from typing import List, Union
from . import _dispatch
from ._classes import TensorBlock, TensorMap, check_isinstance, torch_jit_is_scripting
from ._utils import (
_check_blocks_raise,
_check_same_gradients_raise,
_check_same_keys_raise,
)
[docs]
def divide(A: TensorMap, B: Union[float, int, TensorMap]) -> TensorMap:
if not torch_jit_is_scripting():
if not check_isinstance(A, TensorMap):
raise TypeError(f"`A` must be a metatensor TensorMap, not {type(A)}")
r"""Return a new :class:`TensorMap` with the values being the element-wise
division of ``A`` and ``B``.
If ``B`` is a :py:class:`TensorMap` it has to have the same metadata as ``A``.
If gradients are present in ``A``:
* ``B`` is a scalar then:
.. math::
\nabla(A / B) = \nabla A / B
* ``B`` is a :py:class:`TensorMap` with the same metadata of ``A``.
The multiplication is performed with the rule of the derivatives:
.. math::
\nabla(A / B) =(B*\nabla A-A*\nabla B)/B^2
:param A: First :py:class:`TensorMap` for the division.
:param B: Second instance for the division. Parameter can be a scalar
or a :py:class:`TensorMap`. In the latter case ``B`` must have the same
metadata of ``A``.
:return: New :py:class:`TensorMap` with the same metadata as ``A``.
"""
blocks: List[TensorBlock] = []
if torch_jit_is_scripting():
is_tensor_map = isinstance(B, TensorMap)
else:
is_tensor_map = check_isinstance(B, TensorMap)
if isinstance(B, (float, int)):
B = float(B)
for block_A in A.blocks():
blocks.append(_divide_block_constant(block=block_A, constant=B))
elif is_tensor_map:
_check_same_keys_raise(A, B, "divide")
for key, block_A in A.items():
block_B = B.block(key)
_check_blocks_raise(
block_A,
block_B,
fname="divide",
)
_check_same_gradients_raise(
block_A,
block_B,
fname="divide",
)
blocks.append(_divide_block_block(block_1=block_A, block_2=block_B))
else:
if torch_jit_is_scripting():
extra = ""
else:
extra = f", not {type(B)}"
raise TypeError("`B` must be a metatensor TensorMap or a scalar value" + extra)
return TensorMap(A.keys, blocks)
def _divide_block_constant(block: TensorBlock, constant: float) -> TensorBlock:
values = block.values / constant
result_block = TensorBlock(
values=values,
samples=block.samples,
components=block.components,
properties=block.properties,
)
for parameter, gradient in block.gradients():
if len(gradient.gradients_list()) != 0:
raise NotImplementedError("gradients of gradients are not supported")
result_block.add_gradient(
parameter=parameter,
gradient=TensorBlock(
values=gradient.values / constant,
samples=gradient.samples,
components=gradient.components,
properties=gradient.properties,
),
)
return result_block
def _divide_block_block(block_1: TensorBlock, block_2: TensorBlock) -> TensorBlock:
values = block_1.values / block_2.values
result_block = TensorBlock(
values=values,
samples=block_1.samples,
components=block_1.components,
properties=block_1.properties,
)
for parameter_1, gradient_1 in block_1.gradients():
gradient_2 = block_2.gradient(parameter_1)
if len(gradient_1.gradients_list()) != 0:
raise NotImplementedError("gradients of gradients are not supported")
if len(gradient_2.gradients_list()) != 0:
raise NotImplementedError("gradients of gradients are not supported")
assert gradient_1.values.shape == gradient_2.values.shape
assert gradient_1.samples == gradient_2.samples
_shape: List[int] = []
for c in block_1.components:
_shape.append(len(c))
_shape.append(len(block_1.properties))
# we find the difference between the number of components
# of the gradients and the values and then use it to create
# empty dimensions for broadcasting
diff_components = len(gradient_1.values.shape) - len(block_1.values.shape)
gradient_samples_to_values_samples_1 = gradient_1.samples.column("sample")
gradient_samples_to_values_samples_2 = gradient_2.samples.column("sample")
gradient_values = -block_1.values[
_dispatch.to_index_array(gradient_samples_to_values_samples_1)
].reshape(
[-1] + [1] * diff_components + _shape
) * gradient_2.values / block_2.values[
_dispatch.to_index_array(gradient_samples_to_values_samples_2)
].reshape(
[-1] + [1] * diff_components + _shape
) ** 2 + gradient_1.values / block_2.values[
_dispatch.to_index_array(gradient_samples_to_values_samples_2)
].reshape(
[-1] + [1] * diff_components + _shape
)
result_block.add_gradient(
parameter=parameter_1,
gradient=TensorBlock(
values=gradient_values,
samples=gradient_1.samples,
components=gradient_1.components,
properties=gradient_1.properties,
),
)
return result_block