises ValueError: If the labels are not consistent with the tensor.
"""
super().__init__()
if labels is not None:
self.labels = labels
else:
self._labels = {}
@property
def full_labels(self):
"""
Returns the full labels of the tensor, even for the dimensions that are
not labeled.
:return: The full labels of the tensor
:rtype: dict
"""
to_return_dict = {}
shape_tensor = self.shape
for i, value in enumerate(shape_tensor):
if i in self._labels:
to_return_dict[i] = self._labels[i]
else:
to_return_dict[i] = {"dof": range(value), "name": i}
return to_return_dict
@property
def stored_labels(self):
"""
Returns the labels stored inside the instance.
:return: The labels stored inside the instance.
:rtype: dict
"""
return self._labels
@property
def labels(self):
"""
Returns the labels of the last dimension of the instance.
:return: labels of last dimension
:rtype: list
"""
if self.ndim - 1 in self._labels:
return self._labels[self.ndim - 1]["dof"]
return None
@labels.setter
def labels(self, labels):
"""
Set labels stored insider the instance by checking the type of the
input labels and handling it accordingly. The following types are
accepted:
- **list**: The list of labels is assigned to the last dimension.
- **dict**: The dictionary of labels is assigned to the tensor.
- **str**: The string is assigned to the last dimension.
:param labels: Labels to assign to the class variable _labels.
:type labels: str | list[str] | dict
"""
if not hasattr(self, "_labels"):
self._labels = {}
if isinstance(labels, dict):
self._init_labels_from_dict(labels)
elif isinstance(labels, (list, range)):
self._init_labels_from_list(labels)
elif isinstance(labels, str):
labels = [labels]
self._init_labels_from_list(labels)
else:
raise ValueError("labels must be list, dict or string.")
[docs]
def _init_labels_from_dict(self, labels):
"""
Store the internal label representation according to the values
passed as input.
:param dict labels: The label(s) to update.
:raises ValueError: If the dof list contains duplicates or the number of
dof does not match the tensor shape.
"""
tensor_shape = self.shape
def validate_dof(dof_list, dim_size):
"""Validate the 'dof' list for uniqueness and size."""
if len(dof_list) != len(set(dof_list)):
raise ValueError("dof must be unique")
if len(dof_list) != dim_size:
raise ValueError(
f"Number of dof ({len(dof_list)}) does not match "
f"tensor shape ({dim_size})"
)
for dim, label in labels.items():
if isinstance(label, dict):
if "name" not in label:
label["name"] = dim
if "dof" not in label:
label["dof"] = range(tensor_shape[dim])
if "dof" in label and "name" in label:
dof = label["dof"]
dof_list = dof if isinstance(dof, (list, range)) else [dof]
if not isinstance(dof_list, (list, range)):
raise ValueError(
f"'dof' should be a list or range, not"
f" {type(dof_list)}"
)
validate_dof(dof_list, tensor_shape[dim])
else:
raise ValueError(
"Labels dictionary must contain either "
" both 'name' and 'dof' keys"
)
else:
raise ValueError(
f"Invalid label format for {dim}: Expected "
f"list or dictionary, got {type(label)}"
)
# Assign validated label data to internal labels
self._labels[dim] = label
[docs]
def _init_labels_from_list(self, labels):
"""
Given a list of dof, this method update the internal label
representation by assigning the dof to the last dimension.
:param labels: The label(s) to update.
:type labels: list
"""
# Create a dict with labels
last_dim_labels = {
self.ndim - 1: {"dof": labels, "name": self.ndim - 1}
}
self._init_labels_from_dict(last_dim_labels)
def __str__(self):
"""
The string representation of the
:class:`~pina.label_tensor.LabelTensor`.
:return: String representation of the
:class:`~pina.label_tensor.LabelTensor` instance.
:rtype: str
"""
s = ""
for key, value in self._labels.items():
s += f"{key}: {value}\n"
s += "\n"
s += self.tensor.__str__()
return s
[docs]
@staticmethod
def cat(tensors, dim=0):
"""
Concatenate a list of tensors along a specified dimension. For more
details, see :meth:`torch.cat`.
:param list[LabelTensor] tensors:
:class:`~pina.label_tensor.LabelTensor` instances to concatenate
:param int dim: Dimensions on which you want to perform the operation
(default is 0)
:return: A new :class:`LabelTensor` instance obtained by concatenating
the input instances.
