• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1torch.func
2==========
3
4.. currentmodule:: torch.func
5
6torch.func, previously known as "functorch", is
7`JAX-like <https://github.com/google/jax>`_ composable function transforms for PyTorch.
8
9.. note::
10   This library is currently in `beta <https://pytorch.org/blog/pytorch-feature-classification-changes/#beta>`_.
11   What this means is that the features generally work (unless otherwise documented)
12   and we (the PyTorch team) are committed to bringing this library forward. However, the APIs
13   may change under user feedback and we don't have full coverage over PyTorch operations.
14
15   If you have suggestions on the API or use-cases you'd like to be covered, please
16   open an GitHub issue or reach out. We'd love to hear about how you're using the library.
17
18What are composable function transforms?
19----------------------------------------
20
21- A "function transform" is a higher-order function that accepts a numerical function
22  and returns a new function that computes a different quantity.
23
24- :mod:`torch.func` has auto-differentiation transforms (``grad(f)`` returns a function that
25  computes the gradient of ``f``), a vectorization/batching transform (``vmap(f)``
26  returns a function that computes ``f`` over batches of inputs), and others.
27
28- These function transforms can compose with each other arbitrarily. For example,
29  composing ``vmap(grad(f))`` computes a quantity called per-sample-gradients that
30  stock PyTorch cannot efficiently compute today.
31
32Why composable function transforms?
33-----------------------------------
34
35There are a number of use cases that are tricky to do in PyTorch today:
36
37- computing per-sample-gradients (or other per-sample quantities)
38- running ensembles of models on a single machine
39- efficiently batching together tasks in the inner-loop of MAML
40- efficiently computing Jacobians and Hessians
41- efficiently computing batched Jacobians and Hessians
42
43Composing :func:`vmap`, :func:`grad`, and :func:`vjp` transforms allows us to express the above without designing a separate subsystem for each.
44This idea of composable function transforms comes from the `JAX framework <https://github.com/google/jax>`_.
45
46Read More
47---------
48
49.. toctree::
50   :maxdepth: 2
51
52   func.whirlwind_tour
53   func.api
54   func.ux_limitations
55   func.migrating
56