# oneflow¶

## oneflow¶

Licensed under the Apache License, Version 2.0 (the “License”); you may not use this file except in compliance with the License. You may obtain a copy of the License at

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an “AS IS” BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

class oneflow.BoolTensor

Creates a Tensor with the dtype of oneflow.bool and the device on cpu, it has the same parameters as oneflow.Tensor()

class oneflow.ByteTensor

Creates a Tensor with the dtype of oneflow.uint8 and the device on cpu, it has the same parameters as oneflow.Tensor()

class oneflow.CharTensor

Creates a Tensor with the dtype of oneflow.int8 and the device on cpu, it has the same parameters as oneflow.Tensor()

class oneflow.DoubleTensor

Creates a Tensor with the dtype of oneflow.float64 and the device on cpu, it has the same parameters as oneflow.Tensor()

class oneflow.FloatTensor

Creates a Tensor with the dtype of oneflow.float32 and the device on cpu, it has the same parameters as oneflow.Tensor()

class oneflow.HalfTensor

Creates a Tensor with the dtype of oneflow.float16 and the device on cpu, it has the same parameters as oneflow.Tensor()

class oneflow.IntTensor

Creates a Tensor with the dtype of oneflow.int32 and the device on cpu, it has the same parameters as oneflow.Tensor()

class oneflow.LongTensor

Creates a Tensor with the dtype of oneflow.int64 and the device on cpu, it has the same parameters as oneflow.Tensor()

oneflow.abs()

Return the absolute value of each element in input tensor:math:y = |x| element-wise.

Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> x = flow.tensor(np.array([-1, 2, -3, 4]).astype(np.float32))
>>> flow.abs(x)
tensor([1., 2., 3., 4.], dtype=oneflow.float32)

oneflow.acos()

Returns a new tensor with the inverse cosine of the elements of input.

$\text{out}_{i} = \arccos(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> arr = np.array([0.5, 0.6, 0.7])
>>> input = flow.tensor(arr, dtype=flow.float32)
>>> output = flow.acos(input)
>>> output
tensor([1.0472, 0.9273, 0.7954], dtype=oneflow.float32)

oneflow.acosh()

Returns a new tensor with the inverse hyperbolic cosine of the elements of input.

$\text{out}_{i} = \cosh^{-1}(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> x1 = flow.tensor(np.array([2, 3, 4]).astype(np.float32))
>>> out1 = flow.acosh(x1)
>>> out1
tensor([1.3170, 1.7627, 2.0634], dtype=oneflow.float32)
>>> x2 = flow.tensor(np.array([1.5, 2.6, 3.7]).astype(np.float32),device=flow.device('cuda'))
>>> out2 = flow.acosh(x2)
>>> out2
tensor([0.9624, 1.6094, 1.9827], device='cuda:0', dtype=oneflow.float32)

oneflow.adaptive_avg_pool1d(input, output_size)

Applies a 1D adaptive average pooling over an input signal composed of several input planes.

Parameters
• input – input tensor

• output_size – the target output size (single integer)

oneflow.adaptive_avg_pool2d(input, output_size)

Applies a 2D adaptive average pooling over an input signal composed of several input planes.

Parameters
• input – input tensor

• output_size – the target output size (single integer or double-integer tuple)

oneflow.adaptive_avg_pool3d(input, output_size)

Applies a 3D adaptive average pooling over an input signal composed of several input planes.

Parameters
• input – input tensor

• output_size – the target output size (single integer or triple-integer tuple)

oneflow.add(input, other, *, alpha=1)Tensor

Adds other, scaled by alpha, to input. Scalar and broadcast promotation are supported.

$out = input + alpha \times other$
Parameters
• input (Union[int, float, oneflow.Tensor]) – the input tensor.

• other (Union[int, float, oneflow.Tensor]) – the tensor or number to add to input.

Keyword Arguments

alpha (Number, optional) – the multiplier for other.

Returns

the output Tensor.