:rtype: LabelTensor
:raises ValueError: either number dof or dimensions names differ.
"""
if not tensors:
return [] # Handle empty list
if len(tensors) == 1:
return tensors[0] # Return single tensor as-is
# Perform concatenation
cat_tensor = torch.cat(tensors, dim=dim)
tensors_labels = [tensor.stored_labels for tensor in tensors]
# Check label consistency across tensors, excluding the
# concatenation dimension
for key in tensors_labels[0]:
if key != dim:
if any(
tensors_labels[i][key] != tensors_labels[0][key]
for i in range(len(tensors_labels))
):
raise RuntimeError(
f"Tensors must have the same labels along all "
f"dimensions except {dim}."
)
# Copy and update the 'dof' for the concatenation dimension
cat_labels = {k: copy(v) for k, v in tensors_labels[0].items()}
# Update labels if the concatenation dimension has labels
if dim in tensors[0].stored_labels:
if dim in cat_labels:
cat_dofs = [label[dim]["dof"] for label in tensors_labels]
cat_labels[dim]["dof"] = sum(cat_dofs, [])
else:
cat_labels = tensors[0].stored_labels
# Assign updated labels to the concatenated tensor
cat_tensor._labels = cat_labels
return cat_tensor
[docs]
@staticmethod
def stack(tensors):
"""
Stacks a list of tensors along a new dimension. For more details, see
:meth:`torch.stack`.
:param list[LabelTensor] tensors: A list of tensors to stack.
All tensors must have the same shape.
:return: A new :class:`~pina.label_tensor.LabelTensor` instance obtained
by stacking the input tensors.
:rtype: LabelTensor
"""
# Perform stacking in torch
new_tensor = torch.stack(tensors)
# Increase labels keys by 1
labels = tensors[0]._labels
labels = {key + 1: value for key, value in labels.items()}
new_tensor._labels = labels
return new_tensor
[docs]
def requires_grad_(self, mode=True):
"""
Override the :meth:`~torch.Tensor.requires_grad_` method to handle
the labels in the new tensor.
For more details, see :meth:`~torch.Tensor.requires_grad_`.
:param bool mode: A boolean value indicating whether the tensor should
track gradients.If `True`, the tensor will track gradients;
if `False`, it will not.
:return: The :class:`~pina.label_tensor.LabelTensor` itself with the
updated ``requires_grad`` state and retained labels.
:rtype: LabelTensor
"""
lt = super().requires_grad_(mode)
lt._labels = self._labels
return lt
@property
def dtype(self):
"""
Give the ``dtype`` of the tensor. For more details, see
:meth:`torch.dtype`.
:return: The data type of the tensor.
:rtype: torch.dtype
"""
return super().dtype
[docs]
def to(self, *args, **kwargs):
"""
Performs Tensor dtype and/or device conversion. For more details, see
:meth:`torch.Tensor.to`.
:return: A new :class:`~pina.label_tensor.LabelTensor` instance with the
updated dtype and/or device and retained labels.
:rtype: LabelTensor
"""
lt = super().to(*args, **kwargs)
lt._labels = self._labels
return lt
[docs]
def clone(self, *args, **kwargs):
"""
Clone the :class:`~pina.label_tensor.LabelTensor`. For more details, see
:meth:`torch.Tensor.clone`.
:return: A new :class:`~pina.label_tensor.LabelTensor` instance with the
same data and labels but allocated in a different memory location.
:rtype: LabelTensor
"""
out = LabelTensor(
super().clone(*args, **kwargs), deepcopy(self._labels)
)
return out
[docs]
def append(self, tensor, mode="std"):
"""
Appends a given tensor to the current tensor along the last dimension.
This method supports two types of appending operations:
1. **Standard append** ("std"): Concatenates the input tensor with the \
current tensor along the last dimension.
2. **Cross append** ("cross"): Creates a cross-product of the current \
tensor and the input tensor.
:param tensor: The tensor to append to the current tensor.
:type tensor: LabelTensor
:param mode: The append mode to use. Defaults to ``st``.
:type mode: str, optional
:return: A new :class:`LabelTensor` instance obtained by appending the
input tensor.
