from copy import deepcopy
from typing import List, Optional, Union
import torch
from metatensor.operations import _dispatch
from .._backend import Labels, LabelsEntry, TensorBlock, TensorMap
from ._module import Module
@torch.jit.interface
class ModuleMapInterface(torch.nn.Module):
"""
This interface required for TorchScript to index the :py:class:`torch.nn.ModuleDict`
with non-literals in ModuleMap. Any module that is used with ModuleMap must
implement this interface to be TorchScript compilable.
Note that the *typings and argument names must match exactly* so that an interface
is correctly implemented.
Reference
---------
https://github.com/pytorch/pytorch/pull/45716
"""
def forward(self, input: torch.Tensor) -> torch.Tensor:
pass
[docs]
class ModuleMap(Module):
"""
A class that imitates :py:class:`torch.nn.ModuleDict`. In its forward function the
module at position `i` given on construction by :param modules: is applied to the
tensor block that corresponding to the`i`th key in :param in_keys:.
:param in_keys:
A :py:class:`metatensor.Labels` object with the keys of the module map that are
assumed to be in the input tensor map in the :py:meth:`forward` function.
:param modules:
A sequence of modules applied in the :py:meth:`forward` function on the input
:py:class:`TensorMap`. Each module corresponds to one :py:class:`LabelsEntry` in
:param in_keys: that determines on which :py:class:`TensorBlock` the module is
applied on. :param modules: and :param in_keys: must match in length.
:param out_properties:
A list of labels that is used to determine the properties labels of the
output. Because a module could change the number of properties, the labels of
the properties cannot be persevered. By default the output properties are
relabeled using Labels.range with "_" as key.
>>> import torch
>>> import numpy as np
>>> from copy import deepcopy
>>> from metatensor import Labels, TensorBlock, TensorMap
>>> from metatensor.learn.nn import ModuleMap
Create simple block
>>> block_1 = TensorBlock(
... values=torch.tensor(
... [
... [1.0, 2.0, 4.0],
... [3.0, 5.0, 6.0],
... ]
... ),
... samples=Labels(
... ["system", "atom"],
... np.array(
... [
... [0, 0],
... [0, 1],
... ]
... ),
... ),
... components=[],
... properties=Labels(["properties"], np.array([[0], [1], [2]])),
... )
>>> block_2 = TensorBlock(
... values=torch.tensor(
... [
... [5.0, 8.0, 2.0],
... [1.0, 2.0, 8.0],
... ]
... ),
... samples=Labels(
... ["system", "atom"],
... np.array(
... [
... [0, 0],
... [0, 1],
... ]
... ),
... ),
... components=[],
... properties=Labels(["properties"], np.array([[3], [4], [5]])),
... )
>>> keys = Labels(names=["key"], values=np.array([[0], [1]]))
>>> tensor = TensorMap(keys, [block_1, block_2])
Create modules
>>> linear = torch.nn.Linear(3, 1, bias=False)
>>> with torch.no_grad():
... _ = linear.weight.copy_(torch.tensor([1.0, 1.0, 1.0]))
>>> modules = [linear, deepcopy(linear)]
>>> # you could also extend the module by some nonlinear activation function
Create ModuleMap from this ModucDict and apply it
>>> module_map = ModuleMap(tensor.keys, modules)
>>> out = module_map(tensor)
>>> out
TensorMap with 2 blocks
keys: key
0
1
>>> out[0].values
tensor([[ 7.],
[14.]], grad_fn=<MmBackward0>)
>>> out[1].values
tensor([[15.],
[11.]], grad_fn=<MmBackward0>)
Let's look at the metadata
>>> tensor[0]
TensorBlock
samples (2): ['system', 'atom']
components (): []
properties (3): ['properties']
gradients: None
>>> out[0]
TensorBlock
samples (2): ['system', 'atom']
components (): []
properties (1): ['_']
gradients: None
It got completely lost because we cannot know in general what the output is.
You can add in the initialization of the ModuleMap a TensorMap that contains
the intended output Labels.
