import copy
import ctypes
import pathlib
import warnings
from pickle import PickleBuffer
from typing import BinaryIO, Dict, List, Sequence, Union
import numpy as np
from . import data
from ._c_api import c_uintptr_t, mts_block_t, mts_labels_t
from ._c_lib import _get_library
from .block import TensorBlock
from .data import Device, DeviceWarning, DType
from .labels import Labels, LabelsEntry
from .status import _check_pointer
from .utils import _to_arguments_parse
[docs]
class TensorMap:
"""
A TensorMap is the main user-facing class of this library, and can store any kind of
data used in atomistic machine learning similar to a Python :py:class:`dict`.
A tensor map contains a list of :py:class:`TensorBlock`, each one associated with a
key. Blocks can either be accessed one by one with the :py:func:`TensorMap.block`
function, or by iterating over the tensor map itself:
.. code-block:: python
for block in tensor:
...
The corresponding keys can be included in the loop by using the ``items()`` method
of a :py:func:`TensorMap`:
.. code-block:: python
for key, block in tensor.items():
...
A tensor map provides functions to move some of these keys to the samples or
properties labels of the blocks, moving from a sparse representation of the data to
a dense one.
"""
def __init__(self, keys: Labels, blocks: Sequence[TensorBlock]):
"""
:param keys: keys associated with each block
:param blocks: set of blocks containing the actual data
"""
if not isinstance(keys, Labels):
raise TypeError(f"`keys` must be metatensor Labels, not {type(keys)}")
blocks = list(blocks)
for block in blocks:
if not isinstance(block, TensorBlock):
raise TypeError(
"`blocks` elements must be metatensor TensorBlock, "
f"not {type(block)}"
)
self._lib = _get_library()
blocks_array_t = ctypes.POINTER(mts_block_t) * len(blocks)
blocks_array = blocks_array_t(*[block._ptr for block in blocks])
for block in blocks:
if block._parent is not None:
raise ValueError(
"can not use blocks from another TensorMap in a new one, "
"use TensorBlock.copy() to make a copy of each block first"
)
block_origin = data.data_origin(block._raw_values)
first_block_origin = data.data_origin(blocks[0]._raw_values)
if block_origin != first_block_origin:
raise ValueError(
"all blocks in a TensorMap must have the same origin, "
f"got '{data.data_origin_name(first_block_origin)}' "
f"and '{data.data_origin_name(block_origin)}'"
)
if block.device != blocks[0].device:
raise ValueError(
"all blocks in a TensorMap must have the same device, "
f"got '{blocks[0].device}' and '{block.device}'"
)
if block.dtype != blocks[0].dtype:
raise ValueError(
"all blocks in a TensorMap must have the same dtype, "
f"got {blocks[0].dtype} and {block.dtype}"
)
if len(blocks) > 0 and not data.array_device_is_cpu(blocks[0].values):
warnings.warn(
"Blocks values and keys for this TensorMap are on different devices: "
f"keys are always on CPU, and blocks values are on device "
f"'{blocks[0].device}'. If you are using PyTorch and need the labels "
f"to also be on {blocks[0].device}, you should use "
"`metatensor.torch.TensorMap`.",
category=DeviceWarning,
stacklevel=2,
)
# all blocks are moved into the tensor map, assign NULL to `block._ptr` to
# prevent accessing invalid data from Python and double free
for block in blocks:
block._move_ptr()
self._ptr = self._lib.mts_tensormap(
keys._as_mts_labels_t(), blocks_array, len(blocks)
)
_check_pointer(self._ptr)
for block in blocks:
block._is_inside_map = True
@staticmethod
def _from_ptr(ptr):
"""Create a tensor map from a pointer owning its data"""
_check_pointer(ptr)
obj = TensorMap.__new__(TensorMap)
obj._lib = _get_library()
obj._ptr = ptr
obj._blocks = []
return obj
def __del__(self):
if hasattr(self, "_lib") and self._lib is not None and hasattr(self, "_ptr"):
self._lib.mts_tensormap_free(self._ptr)
def __copy__(self):
raise ValueError(
"shallow copies of TensorMap are not possible, use a deepcopy instead"
)
def __deepcopy__(self, _memodict):
new_ptr = self._lib.mts_tensormap_copy(self._ptr)
return TensorMap._from_ptr(new_ptr)
[docs]
def copy(self) -> "TensorMap":
"""
Get a deep copy of this TensorMap, including all the data and metadata
"""
return copy.deepcopy(self)
def __len__(self):
return len(self.keys)
def __repr__(self) -> str:
return self.print(4)
def __str__(self) -> str:
return self.print(-1)
def __getitem__(self, selection) -> TensorBlock:
"""This is equivalent to self.block(selection)"""
return self.block(selection)
# ===== Serialization support ===== #
@classmethod
def _from_pickle(cls, buffer: Union[bytes, bytearray]):
"""
Passed to pickler to reconstruct TensorMap from bytes object
"""