:rtype: LabelTensor
:raises ValueError: If the mode is not "std" or "cross".
"""
if mode == "std":
# Call cat on last dimension
new_label_tensor = LabelTensor.cat(
[self, tensor], dim=self.ndim - 1
)
return new_label_tensor
if mode == "cross":
# Crete tensor and call cat on last dimension
tensor1 = self
tensor2 = tensor
n1 = tensor1.shape[0]
n2 = tensor2.shape[0]
tensor1 = LabelTensor(tensor1.repeat(n2, 1), labels=tensor1.labels)
tensor2 = LabelTensor(
tensor2.repeat_interleave(n1, dim=0), labels=tensor2.labels
)
new_label_tensor = LabelTensor.cat(
[tensor1, tensor2], dim=self.ndim - 1
)
return new_label_tensor
raise ValueError('mode must be either "std" or "cross"')
[docs]
@staticmethod
def vstack(tensors):
"""
Stack tensors vertically. For more details, see :meth:`torch.vstack`.
:param list of LabelTensor label_tensors: The
:class:`~pina.label_tensor.LabelTensor` instances to stack. They
need to have equal labels.
:return: A new :class:`~pina.label_tensor.LabelTensor` instance obtained
by stacking the input tensors vertically.
:rtype: LabelTensor
"""
return LabelTensor.cat(tensors, dim=0)
# This method is used to update labels
[docs]
def _update_single_label(self, index, dim):
"""
Update the labels of the tensor based on the index (or list of indices).
:param index: Index of dof to retain.
:type index: int | slice | list[int] | tuple[int] | torch.Tensor
:param int dim: Dimension of the indexes in the original tensor.
:return: The updated labels for the specified dimension.
:rtype: list[int]
:raises: ValueError: If the index type is not supported.
"""
old_dof = self._labels[dim]["dof"]
# Handle slicing
if isinstance(index, slice):
new_dof = old_dof[index]
# Handle single integer index
elif isinstance(index, int):
new_dof = [old_dof[index]]
# Handle lists or tensors
elif isinstance(index, (list, torch.Tensor)):
# Handle list of bools
if isinstance(index, torch.Tensor) and index.dtype == torch.bool:
index = index.nonzero().squeeze()
new_dof = (
[old_dof[i] for i in index]
if isinstance(old_dof, list)
else index
)
else:
raise NotImplementedError(
f"Unsupported index type: {type(index)}. Expected slice, int, "
f"list, or torch.Tensor."
)
return new_dof
def __getitem__(self, index):
""" "
Override the __getitem__ method to handle the labels of the
:class:`~pina.label_tensor.LabelTensor` instance. It first performs
__getitem__ operation on the :class:`torch.Tensor` part of the instance,
then updates the labels based on the index.
:param index: The index used to access the item
:type index: int | str | tuple of int | list ot int | torch.Tensor
:return: A new :class:`~pina.label_tensor.LabelTensor` instance obtained
`__getitem__` operation on :class:`torch.Tensor` part of the
instance, with the updated labels.
:rtype: LabelTensor
:raises KeyError: If an invalid label index is provided.
:raises IndexError: If an invalid index is accessed in the tensor.
"""
# Handle string index
if isinstance(index, str) or (
isinstance(index, (tuple, list))
and all(isinstance(i, str) for i in index)
):
return self.extract(index)
# Retrieve selected tensor and labels
selected_tensor = super().__getitem__(index)
if not hasattr(self, "_labels"):
return selected_tensor
original_labels = self._labels
updated_labels = copy(original_labels)
# Ensure the index is iterable
if not isinstance(index, tuple):
index = [index]
# Update labels based on the index
offset = 0
removed = 0
for dim, idx in enumerate(index):
if dim in original_labels:
if isinstance(idx, int):
# Compute the working dimension considering the removed
# dimensions due to int index on a non labled dimension
dim_ = dim - removed
selected_tensor = selected_tensor.unsqueeze(dim_)
if idx != slice(None):
# Update the labels for the selected dimension
updated_labels[offset] = {
"dof": self._update_single_label(idx, dim),
"name": original_labels[dim]["name"],
}
else:
# Adjust label keys if dimension is reduced (case of integer
# index on a non-labeled dimension)
if isinstance(idx, int):
updated_labels = {
key - 1 if key > dim else key: value
for key, value in updated_labels.items()
}
removed += 1
continue
offset += 1
# Update the selected tensor's labels
selected_tensor._labels = updated_labels
return selected_tensor
[docs]
def sort_labels(self, dim=None):
"""
Sort the labels along the specified dimension and apply. It applies the
same sorting to the tensor part of the instance.
