torch.testing¶
Warning
This module is a beta release, and its interfaces and functionality may change without warning in future PyTorch releases.
-
torch.testing.
assert_close
(actual, expected, *, allow_subclasses=True, rtol=None, atol=None, equal_nan=False, check_device=True, check_dtype=True, check_layout=True, check_stride=False, msg=None)[source]¶ Asserts that
actual
andexpected
are close.If
actual
andexpected
are strided, non-quantized, real-valued, and finite, they are considered close ifand they have the same
device
(ifcheck_device
isTrue
), samedtype
(ifcheck_dtype
isTrue
), and the same stride (ifcheck_stride
isTrue
). Non-finite values (-inf
andinf
) are only considered close if and only if they are equal.NaN
’s are only considered equal to each other ifequal_nan
isTrue
.If
actual
andexpected
are sparse (either having COO or CSR layout), their strided members are checked individually. Indices, namelyindices
for COO orcrow_indices
andcol_indices
for CSR layout, are always checked for equality whereas the values are checked for closeness according to the definition above.If
actual
andexpected
are quantized, they are considered close if they have the sameqscheme()
and the result ofdequantize()
is close according to the definition above.actual
andexpected
can beTensor
’s or any tensor-or-scalar-likes from whichtorch.Tensor
’s can be constructed withtorch.as_tensor()
. Except for Python scalars the input types have to be directly related. In addition,actual
andexpected
can beSequence
’s orMapping
’s in which case they are considered close if their structure matches and all their elements are considered close according to the above definition.Note
Python scalars are an exception to the type relation requirement, because their
type()
, i.e.int
,float
, andcomplex
, is equivalent to thedtype
of a tensor-like. Thus, Python scalars of different types can be checked, but requirecheck_dtype=False
.- Parameters
actual (Any) – Actual input.
expected (Any) – Expected input.
allow_subclasses (bool) – If
True
(default) and except for Python scalars, inputs of directly related types are allowed. Otherwise type equality is required.rtol (Optional[float]) – Relative tolerance. If specified
atol
must also be specified. If omitted, default values based on thedtype
are selected with the below table.atol (Optional[float]) – Absolute tolerance. If specified
rtol
must also be specified. If omitted, default values based on thedtype
are selected with the below table.equal_nan (Union[bool, str]) – If
True
, twoNaN
values will be considered equal.check_device (bool) – If
True
(default), asserts that corresponding tensors are on the samedevice
. If this check is disabled, tensors on differentdevice
’s are moved to the CPU before being compared.check_dtype (bool) – If
True
(default), asserts that corresponding tensors have the samedtype
. If this check is disabled, tensors with differentdtype
’s are promoted to a commondtype
(according totorch.promote_types()
) before being compared.check_layout (bool) – If
True
(default), asserts that corresponding tensors have the samelayout
. If this check is disabled, tensors with differentlayout
’s are converted to strided tensors before being compared.check_stride (bool) – If
True
and corresponding tensors are strided, asserts that they have the same stride.msg (Optional[str]) – Optional error message to use in case a failure occurs during the comparison.
- Raises
ValueError – If no
torch.Tensor
can be constructed from an input.ValueError – If only
rtol
oratol
is specified.NotImplementedError – If a tensor is a meta tensor. This is a temporary restriction and will be relaxed in the future.
AssertionError – If corresponding inputs are not Python scalars and are not directly related.
AssertionError – If
allow_subclasses
isFalse
, but corresponding inputs are not Python scalars and have different types.AssertionError – If the inputs are
Sequence
’s, but their length does not match.AssertionError – If the inputs are
Mapping
’s, but their set of keys do not match.AssertionError – If corresponding tensors do not have the same
shape
.AssertionError – If
check_layout
isTrue
, but corresponding tensors do not have the samelayout
.AssertionError – If only one of corresponding tensors is quantized.
AssertionError – If corresponding tensors are quantized, but have different
qscheme()
’s.AssertionError – If
check_device
isTrue
, but corresponding tensors are not on the samedevice
.AssertionError – If
check_dtype
isTrue
, but corresponding tensors do not have the samedtype
.AssertionError – If
check_stride
isTrue
, but corresponding strided tensors do not have the same stride.AssertionError – If the values of corresponding tensors are not close according to the definition above.