"""
def __init__(
self,
in_keys: Labels,
modules: List[Module],
out_properties: Optional[List[Labels]] = None,
):
super().__init__()
self.module_list = torch.nn.ModuleList(modules)
if len(in_keys) != len(modules):
raise ValueError(
"in_keys and modules must match in length, but found "
f"{len(in_keys) != len(modules)} [len(in_keys) != len(modules)]"
)
self._in_keys: Labels = in_keys
self._out_properties = out_properties
[docs]
@classmethod
def from_module(
cls,
in_keys: Labels,
module: Module,
out_properties: Optional[List[Labels]] = None,
):
"""
A wrapper around one :py:class:`torch.nn.Module` or
:py:class:`metatensor.learn.nn.Module` applying the same type of module on each
tensor block.
:param in_keys:
A :py:class:`metatensor.Labels` object that determines the keys of the
module map that are ass the TensorMaps that are assumed to be in the input
tensor map in the :py:meth:`forward` function.
:param module:
The module that is applied on each block.
:param out_properties:
A list of labels that is used to determine the properties labels of the
output. Because a module could change the number of properties, the labels
of the properties cannot be persevered. By default the output properties are
relabeled using Labels.range with "_" as key.
>>> import torch
>>> import numpy as np
>>> from metatensor import Labels, TensorBlock, TensorMap
>>> block_1 = TensorBlock(
... values=torch.tensor(
... [
... [1.0, 2.0, 4.0],
... [3.0, 5.0, 6.0],
... ]
... ),
... samples=Labels(["system", "atom"], np.array([[0, 0], [0, 1]])),
... components=[],
... properties=Labels(["properties"], np.array([[0], [1], [2]])),
... )
>>> block_2 = TensorBlock(
... values=torch.tensor(
... [
... [5.0, 8.0, 2.0],
... [1.0, 2.0, 8.0],
... ]
... ),
... samples=Labels(["system", "atom"], np.array([[0, 0], [0, 1]])),
... components=[],
... properties=Labels(["properties"], np.array([[0], [1], [2]])),
... )
>>> keys = Labels(names=["key"], values=np.array([[0], [1]]))
>>> tensor = TensorMap(keys, [block_1, block_2])
>>> linear = torch.nn.Linear(3, 1, bias=False)
>>> with torch.no_grad():
... _ = linear.weight.copy_(torch.tensor([1.0, 1.0, 1.0]))
>>> # you could also extend the module by some nonlinear activation function
>>> from metatensor.learn.nn import ModuleMap
>>> module_map = ModuleMap.from_module(tensor.keys, linear)
>>> out = module_map(tensor)
>>> out[0].values
tensor([[ 7.],
[14.]], grad_fn=<MmBackward0>)
>>> out[1].values
tensor([[15.],
[11.]], grad_fn=<MmBackward0>)
"""
module = deepcopy(module)
modules = []
for _ in range(len(in_keys)):
modules.append(deepcopy(module))
return cls(in_keys, modules, out_properties)
[docs]
def forward(self, tensor: TensorMap) -> TensorMap:
"""
Apply the modules on each block in ``tensor``. ``tensor`` must have the same
set of keys as the modules used to initialize this :py:class:`ModuleMap`.
:param tensor:
input tensor map
"""
out_blocks: List[TensorBlock] = []
for key, block in tensor.items():
out_block = self._forward_block(key, block)
for parameter, gradient in block.gradients():
if len(gradient.gradients_list()) != 0:
raise NotImplementedError(
"gradients of gradients are not supported"
)
out_block.add_gradient(
parameter=parameter,
gradient=self._forward_block(key, gradient),
)
out_blocks.append(out_block)
return TensorMap(tensor.keys, out_blocks)
def _forward_block(self, key: LabelsEntry, block: TensorBlock) -> TensorBlock:
module_idx = self.in_keys.position(key)
if module_idx is None:
raise KeyError(f"key {key} not found in modules.")
module: ModuleMapInterface = self.module_list[module_idx]
out_values = module.forward(block.values)
if self._out_properties is None:
# we do not use `Labels.range` because of metatensor/issues/410
properties = Labels(
"_",
_dispatch.int_array_like(
list(range(out_values.shape[-1])), block.samples.values
).reshape(-1, 1),
# unique because `list(range(...))` produces unique entries
assume_unique=True,
)
else:
properties = self._out_properties[module_idx]
return TensorBlock(
values=out_values,
properties=properties,
components=block.components,
samples=block.samples,
)
@property
def in_keys(self) -> Labels:
"""
Labels that defines the keys this module map expects as input
"""
return self._in_keys
@property
def out_properties(self) -> Union[None, List[Labels]]:
"""
A list of labels that is used to determine properties of the output
"""
return self._out_properties