Skip to content

Latest commit

 

History

History
214 lines (146 loc) · 8.55 KB

FURTHER-DOCUMENTATION.md

File metadata and controls

214 lines (146 loc) · 8.55 KB
Raw

Further documentation

Design goals

torchtyping had a few design goals.

  • Use type annotations. There's a few other libraries out there that do this via, essentially, syntactic sugar around assert statements. I wanted something neater than that.
  • It should be easy to stop using torchtyping. No, really! If it's not for you then it's easy to remove afterwards. Using torchtyping isn't something you should have to bake into your code; just replace from torchtyping import TensorType with TensorType = list (as a dummy), and your code should still all run.
  • The runtime type checking should be optional. Runtime checks obviously impose a performance penalty. Integrating with typeguard accomplishes this perfectly, in particular through its option to only activate when running tests (my favourite choice).
  • torchtyping should be human-readable. A big part of using type annotations in Python code is to document -- for whoever's reading it -- what is expected. (Particularly valuable on large codebases with several developers.) torchtyping's syntax, and the use of type annotations over some other mechanism, is deliberately chosen to fulfill this goal.

FAQ

The runtime checking isn't working!

First make sure that you're calling torchtyping.patch_typeguard.

Then make sure that you've enabled typeguard, either by decorating the function with typeguard.typechecked, or by using typeguard.importhook.install_import_hook, or by using the pytest command line flags listed in the main README.

Make sure that function you're checking is defined after calling torchtyping.patch_typeguard.

If you have done all of that, then feel free to raise an issue.

flake8 is giving spurious warnings.

Running flake8 will produce spurious warnings for annotations using strings: TensorType["batch"] gives F821 undefined name 'batch'.

You can silence these en-masse just by creating a dummy batch = None anywhere in the file. (Or by placing # noqa: F821 on the relevant lines.)

Does this work with mypy?

Mostly. You'll need to tell mypy not to think too hard about torchtyping, by annotating its import statements with:

from torchtyping import TensorType  # type: ignore

This is because the functionality provided by torchtyping is currently beyond what mypy is capable of representing/understanding. (See also the links at the end for further material on this.)

Additionally mypy has a bug which causes it crash on any file using the str: int or str: ... notation, as in TensorType["batch": 10]. This can be worked around by skipping the file, by creating a filename.pyi file in the same directory. See also the corresponding mypy issue.

Are nested annotations of the form Blahblah[Moreblah[TensorType[...]]] supported?

Yes.

Are multiple ... supported?

Yes. For example:

def func(x:  TensorType["dim1": ..., "dim2": ...],
         y:  TensorType["dim2": ...]
        ) -> TensorType["dim1": ...]:
    sum_dims = [-i - 1 for i in range(y.dim())]
    return (x * y).sum(dim=sum_dims)

TensorType[float] corresponds tofloat32 but torch.rand(2).to(float) produces float64.

This is a deliberate asymmetry. TensorType[float] corresponds to torch.get_default_dtype(), as a convenience, but .to(float) always corresponds to float64.

How to indicate a scalar Tensor, i.e. one with zero dimensions?

TensorType[()]. Equivalently TensorType[(), float], etc.

Support for TensorFlow/JAX/etc?

Not at the moment. The library is called torchtyping after all. There are alternatives for these libraries.

Custom extensions

Writing custom extensions is a breeze. Checking extra properties is done by subclassing torchtyping.TensorDetail, and passing instances of your detail to torchtyping.TensorType. For example this checks that the tensor has an additional attribute foo, which must be a string with value "good-foo":

from torch import rand, Tensor
from torchtyping import TensorDetail, TensorType
from typeguard import typechecked

# Write the extension

class FooDetail(TensorDetail):
    def __init__(self, foo):
        super().__init__()
        self.foo = foo
        
    def check(self, tensor: Tensor) -> bool:
        return hasattr(tensor, "foo") and tensor.foo == self.foo

    # reprs used in error messages when the check is failed
    
    def __repr__(self) -> str:
        return f"FooDetail({self.foo})"

    @classmethod
    def tensor_repr(cls, tensor: Tensor) -> str:
        # Should return a representation of the tensor with respect
        # to what this detail is checking
        if hasattr(tensor, "foo"):
            return f"FooDetail({tensor.foo})"
       	else:
            return ""

# Test the extension

@typechecked
def foo_checker(tensor: TensorType[float, FooDetail("good-foo")]):
    pass


def valid_foo():
    x = rand(3)
    x.foo = "good-foo"
    foo_checker(x)


def invalid_foo_one():
    x = rand(3)
    x.foo = "bad-foo"
    foo_checker(x)


def invalid_foo_two():
    x = rand(2).int()
    x.foo = "good-foo"
    foo_checker(x)

As you can see, a detail must supply three methods. The first is a check method, which takes a tensor and checks whether it satisfies the detail. Second is a __repr__, which is used in error messages, to describe the detail that wasn't satisfied. Third is a tensor_repr, which is also used in error messages, to describe what property the tensor had (instead of the desired detail).

Other libraries and resources

torchtyping is one amongst a few libraries trying to do this kind of thing. Here's some links for the curious:

Discussion

  • PEP 646 proposes variadic generics. These are a tool needed for static checkers (like mypy) to be able to do the kind of shape checking that torchtyping does dynamically. However at time of writing Python doesn't yet support this.
  • The Ideas for array shape typing in Python document is a good overview of some of the ways to type check arrays.

Other libraries

  • TensorAnnotations is a library for statically checking JAX or TensorFlow tensor shapes. (It also has some good links on to other discussions around this topic.)
  • nptyping does something very similar to torchtyping, but for numpy.
  • tsanley/tsalib is an alternative dynamic shape checker, but requires a bit of extra setup.
  • TensorGuard is an alternative, using extra function calls rather than type hints.

More Examples

Shape checking:

def func(x: TensorType["batch", 5],
         y: TensorType["batch", 3]):
    # x has shape (batch, 5)
    # y has shape (batch, 3)
    # batch dimension is the same for both
	
def func(x: TensorType[2, -1, -1]):
	# x has shape (2, any_one, any_two)
	# -1 is a special value to represent any size.

Checking arbitrary numbers of batch dimensions:

def func(x: TensorType[..., 2, 3]):
    # x has shape (..., 2, 3)
	
def func(x: TensorType[..., 2, "channels"],
         y: TensorType[..., "channels"]):
    # x has shape (..., 2, channels)
    # y has shape (..., channels)
    # "channels" is checked to be the same size for both arguments.
	
def func(x: TensorType["batch": ..., "channels_x"],
         y: TensorType["batch": ..., "channels_y"]):
    # x has shape (..., channels_x)
    # y has shape (..., channels_y)
    # the ... batch dimensions are checked to be of the same size.

Return value checking:

def func(x: TensorType[3, 4]) -> TensorType[()]:
    # x has shape (3, 4)
    # return has shape ()

Dtype checking:

def func(x: TensorType[float]):
    # x has dtype torch.float32

Checking shape and dtype at the same time:

def func(x: TensorType[3, 4, float]):
    # x has shape (3, 4) and has dtype torch.float32

Checking names for dimensions as per named tensors:

def func(x: TensorType["a": 3, "b", is_named]):
    # x has has names ("a", "b")
    # x has shape (3, Any)

Checking layouts:

def func(x: TensorType[torch.sparse_coo]):
    # x is a sparse tensor with layout sparse_coo