1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
|
functorch
=========
.. currentmodule:: functorch
.. warning::
We've integrated functorch into PyTorch. As the final step of the
integration, the functorch APIs are deprecated as of PyTorch 2.0.
Please use the torch.func APIs instead and see the
`migration guide <https://pytorch.org/docs/main/func.migrating.html>`_
and `docs <https://pytorch.org/docs/main/func.html>`_
for more details.
Function Transforms
-------------------
.. autosummary::
:toctree: generated
:nosignatures:
vmap
grad
grad_and_value
vjp
jvp
jacrev
jacfwd
hessian
functionalize
Utilities for working with torch.nn.Modules
-------------------------------------------
In general, you can transform over a function that calls a ``torch.nn.Module``.
For example, the following is an example of computing a jacobian of a function
that takes three values and returns three values:
.. code-block:: python
model = torch.nn.Linear(3, 3)
def f(x):
return model(x)
x = torch.randn(3)
jacobian = jacrev(f)(x)
assert jacobian.shape == (3, 3)
However, if you want to do something like compute a jacobian over the parameters
of the model, then there needs to be a way to construct a function where the
parameters are the inputs to the function.
That's what :func:`make_functional` and :func:`make_functional_with_buffers` are for:
given a ``torch.nn.Module``, these return a new function that accepts ``parameters``
and the inputs to the Module's forward pass.
.. autosummary::
:toctree: generated
:nosignatures:
make_functional
make_functional_with_buffers
combine_state_for_ensemble
If you're looking for information on fixing Batch Norm modules, please follow the
guidance here
.. toctree::
:maxdepth: 1
batch_norm
|