from .io import create_numpy_array, load_buffer_custom_array
# TODO: make it so when saving data in torch tensors, we load back data in torch
# tensors.
return load_buffer_custom_array(buffer, create_numpy_array)
def __reduce_ex__(self, protocol: int):
"""
Used by the Pickler to dump TensorMap object to bytes object. When protocol >= 5
it supports PickleBuffer which reduces number of copies needed
"""
from .io import _save_tensor_buffer_raw
buffer = _save_tensor_buffer_raw(self)
if protocol >= 5:
return self._from_pickle, (PickleBuffer(buffer),)
else:
return self._from_pickle, (buffer.raw,)
[docs]
@staticmethod
def load(file: Union[str, pathlib.Path, BinaryIO], use_numpy=False) -> "TensorMap":
"""
Load a serialized :py:class:`TensorMap` from a file or a buffer, calling
:py:func:`metatensor.load`.
:param file: file path or file object to load from
:param use_numpy: should we use the numpy loader or metatensor's. See
:py:func:`metatensor.load` for more information.
"""
from .io import load
return load(file=file, use_numpy=use_numpy)
[docs]
@staticmethod
def load_buffer(
buffer: Union[bytes, bytearray, memoryview],
use_numpy=False,
) -> "TensorMap":
"""
Load a serialized :py:class:`TensorMap` from a buffer, calling
:py:func:`metatensor.io.load_buffer`.
:param buffer: in-memory buffer containing the data
:param use_numpy: should we use the numpy loader or metatensor's. See
:py:func:`metatensor.load` for more information.
"""
from .io import load_buffer
return load_buffer(buffer=buffer)
[docs]
def save(self, file: Union[str, pathlib.Path, BinaryIO], use_numpy=False):
"""
Save this :py:class:`TensorMap` to a file or a buffer, calling
:py:func:`metatensor.save`.
:param file: file path or file object to save to
:param use_numpy: should we use the numpy serializer or metatensor's. See
:py:func:`metatensor.save` for more information.
"""
from .io import save
return save(file=file, data=self, use_numpy=use_numpy)
[docs]
def save_buffer(self, use_numpy=False) -> memoryview:
"""
Save this :py:class:`TensorMap` to an in-memory buffer, calling
:py:func:`metatensor.io.save_buffer`.
:param use_numpy: should we use numpy serialization or metatensor's. See
:py:func:`metatensor.save` for more information.
"""
from .io import save_buffer
return save_buffer(data=self, use_numpy=use_numpy)
# ===== Math functions, implemented using metatensor-operations ===== #
def __eq__(self, other):
from metatensor.operations import equal
return equal(self, other)
def __ne__(self, other):
from metatensor.operations import equal
return not equal(self, other)
def __add__(self, other):
from metatensor.operations import add
return add(self, other)
def __sub__(self, other):
from metatensor.operations import subtract
return subtract(self, other)
def __mul__(self, other):
from metatensor.operations import multiply
return multiply(self, other)
def __matmul__(self, other):
from metatensor.operations import dot
return dot(self, other)
def __truediv__(self, other):
from metatensor.operations import divide
return divide(self, other)
def __pow__(self, other):
from metatensor.operations import pow
return pow(self, other)
def __neg__(self):
from metatensor.operations import multiply
return multiply(self, -1)
def __pos__(self):
return self
# ===== Data manipulation ===== #
@property
def keys(self) -> Labels:
"""The set of keys labeling the blocks in this tensor map."""