The following table displays the default
rtol
andatol
for differentdtype
’s. In case of mismatchingdtype
’s, the maximum of both tolerances is used.dtype
rtol
atol
float16
1e-3
1e-5
bfloat16
1.6e-2
1e-5
float32
1.3e-6
1e-5
float64
1e-7
1e-7
complex64
1.3e-6
1e-5
complex128
1e-7
1e-7
quint8
1.3e-6
1e-5
quint2x4
1.3e-6
1e-5
quint4x2
1.3e-6
1e-5
qint8
1.3e-6
1e-5
qint32
1.3e-6
1e-5
other
0.0
0.0
Note
assert_close()
is highly configurable with strict default settings. Users are encouraged topartial()
it to fit their use case. For example, if an equality check is needed, one might define anassert_equal
that uses zero tolrances for everydtype
by default:>>> import functools >>> assert_equal = functools.partial(torch.testing.assert_close, rtol=0, atol=0) >>> assert_equal(1e-9, 1e-10) Traceback (most recent call last): ... AssertionError: Scalars are not equal! Absolute difference: 9.000000000000001e-10 Relative difference: 9.0
Examples
>>> # tensor to tensor comparison >>> expected = torch.tensor([1e0, 1e-1, 1e-2]) >>> actual = torch.acos(torch.cos(expected)) >>> torch.testing.assert_close(actual, expected)
>>> # scalar to scalar comparison >>> import math >>> expected = math.sqrt(2.0) >>> actual = 2.0 / math.sqrt(2.0) >>> torch.testing.assert_close(actual, expected)
>>> # numpy array to numpy array comparison >>> import numpy as np >>> expected = np.array([1e0, 1e-1, 1e-2]) >>> actual = np.arccos(np.cos(expected)) >>> torch.testing.assert_close(actual, expected)
>>> # sequence to sequence comparison >>> import numpy as np >>> # The types of the sequences do not have to match. They only have to have the same >>> # length and their elements have to match. >>> expected = [torch.tensor([1.0]), 2.0, np.array(3.0)] >>> actual = tuple(expected) >>> torch.testing.assert_close(actual, expected)
>>> # mapping to mapping comparison >>> from collections import OrderedDict >>> import numpy as np >>> foo = torch.tensor(1.0) >>> bar = 2.0 >>> baz = np.array(3.0) >>> # The types and a possible ordering of mappings do not have to match. They only >>> # have to have the same set of keys and their elements have to match. >>> expected = OrderedDict([("foo", foo), ("bar", bar), ("baz", baz)]) >>> actual = {"baz": baz, "bar": bar, "foo": foo} >>> torch.testing.assert_close(actual, expected)
>>> expected = torch.tensor([1.0, 2.0, 3.0]) >>> actual = expected.clone() >>> # By default, directly related instances can be compared >>> torch.testing.assert_close(torch.nn.Parameter(actual), expected) >>> # This check can be made more strict with allow_subclasses=False >>> torch.testing.assert_close( ... torch.nn.Parameter(actual), expected, allow_subclasses=False ... ) Traceback (most recent call last): ... TypeError: No comparison pair was able to handle inputs of type <class 'torch.nn.parameter.Parameter'> and <class 'torch.Tensor'>. >>> # If the inputs are not directly related, they are never considered close >>> torch.testing.assert_close(actual.numpy(), expected) Traceback (most recent call last): ... TypeError: No comparison pair was able to handle inputs of type <class 'numpy.ndarray'> and <class 'torch.Tensor'>. >>> # Exceptions to these rules are Python scalars. They can be checked regardless of >>> # their type if check_dtype=False. >>> torch.testing.assert_close(1.0, 1, check_dtype=False)
>>> # NaN != NaN by default. >>> expected = torch.tensor(float("Nan")) >>> actual = expected.clone() >>> torch.testing.assert_close(actual, expected) Traceback (most recent call last): ... AssertionError: Scalars are not close! Absolute difference: nan (up to 1e-05 allowed) Relative difference: nan (up to 1.3e-06 allowed) >>> torch.testing.assert_close(actual, expected, equal_nan=True)
>>> expected = torch.tensor([1.0, 2.0, 3.0]) >>> actual = torch.tensor([1.0, 4.0, 5.0]) >>> # The default error message can be overwritten. >>> torch.testing.assert_close(actual, expected, msg="Argh, the tensors are not close!") Traceback (most recent call last): ... AssertionError: Argh, the tensors are not close!
-
torch.testing.
make_tensor
(shape, device, dtype, *, low=None, high=None, requires_grad=False, noncontiguous=False, exclude_zero=False)[source]¶ Creates a tensor with the given
shape
,device
, anddtype
, and filled with values uniformly drawn from[low, high)
.If
low
orhigh
are specified and are outside the range of thedtype
’s representable finite values then they are clamped to the lowest or highest representable finite value, respectively. IfNone
, then the following table describes the default values forlow
andhigh
, which depend ondtype
.dtype
low
high
boolean type
0
2
unsigned integral type
0
10
signed integral types
-9
10
floating types
-9
9
complex types
-9
9
- Parameters
shape (Tuple[int, ..]) – A sequence of integers defining the shape of the output tensor.
device (Union[str, torch.device]) – The device of the returned tensor.
dtype (
torch.dtype
) – The data type of the returned tensor.low (Optional[Number]) – Sets the lower limit (inclusive) of the given range. If a number is provided it is clamped to the least representable finite value of the given dtype. When
None
(default), this value is determined based on thedtype
(see the table above). Default:None
.high (Optional[Number]) – Sets the upper limit (exclusive) of the given range. If a number is provided it is clamped to the greatest representable finite value of the given dtype. When
None
(default) this value is determined based on thedtype
(see the table above). Default:None
.requires_grad (Optional[bool]) – If autograd should record operations on the returned tensor. Default:
False
.noncontiguous (Optional[bool]) – If True, the returned tensor will be noncontiguous. This argument is ignored if the constructed tensor has fewer than two elements.
exclude_zero (Optional[bool]) – If
True
then zeros are replaced with the dtype’s small positive value depending on thedtype
. For bool and integer types zero is replaced with one. For floating point types it is replaced with the dtype’s smallest positive normal number (the “tiny” value of thedtype
’sfinfo()
object), and for complex types it is replaced with a complex number whose real and imaginary parts are both the smallest positive normal number representable by the complex type. DefaultFalse
.
- Raises
ValueError – if
requires_grad=True
is passed for integral dtypeValueError – If
low > high
.ValueError – If either
low
orhigh
isnan
.TypeError – If
dtype
isn’t supported by this function.
Examples
>>> from torch.testing import make_tensor >>> # Creates a float tensor with values in [-1, 1) >>> make_tensor((3,), device='cpu', dtype=torch.float32, low=-1, high=1) tensor([ 0.1205, 0.2282, -0.6380]) >>> # Creates a bool tensor on CUDA >>> make_tensor((2, 2), device='cuda', dtype=torch.bool) tensor([[False, False], [False, True]], device='cuda:0')