functorch

Why functorch? | Install guide | Transformations | Documentation | Future Plans

This library is currently under heavy development - if you have suggestions on the API or use-cases you'd like to be covered, please open an github issue or reach out. We'd love to hear about how you're using the library.

functorch is JAX-like composable function transforms for PyTorch.

It aims to provide composable vmap and grad transforms that work with PyTorch modules and PyTorch autograd with good eager-mode performance.

In addition, there is experimental functionality to trace through these transformations using FX in order to capture the results of these transforms ahead of time. This would allow us to compile the results of vmap or grad to improve performance.

Why composable function transforms?

There are a number of use cases that are tricky to do in PyTorch today:

• computing per-sample-gradients (or other per-sample quantities)
• running ensembles of models on a single machine
• efficiently batching together tasks in the inner-loop of MAML
• efficiently computing Jacobians and Hessians
• efficiently computing batched Jacobians and Hessians

Composing vmap, grad, vjp, and jvp transforms allows us to express the above without designing a separate subsystem for each. This idea of composable function transforms comes from the JAX framework.

Install

There are two ways to install functorch:

1. functorch from source
2. functorch beta (compatible with recent PyTorch releases)

We recommend trying out the functorch beta first.

Installing functorch from source

import torch
from functorch import vmap
x = torch.randn(3)
y = vmap(torch.sin)(x)
assert torch.allclose(y, x.sin())

pytest test/test_vmap.py -v
pytest test/test_eager_transforms.py -v

pip install networkx

Installing functorch beta (compatible with recent PyTorch releases)

pip install functorch

import torch
from functorch import vmap
x = torch.randn(3)
y = vmap(torch.sin)(x)
assert torch.allclose(y, x.sin())

What are the transforms?

Right now, we support the following transforms:

• grad, vjp, jvp,
• jacrev, jacfwd, hessian
• vmap

Furthermore, we have some utilities for working with PyTorch modules.

• make_functional(model)
• make_functional_with_buffers(model)

vmap

Note: vmap imposes restrictions on the code that it can be used on. For more details, please read its docstring.

vmap(func)(*inputs) is a transform that adds a dimension to all Tensor operations in func. vmap(func) returns a new function that maps func over some dimension (default: 0) of each Tensor in inputs.

vmap is useful for hiding batch dimensions: one can write a function func that runs on examples and then lift it to a function that can take batches of examples with vmap(func), leading to a simpler modeling experience:

from functorch import vmap
batch_size, feature_size = 3, 5
weights = torch.randn(feature_size, requires_grad=True)

def model(feature_vec):
    # Very simple linear model with activation
    assert feature_vec.dim() == 1
    return feature_vec.dot(weights).relu()

examples = torch.randn(batch_size, feature_size)
result = vmap(model)(examples)

grad

grad(func)(*inputs) assumes func returns a single-element Tensor. It compute the gradients of the output of func w.r.t. to inputs[0].

from functorch import grad
x = torch.randn([])
cos_x = grad(lambda x: torch.sin(x))(x)
assert torch.allclose(cos_x, x.cos())

# Second-order gradients
neg_sin_x = grad(grad(lambda x: torch.sin(x)))(x)
assert torch.allclose(neg_sin_x, -x.sin())

When composed with vmap, grad can be used to compute per-sample-gradients:

from functorch import vmap
batch_size, feature_size = 3, 5

def model(weights,feature_vec):
    # Very simple linear model with activation
    assert feature_vec.dim() == 1
    return feature_vec.dot(weights).relu()

def compute_loss(weights, example, target):
    y = model(weights, example)
    return ((y - target) ** 2).mean()  # MSELoss

weights = torch.randn(feature_size, requires_grad=True)
examples = torch.randn(batch_size, feature_size)
targets = torch.randn(batch_size)
inputs = (weights,examples, targets)
grad_weight_per_example = vmap(grad(compute_loss), in_dims=(None, 0, 0))(*inputs)

vjp

The vjp transform applies func to inputs and returns a new function that computes vjps given some cotangents Tensors.

from functorch import vjp
outputs, vjp_fn = vjp(func, inputs); vjps = vjp_fn(*cotangents)

jvp

The jvp transforms computes Jacobian-vector-products and is also known as "forward-mode AD". It is not a higher-order function unlike most other transforms, but it returns the outputs of func(inputs) as well as the jvps.