result = mts_labels_t()
self._lib.mts_tensormap_keys(self._ptr, result)
return Labels._from_mts_labels_t(result)
[docs]
def block_by_id(self, index: int) -> TensorBlock:
"""
Get the block at ``index`` in this :py:class:`TensorMap`.
:param index: index of the block to retrieve
"""
if index >= len(self):
# we need to raise IndexError to make sure TensorMap supports iterations
# over blocks with `for block in tensor:` which calls `__getitem__` with
# integers from 0 to whenever IndexError is raised.
raise IndexError(
f"block index out of bounds: we have {len(self)} blocks but the "
f"index is {index}"
)
block = ctypes.POINTER(mts_block_t)()
self._lib.mts_tensormap_block_by_id(self._ptr, block, index)
return TensorBlock._from_ptr(block, parent=self)
[docs]
def blocks_by_id(self, indices: Sequence[int]) -> TensorBlock:
"""
Get the blocks with the given ``indices`` in this :py:class:`TensorMap`.
:param indices: indices of the block to retrieve
"""
return [self.block_by_id(i) for i in indices]
[docs]
def blocks_matching(self, selection: Labels) -> List[int]:
"""
Get a (possibly empty) list of block indexes matching the ``selection``.
This function finds all keys in this :py:class:`TensorMap` with the same values
as ``selection`` for the dimensions/names contained in the ``selection``; and
return the corresponding indexes.
The ``selection`` should contain a single entry.
"""
block_indexes = ctypes.ARRAY(c_uintptr_t, len(self.keys))()
count = c_uintptr_t(block_indexes._length_)
self._lib.mts_tensormap_blocks_matching(
self._ptr,
block_indexes,
count,
selection._as_mts_labels_t(),
)
result = []
for i in range(count.value):
result.append(int(block_indexes[i]))
return result
[docs]
def block(
self,
selection: Union[None, int, Labels, LabelsEntry, Dict[str, int]] = None,
**kwargs,
) -> TensorBlock:
"""
Get the single block in this :py:class:`TensorMap` matching the ``selection``.
When ``selection`` is an ``int``, this is equivalent to
:py:func:`TensorMap.block_by_id`.
When ``selection`` is an :py:class:`Labels`, it should only contain a single
entry, which will be used for the selection.
When ``selection`` is a ``Dict[str, int]``, it is converted into a single single
:py:class:`LabelsEntry` (the dict keys becoming the names and the dict values
being joined together to form the :py:class:`LabelsEntry` values), which is then
used for the selection.
When ``selection`` is a :py:class:`LabelsEntry`, this function finds the key in
this :py:class:`TensorMap` with the same values as ``selection`` for the
dimensions/names contained in the ``selection``; and return the corresponding
indexes.
If ``selection`` is :py:obj:`None`, the selection can be passed as keyword
arguments, which will be converted to a ``Dict[str, int]``.
