.. automodule:: torch.masked
.. automodule:: torch.masked.maskedtensor
.. currentmodule:: torch
Warning
The PyTorch API of masked tensors is in the prototype stage and may or may not change in the future.
MaskedTensor serves as an extension to :class:`torch.Tensor` that provides the user with the ability to:
- use any masked semantics (e.g. variable length tensors, nan* operators, etc.)
- differentiate between 0 and NaN gradients
- various sparse applications (see tutorial below)
"Specified" and "unspecified" have a long history in PyTorch without formal semantics and certainly without consistency; indeed, MaskedTensor was born out of a build up of issues that the vanilla :class:`torch.Tensor` class could not properly address. Thus, a primary goal of MaskedTensor is to become the source of truth for said "specified" and "unspecified" values in PyTorch where they are a first class citizen instead of an afterthought. In turn, this should further unlock sparsity's potential, enable safer and more consistent operators, and provide a smoother and more intuitive experience for users and developers alike.
A MaskedTensor is a tensor subclass that consists of 1) an input (data), and 2) a mask. The mask tells us which entries from the input should be included or ignored.
By way of example, suppose that we wanted to mask out all values that are equal to 0 (represented by the gray) and take the max:
On top is the vanilla tensor example while the bottom is MaskedTensor where all the 0's are masked out. This clearly yields a different result depending on whether we have the mask, but this flexible structure allows the user to systematically ignore any elements they'd like during computation.
There are already a number of existing tutorials that we've written to help users onboard, such as:
- Overview - the place to start for new users, discusses how to use MaskedTensors and why they're useful
- Sparsity - MaskedTensor supports sparse COO and CSR data and mask Tensors
- Adagrad sparse semantics - a practical example of how MaskedTensor can simplify sparse semantics and implementations
- Advanced semantics - discussion on why certain decisions were made (e.g. requiring masks to match for binary/reduction operations), differences with NumPy's MaskedArray, and reduction semantics
Unary operators are operators that only contain only a single input. Applying them to MaskedTensors is relatively straightforward: if the data is masked out at a given index, we apply the operator, otherwise we'll continue to mask out the data.
The available unary operators are:
.. autosummary:: :toctree: generated :nosignatures: abs absolute acos arccos acosh arccosh angle asin arcsin asinh arcsinh atan arctan atanh arctanh bitwise_not ceil clamp clip conj_physical cos cosh deg2rad digamma erf erfc erfinv exp exp2 expm1 fix floor frac lgamma log log10 log1p log2 logit i0 isnan nan_to_num neg negative positive pow rad2deg reciprocal round rsqrt sigmoid sign sgn signbit sin sinc sinh sqrt square tan tanh trunc
The available inplace unary operators are all of the above except:
.. autosummary:: :toctree: generated :nosignatures: angle positive signbit isnan
As you may have seen in the tutorial, :class:`MaskedTensor` also has binary operations implemented with the caveat that the masks in the two MaskedTensors must match or else an error will be raised. As noted in the error, if you need support for a particular operator or have proposed semantics for how they should behave instead, please open an issue on GitHub. For now, we have decided to go with the most conservative implementation to ensure that users know exactly what is going on and are being intentional about their decisions with masked semantics.
The available binary operators are:
.. autosummary:: :toctree: generated :nosignatures: add atan2 arctan2 bitwise_and bitwise_or bitwise_xor bitwise_left_shift bitwise_right_shift div divide floor_divide fmod logaddexp logaddexp2 mul multiply nextafter remainder sub subtract true_divide eq ne le ge greater greater_equal gt less_equal lt less maximum minimum fmax fmin not_equal
The available inplace binary operators are all of the above except:
.. autosummary:: :toctree: generated :nosignatures: logaddexp logaddexp2 equal fmin minimum fmax
The following reductions are available (with autograd support). For more information, the Overview tutorial details some examples of reductions, while the Advanced semantics tutorial has some further in-depth discussions about how we decided on certain reduction semantics.
.. autosummary:: :toctree: generated :nosignatures: sum mean amin amax argmin argmax prod all norm var std
We've included a number of view and select functions as well; intuitively, these operators will apply to both the data and the mask and then wrap the result in a :class:`MaskedTensor`. For a quick example, consider :func:`select`:
>>> data = torch.arange(12, dtype=torch.float).reshape(3, 4) >>> data tensor([[ 0., 1., 2., 3.], [ 4., 5., 6., 7.], [ 8., 9., 10., 11.]]) >>> mask = torch.tensor([[True, False, False, True], [False, True, False, False], [True, True, True, True]]) >>> mt = masked_tensor(data, mask) >>> data.select(0, 1) tensor([4., 5., 6., 7.]) >>> mask.select(0, 1) tensor([False, True, False, False]) >>> mt.select(0, 1) MaskedTensor( [ --, 5.0000, --, --] )
The following ops are currently supported:
.. autosummary:: :toctree: generated :nosignatures: atleast_1d broadcast_tensors broadcast_to cat chunk column_stack dsplit flatten hsplit hstack kron meshgrid narrow ravel select split t transpose vsplit vstack Tensor.expand Tensor.expand_as Tensor.reshape Tensor.reshape_as Tensor.view