from functorch import jvp
x = torch.randn(5)
y = torch.randn(5)
f = lambda x, y: (x * y)
_, output = jvp(f, (x, y), (torch.ones(5), torch.ones(5)))
assert torch.allclose(output, x + y)

jacrev, jacfwd, and hessian

The jacrev transform returns a new function that takes in x and returns the Jacobian of torch.sin with respect to x using reverse-mode AD.

from functorch import jacrev
x = torch.randn(5)
jacobian = jacrev(torch.sin)(x)
expected = torch.diag(torch.cos(x))
assert torch.allclose(jacobian, expected)

Use jacrev to compute the jacobian. This can be composed with vmap to produce batched jacobians:

x = torch.randn(64, 5)
jacobian = vmap(jacrev(torch.sin))(x)
assert jacobian.shape == (64, 5, 5)

jacfwd is a drop-in replacement for jacrev that computes Jacobians using forward-mode AD:

from functorch import jacfwd
x = torch.randn(5)
jacobian = jacfwd(torch.sin)(x)
expected = torch.diag(torch.cos(x))
assert torch.allclose(jacobian, expected)

Composing jacrev with itself or jacfwd can produce hessians:

def f(x):
  return x.sin().sum()

x = torch.randn(5)
hessian0 = jacrev(jacrev(f))(x)
hessian1 = jacfwd(jacrev(f))(x)

The hessian is a convenience function that combines jacfwd and jacrev:

from functorch import hessian

def f(x):
  return x.sin().sum()

x = torch.randn(5)
hess = hessian(f)(x)

Tracing through the transformations

We can also trace through these transformations in order to capture the results as new code using make_fx. There is also experimental integration with the NNC compiler (only works on CPU for now!).

from functorch import make_fx, grad
def f(x):
    return torch.sin(x).sum()
x = torch.randn(100)
grad_f = make_fx(grad(f))(x)
print(grad_f.code)

def forward(self, x_1):
    sin = torch.ops.aten.sin(x_1)
    sum_1 = torch.ops.aten.sum(sin, None);  sin = None
    cos = torch.ops.aten.cos(x_1);  x_1 = None
    _tensor_constant0 = self._tensor_constant0
    mul = torch.ops.aten.mul(_tensor_constant0, cos);  _tensor_constant0 = cos = None
    return mul

Working with NN modules: make_functional and friends

Sometimes you may want to perform a transform with respect to the parameters and/or buffers of an nn.Module. This can happen for example in:

• model ensembling, where all of your weights and buffers have an additional dimension
• per-sample-gradient computation where you want to compute per-sample-grads of the loss with respect to the model parameters

Our solution to this right now is an API that, given an nn.Module, creates a stateless version of it that can be called like a function.

• make_functional(model) returns a functional version of model and the model.parameters()
• make_functional_with_buffers(model) returns a functional version of model and the model.parameters() and model.buffers().

Here's an example where we compute per-sample-gradients using an nn.Linear layer:

import torch
from functorch import make_functional, vmap, grad

model = torch.nn.Linear(3, 3)
data = torch.randn(64, 3)
targets = torch.randn(64, 3)

func_model, params = make_functional(model)

def compute_loss(params, data, targets):
    preds = func_model(params, data)
    return torch.mean((preds - targets) ** 2)

per_sample_grads = vmap(grad(compute_loss), (None, 0, 0))(params, data, targets)

If you're making an ensemble of models, you may find combine_state_for_ensemble useful.

Documentation

For more documentation, see our docs website.

Debugging

torch._C._functorch.dump_tensor: Dumps dispatch keys on stack torch._C._functorch._set_vmap_fallback_warning_enabled(False) if the vmap warning spam bothers you.

Future Plans

In the end state, we'd like to upstream this into PyTorch once we iron out the design details. To figure out the details, we need your help -- please send us your use cases by starting a conversation in the issue tracker or trying our project out.

License

Functorch has a BSD-style license, as found in the LICENSE file.

Citing functorch

If you use functorch in your publication, please cite it by using the following BibTeX entry.

@Misc{functorch2021,
  author =       {Horace He, Richard Zou},
  title =        {functorch: JAX-like composable function transforms for PyTorch},
  howpublished = {\url{https://github.com/pytorch/functorch}},
  year =         {2021}
}