:param int dim: The dimension along which to sort the labels.
If ``None``, the last dimension is used.
:return: A new tensor with sorted labels along the specified dimension.
:rtype: LabelTensor
"""
def arg_sort(lst):
return sorted(range(len(lst)), key=lambda x: lst[x])
if dim is None:
dim = self.ndim - 1
if self.shape[dim] == 1:
return self
labels = self.stored_labels[dim]["dof"]
sorted_index = arg_sort(labels)
# Define an indexer to sort the tensor along the specified dimension
indexer = [slice(None)] * self.ndim
# Assigned the sorted index to the specified dimension
indexer[dim] = sorted_index
return self[tuple(indexer)]
def __deepcopy__(self, memo):
"""
Creates a deep copy of the object. For more details, see
:meth:`copy.deepcopy`.
:param memo: LabelTensor object to be copied.
:type memo: LabelTensor
:return: A deep copy of the original LabelTensor object.
:rtype: LabelTensor
"""
cls = self.__class__
result = cls(deepcopy(self.tensor), deepcopy(self.stored_labels))
return result
[docs]
def permute(self, *dims):
"""
Permutes the dimensions of the tensor and the associated labels
accordingly. For more details, see :meth:`torch.Tensor.permute`.
:param dims: The dimensions to permute the tensor to.
:type dims: tuple[int] | list[int]
:return: A new object with permuted dimensions and reordered labels.
:rtype: LabelTensor
"""
# Call the base class permute method
tensor = super().permute(*dims)
# Update lables
labels = self._labels
keys_list = list(*dims)
labels = {keys_list.index(k): v for k, v in labels.items()}
# Assign labels to the new tensor
tensor._labels = labels
return tensor
[docs]
def detach(self):
"""
Detaches the tensor from the computation graph and retains the stored
labels. For more details, see :meth:`torch.Tensor.detach`.
:return: A new tensor detached from the computation graph.
:rtype: LabelTensor
"""
lt = super().detach()
# Copy the labels to the new tensor only if present
if hasattr(self, "_labels"):
lt._labels = self.stored_labels
return lt
[docs]
@staticmethod
def summation(tensors):
"""
Computes the summation of a list of
:class:`~pina.label_tensor.LabelTensor` instances.
:param list[LabelTensor] tensors: A list of tensors to sum. All
tensors must have the same shape and labels.
:return: A new `LabelTensor` containing the element-wise sum of the
input tensors.
:rtype: LabelTensor
:raises ValueError: If the input `tensors` list is empty.
:raises RuntimeError: If the tensors have different shapes and/or
mismatched labels.
"""
if not tensors:
raise ValueError("The tensors list must not be empty.")
if len(tensors) == 1:
return tensors[0]
# Initialize result tensor and labels
data = torch.zeros_like(tensors[0].tensor).to(tensors[0].device)
last_dim_labels = []
# Accumulate tensors
for tensor in tensors:
data += tensor.tensor
last_dim_labels.append(tensor.labels)
# Construct last dimension labels
last_dim_labels = ["+".join(items) for items in zip(*last_dim_labels)]
# Update the labels for the resulting tensor
labels = {k: copy(v) for k, v in tensors[0].stored_labels.items()}
labels[tensors[0].ndim - 1] = {
"dof": last_dim_labels,
"name": tensors[0].name,
}
return LabelTensor(data, labels)
[docs]
def reshape(self, *shape):
"""
Override the reshape method to update the labels of the tensor.
For more details, see :meth:`torch.Tensor.reshape`.
:param tuple of int shape: The new shape of the tensor.
:return: A new :class:`~pina.label_tensor.LabelTensor` instance with the
updated shape and labels.
:rtype: LabelTensor
"""
# As for now the reshape method is used only in the context of the
# dataset, the labels are not
tensor = super().reshape(*shape)
if not hasattr(self, "_labels") or shape != (-1, *self.shape[2:]):
return tensor
tensor.labels = self.labels
return tensor