:param selection: description of the block to extract
>>> from metatensor import TensorMap, TensorBlock, Labels
>>> keys = Labels(["key_1", "key_2"], np.array([[0, 0], [6, 8]]))
>>> block_1 = TensorBlock(
... values=np.full((3, 5), 1.0),
... samples=Labels.range("sample", 3),
... components=[],
... properties=Labels.range("property", 5),
... )
>>> block_2 = TensorBlock(
... values=np.full((5, 3), 2.0),
... samples=Labels.range("sample", 5),
... components=[],
... properties=Labels.range("property", 3),
... )
>>> tensor = TensorMap(keys, [block_1, block_2])
>>> # numeric index selection, this gives a block by its position
>>> block = tensor.block(0)
>>> block
TensorBlock
samples (3): ['sample']
components (): []
properties (5): ['property']
gradients: None
>>> # This is the first block
>>> block.values.mean()
1.0
>>> # use a single key entry (i.e. LabelsEntry) for the selection
>>> tensor.block(tensor.keys[0]).values.mean()
1.0
>>> # Labels with a single entry selection
>>> labels = Labels(names=["key_1", "key_2"], values=np.array([[6, 8]]))
>>> tensor.block(labels).values.mean()
2.0
>>> # keyword arguments selection
>>> tensor.block(key_1=0, key_2=0).values.mean()
1.0
>>> # dictionary selection
>>> tensor.block({"key_1": 6, "key_2": 8}).values.mean()
2.0
"""
if selection is None:
return self.block(kwargs)
elif isinstance(selection, int):
return self.block_by_id(selection)
else:
selection = _normalize_selection(selection)
matching = self.blocks_matching(selection)
if len(matching) == 0:
if len(self.keys) == 0:
raise ValueError("there are no blocks in this TensorMap")
else:
raise ValueError(
f"couldn't find any block matching {selection[0].print()}"
)
elif len(matching) > 1:
raise ValueError(
f"more than one block matched {selection[0].print()}, "
"use `TensorMap.blocks` to get all of them"
)
else:
return self.block_by_id(matching[0])
[docs]
def blocks(
self,
selection: Union[
None, Sequence[int], int, Labels, LabelsEntry, Dict[str, int]
] = None,
**kwargs,
) -> List[TensorBlock]:
"""
Get the blocks in this :py:class:`TensorMap` matching the ``selection``.
When ``selection`` is ``None`` (the default), all blocks are returned.
When ``selection`` is an ``int``, this is equivalent to
:py:func:`TensorMap.block_by_id`; and for a ``List[int]`` this is equivalent to
:py:func:`TensorMap.blocks_by_id`.
When ``selection`` is an :py:class:`Labels`, it should only contain a single
entry, which will be used for the selection.
When ``selection`` is a ``Dict[str, int]``, it is converted into a single single
:py:class:`LabelsEntry` (the dict keys becoming the names and the dict values
being joined together to form the :py:class:`LabelsEntry` values), which is then
used for the selection.
When ``selection`` is a :py:class:`LabelsEntry`, this function finds all keys in
this :py:class:`TensorMap` with the same values as ``selection`` for the
dimensions/names contained in the ``selection``; and return the corresponding
blocks.
If ``selection`` is :py:obj:`None`, the selection can be passed as keyword
arguments, which will be converted to a ``Dict[str, int]``.
:param selection: description of the blocks to extract
"""
if selection is None:
return self.blocks(kwargs)
elif isinstance(selection, int):
return [self.block_by_id(selection)]
else:
selection = _normalize_selection(selection)
matching = self.blocks_matching(selection)
if len(self.keys) == 0:
# return an empty list here instead of the top of this function to make sure
# the selection was validated
return []
if len(matching) == 0:
raise ValueError(
f"Couldn't find any block matching '{selection[0].print()}'"
)
else:
return self.blocks_by_id(matching)
[docs]
def items(self):
"""get an iterator over (key, block) pairs in this :py:class:`TensorMap`"""
keys = self.keys
for i, key in enumerate(keys):
yield key, self.block_by_id(i)
[docs]
def keys_to_samples(
self,
keys_to_move: Union[str, Sequence[str]],
*,
sort_samples=True,
) -> "TensorMap":
"""
Merge blocks along the samples axis, adding ``keys_to_move`` to the end of the
samples labels dimensions.
This function will remove ``keys_to_move`` from the keys, and find all blocks
with the same remaining keys values. It will then merge these blocks along the
samples direction (i.e. do a *vertical* concatenation), adding ``keys_to_move``
to the end of the samples labels dimensions. The values taken by
``keys_to_move`` in the new samples labels will be the values of these
dimensions in the merged blocks' keys.
