Type annotations for a tensor's shape, dtype, names, ...

Turn this:

def batch_outer_product(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
    # x has shape (batch, x_channels)
    # y has shape (batch, y_channels)
    # return has shape (batch, x_channels, y_channels)

    return x.unsqueeze(-1) * y.unsqueeze(-2)

into this:

def batch_outer_product(x:   TensorType["batch", "x_channels"],
                        y:   TensorType["batch", "y_channels"]
                        ) -> TensorType["batch", "x_channels", "y_channels"]:

    return x.unsqueeze(-1) * y.unsqueeze(-2)

with programmatic checking that the shape (dtype, ...) specification is met.

Bye-bye bugs! Say hello to enforced, clear documentation of your code.

If (like me) you find yourself littering your code with comments like # x has shape (batch, hidden_state) or statements like assert x.shape == y.shape , just to keep track of what shape everything is, then this is for you.


pip install torchtyping

Requires Python 3.7+ and PyTorch 1.7.0+.


torchtyping allows for type annotating:

  • shape: size, number of dimensions;
  • dtype (float, integer, etc.);
  • layout (dense, sparse);
  • names of dimensions as per named tensors;
  • arbitrary number of batch dimensions with ...;
  • anything else you like, as torchtyping is highly extensible.

If typeguard is (optionally) installed then at runtime the types can be checked to ensure that the tensors really are of the advertised shape, dtype, etc.


from torch import rand
from torchtyping import TensorType, patch_typeguard
from typeguard import typechecked

patch_typeguard()  # use before @typechecked

def func(x: TensorType["batch"],
         y: TensorType["batch"]) -> TensorType["batch"]:
    return x + y

func(rand(3), rand(3))  # works
func(rand(3), rand(1))
# TypeError: Dimension 'batch' of inconsistent size. Got both 1 and 3.

typeguard also has an import hook that can be used to automatically test an entire module, without needing to manually add @typeguard.typechecked decorators.

If you're not using typeguard then torchtyping.patch_typeguard() can be omitted altogether, and torchtyping just used for documentation purposes. If you're not already using typeguard for your regular Python programming, then strongly consider using it. It's a great way to squash bugs. Both typeguard and torchtyping also integrate with pytest, so if you're concerned about any performance penalty then they can be enabled during tests only.


torchtyping.TensorType[shape, dtype, layout, details]

The core of the library.

Each of shape, dtype, layout, details are optional.

  • The shape argument can be any of:
    • An int: the dimension must be of exactly this size. If it is -1 then any size is allowed.
    • A str: the size of the dimension passed at runtime will be bound to this name, and all tensors checked that the sizes are consistent.
    • A ...: An arbitrary number of dimensions of any sizes.
    • A str: int pair (technically it's a slice), combining both str and int behaviour. (Just a str on its own is equivalent to str: -1.)
    • A str: str pair, in which case the size of the dimension passed at runtime will be bound to both names, and all dimensions with either name must have the same size. (Some people like to use this as a way to associate multiple names with a dimension, for extra documentation purposes.)
    • A str: ... pair, in which case the multiple dimensions corresponding to ... will be bound to the name specified by str, and again checked for consistency between arguments.
    • None, which when used in conjunction with is_named below, indicates a dimension that must not have a name in the sense of named tensors.
    • A None: int pair, combining both None and int behaviour. (Just a None on its own is equivalent to None: -1.)
    • A None: str pair, combining both None and str behaviour. (That is, it must not have a named dimension, but must be of a size consistent with other uses of the string.)
    • A typing.Any: Any size is allowed for this dimension (equivalent to -1).
    • Any tuple of the above. For example.TensorType["batch": ..., "length": 10, "channels", -1]. If you just want to specify the number of dimensions then use for example TensorType[-1, -1, -1] for a three-dimensional tensor.
  • The dtype argument can be any of:
    • torch.float32, torch.float64 etc.
    • int, bool, float, which are converted to their corresponding PyTorch types. float is specifically interpreted as torch.get_default_dtype(), which is usually float32.
  • The layout argument can be either torch.strided or torch.sparse_coo, for dense and sparse tensors respectively.
  • The details argument offers a way to pass an arbitrary number of additional flags that customise and extend torchtyping. Two flags are built-in by default. torchtyping.is_named causes the names of tensor dimensions to be checked, and torchtyping.is_float can be used to check that arbitrary floating point types are passed in. (Rather than just a specific one as with e.g. TensorType[torch.float32].) For discussion on how to customise torchtyping with your own details, see the further documentation.
  • Check multiple things at once by just putting them all together inside a single []. For example TensorType["batch": ..., "length", "channels", float, is_named].

torchtyping integrates with typeguard to perform runtime type checking. torchtyping.patch_typeguard() should be called at the global level, and will patch typeguard to check TensorTypes.

This function is safe to run multiple times. (It does nothing after the first run).

  • If using @typeguard.typechecked, then torchtyping.patch_typeguard() should be called any time before using @typeguard.typechecked. For example you could call it at the start of each file using torchtyping.
  • If using typeguard.importhook.install_import_hook, then torchtyping.patch_typeguard() should be called any time before defining the functions you want checked. For example you could call torchtyping.patch_typeguard() just once, at the same time as the typeguard import hook. (The order of the hook and the patch doesn't matter.)
  • If you're not using typeguard then torchtyping.patch_typeguard() can be omitted altogether, and torchtyping just used for documentation purposes.
pytest --torchtyping-patch-typeguard

torchtyping offers a pytest plugin to automatically run torchtyping.patch_typeguard() before your tests. pytest will automatically discover the plugin, you just need to pass the --torchtyping-patch-typeguard flag to enable it. Packages can then be passed to typeguard as normal, either by using @typeguard.typechecked, typeguard's import hook, or the pytest flag --typeguard-packages="your_package_here".

Further documentation

See the further documentation for:

  • FAQ;
    • Including flake8 and mypy compatibility;
  • How to write custom extensions to torchtyping;
  • Resources and links to other libraries and materials on this topic;
  • More examples.