If ``keys_to_move`` is a set of :py:class:`Labels`, it must be empty
(``keys_to_move.values.shape[0] == 0``), and only the :py:class:`Labels.names`
will be used.
The order of the samples is controlled by ``sort_samples``. If ``sort_samples``
is true, samples are re-ordered to keep them lexicographically sorted. Otherwise
they are kept in the order in which they appear in the blocks.
This function is only implemented when the blocks to merge have the same
properties values.
:param keys_to_move: description of the keys to move
:param sort_samples: whether to sort the merged samples or keep them in the
order in which they appear in the original blocks
:return: a new :py:class:`TensorMap` with merged blocks
"""
keys_to_move = _normalize_keys_to_move(keys_to_move)
ptr = self._lib.mts_tensormap_keys_to_samples(
self._ptr, keys_to_move._as_mts_labels_t(), sort_samples
)
return TensorMap._from_ptr(ptr)
[docs]
def components_to_properties(
self, dimensions: Union[str, Sequence[str]]
) -> "TensorMap":
"""
Move the given ``dimensions`` from the component labels to the property labels
for each block.
:param dimensions: name of the component dimensions to move to the properties
"""
c_dimensions = _list_or_str_to_array_c_char(dimensions)
ptr = self._lib.mts_tensormap_components_to_properties(
self._ptr, c_dimensions, c_dimensions._length_
)
return TensorMap._from_ptr(ptr)
[docs]
def keys_to_properties(
self,
keys_to_move: Union[str, Sequence[str], Labels],
*,
sort_samples=True,
) -> "TensorMap":
"""
Merge blocks along the properties direction, adding ``keys_to_move`` at the
beginning of the properties labels dimensions.
This function will remove ``keys_to_move`` from the keys, and find all blocks
with the same remaining keys values. Then it will merge these blocks along the
properties direction (i.e. do a *horizontal* concatenation).
If ``keys_to_move`` is given as strings, then the new property labels will
**only** contain entries from the existing blocks. For example, merging a block
with key ``a=0`` and properties ``p=1, 2`` with a block with key ``a=2`` and
properties ``p=1, 3`` will produce a block with properties ``a, p = (0, 1), (0,
2), (2, 1), (2, 3)``.
If ``keys_to_move`` is a set of :py:class:`Labels` and it is empty
(``len(keys_to_move) == 0``), the :py:class:`Labels.names` will be used as if
they where passed directly.
Finally, if ``keys_to_move`` is a non empty set of :py:class:`Labels`, the new
properties labels will contains **all** of the entries of ``keys_to_move``
(regardless of the values taken by ``keys_to_move.names`` in the merged blocks'
keys) followed by the existing properties labels. For example, using ``a=2, 3``
in ``keys_to_move``, blocks with properties ``p=1, 2`` will result in ``a, p =
(2, 1), (2, 2), (3, 1), (3, 2)``. If there is no values (no block/missing
sample) for a given property in the merged block, then the value will be set to
zero.
When using a non empty :py:class:`Labels` for ``keys_to_move``, the properties
labels of all the merged blocks must take the same values.
The order of the samples in the merged blocks is controlled by ``sort_samples``.
If ``sort_samples`` is :py:obj:`True`, samples are re-ordered to keep them
lexicographically sorted. Otherwise they are kept in the order in which they
appear in the blocks.
:param keys_to_move: description of the keys to move
:param sort_samples: whether to sort the merged samples or keep them in the
order in which they appear in the original blocks
:return: a new :py:class:`TensorMap` with merged blocks
"""
keys_to_move = _normalize_keys_to_move(keys_to_move)
ptr = self._lib.mts_tensormap_keys_to_properties(
self._ptr, keys_to_move._as_mts_labels_t(), sort_samples
)
return TensorMap._from_ptr(ptr)
@property
def sample_names(self) -> List[str]:
"""
names of the samples dimensions for all blocks in this :py:class:`TensorMap`
"""
if len(self.keys) == 0:
return []
return self.block(0).samples.names
@property
def component_names(self) -> List[str]:
"""
names of the components dimensions for all blocks in this :py:class:`TensorMap`
"""
if len(self.keys) == 0:
return []
return [c.names[0] for c in self.block(0).components]
@property
def property_names(self) -> List[str]:
"""
names of the properties dimensions for all blocks in this :py:class:`TensorMap`
"""
if len(self.keys) == 0:
return []
return self.block(0).properties.names
[docs]
def print(self, max_keys: int) -> str:
"""
Print this :py:class:`TensorMap` to a string, including at most ``max_keys`` in
the output.
:param max_keys: how many keys to include in the output. Use ``-1`` to include
all keys.
"""
result = f"TensorMap with {len(self)} blocks\nkeys:"
result += self.keys.print(max_entries=max_keys, indent=5)
return result
@property
def device(self) -> Device:
"""get the device of all the arrays stored inside this :py:class:`TensorMap`"""
if len(self.keys) == 0:
return "cpu"
return self.block_by_id(0).device
@property
def dtype(self) -> DType:
"""get the dtype of all the arrays stored inside this :py:class:`TensorMap`"""
if len(self.keys) == 0:
return None
return self.block_by_id(0).dtype
[docs]
def to(self, *args, **kwargs) -> "TensorMap":
"""
Move the keys and all the blocks in this :py:class:`TensorMap` to the given
``dtype``, ``device`` and ``arrays`` backend.
The arguments to this function can be given as positional or keyword arguments.
:param dtype: new dtype to use for all arrays. The dtype stays the same if this
is set to ``None``.
:param device: new device to use for all arrays. The device stays the same if
this is set to ``None``.
:param arrays: new backend to use for the arrays. This can be either
``"numpy"``, ``"torch"`` or ``None`` (keeps the existing backend); and must
be given as a keyword argument (``arrays="numpy"``).
"""
arrays = kwargs.pop("arrays", None)
dtype, device = _to_arguments_parse("`TensorMap.to`", *args, **kwargs)
blocks = []
with warnings.catch_warnings():
# do not warn on device mismatch between values/labels here,
# there will be a warning when constructing the TensorMap
warnings.simplefilter("ignore", DeviceWarning)
for block in self.blocks():
blocks.append(block.to(dtype=dtype, device=device, arrays=arrays))
return TensorMap(self.keys, blocks)
def _normalize_keys_to_move(keys_to_move: Union[str, Sequence[str], Labels]) -> Labels:
if isinstance(keys_to_move, str):
keys_to_move = (keys_to_move,)
if not isinstance(keys_to_move, Labels):
for key in keys_to_move:
assert isinstance(key, str)
keys_to_move = Labels(
names=keys_to_move,
values=np.zeros((0, len(keys_to_move))),
)
return keys_to_move
def _list_or_str_to_array_c_char(strings: Union[str, Sequence[str]]):
if isinstance(strings, str):
strings = [strings]
c_strings = ctypes.ARRAY(ctypes.c_char_p, len(strings))()
for i, v in enumerate(strings):
assert isinstance(v, str)
c_strings[i] = v.encode("utf8")
return c_strings
def _can_cast_to_numpy_int(value):
return np.can_cast(value, np.int32, casting="same_kind")
def _normalize_selection(
selection: Union[Labels, LabelsEntry, Dict[str, int]]
) -> Labels:
if isinstance(selection, dict):
for key, value in selection.items():
if isinstance(value, int):
# all good
pass
elif isinstance(value, float) or not _can_cast_to_numpy_int(value):
raise TypeError(
f"expected integer values in selection, got {key}={value} of "
f"type {type(value)}"
)
if len(selection) == 0:
return Labels([], np.empty((0, 0), dtype=np.int32))
else:
return Labels(
list(selection.keys()),
np.array([[np.int32(v) for v in selection.values()]], dtype=np.int32),
)
elif isinstance(selection, Labels):
return selection
elif isinstance(selection, LabelsEntry):
return Labels(selection.names, selection.values.reshape(1, -1))
else:
raise TypeError(f"invalid type for block selection: {type(selection)}")