Return type

oneflow.Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> x = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> y = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> out.shape
(2, 3)

>>> x = 5
>>> y = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> out.shape
(2, 3)

>>> x = flow.tensor(np.random.randn(1,1), dtype=flow.float32)
>>> y = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> out.shape
(2, 3)

# use alpha
>>> x = flow.zeros(2, 3)
>>> y = flow.ones(2, 3)
>>> out = flow.add(x, y, alpha=10)
>>> out
tensor([[10., 10., 10.],
[10., 10., 10.]], dtype=oneflow.float32)

oneflow.addcmul(input, tensor1, tensor2, *, value=1)Tensor

Performs the element-wise multiplication of tensor1 by tensor2, multiply the result by the scalar value and add it to input. The documentation is referenced from: https://pytorch.org/docs/stable/generated/torch.addcmul.html

$\text{out}_i = \text{input}_i + value \times\ \text{tensor1}_i \times\ \text{tensor2}_i$
Parameters
• input (Tensor) – the tensor to be added.

• tensor1 (Tensor) – the tensor to be multiplied.

• tensor2 (Tensor) – the tensor to be multiplied.

Keyword Arguments

value (Number, optional) – multiplier for $$tensor1 * tensor2$$.

Returns

the output Tensor.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow

>>> input = flow.rand(2, 3, 4)
>>> tensor1 = flow.rand(2, 3, 4)
>>> tensor2 = flow.rand(2, 3, 4)
>>> out = flow.addcmul(input, tensor1, tensor2, value=2)
>>> out.size()
oneflow.Size([2, 3, 4])

oneflow.addmm(beta=1, input, alpha=1, mat1, mat2, out=None)Tensor

Performs a matrix multiplication of the matrices mat1 and mat2. The matrix input is added to the final result.

If mat1 is a $$(n \times m)$$ tensor, mat2 is a $$(m \times p)$$ tensor, then input must be broadcastable with a $$(n \times p)$$ tensor and out will be a $$(n \times p)$$ tensor.

alpha and beta are scaling factors on matrix-vector product between mat1 and mat2 and the added matrix input respectively.

$\text{out} = \beta\ \text{input} + \alpha\ (\text{mat1}_i \mathbin{@} \text{mat2}_i)$

For inputs of type float or double, arguments beta and alpha must be real numbers, otherwise they should be integers.

Parameters
• beta (Number, optional) – multiplier for input ($$\beta$$)

• input (Tensor) – matrix to be added

• alpha (Number, optional) – multiplier for $$mat1 @ mat2$$ ($$\alpha$$)

• mat1 (Tensor) – the first matrix to be multiplied

• mat2 (Tensor) – the second matrix to be multiplied

• out (Tensor, optional) – the output tensor.

For example:

>>> import numpy as np
>>> import oneflow as flow
>>> input = flow.tensor(np.array([[1,2,4],[5,11,9.1]]))
>>> mat1 = flow.tensor(np.array([[7.3,1.9,7.3],[10.2,1,5.5]]))
>>> mat2 = flow.tensor(np.array([[7.3,1.9,7.3],[10.2,1,5.5],[3.7,2.2,8.1]]))
>>> output = flow.addmm(input, mat1, mat2)
>>> output
tensor([[100.6800,  33.8300, 126.8700],
[110.0100,  43.4800, 133.6100]], dtype=oneflow.float64)
>>> output.shape
oneflow.Size([2, 3])

>>> input2 = flow.tensor(np.array([1.7]))
>>> mat1 = flow.tensor(np.array([[1,2],[5,9.1],[7.7,1.4]]))
>>> mat2 = flow.tensor(np.array([[1,2,3.7],[5,9.1,6.8]]))
>>> output2 = flow.addmm(input2, mat1, mat2, alpha=1, beta=2)
>>> output2
tensor([[14.4000, 23.6000, 20.7000],
[53.9000, 96.2100, 83.7800],
[18.1000, 31.5400, 41.4100]], dtype=oneflow.float64)
>>> output2.shape
oneflow.Size([3, 3])

oneflow.all(input, dim=None, keepdim=False)Tensor

For each row of input in the given dimension dim, returns True if all element in the row evaluate to True and False otherwise. If the dimension is None, compute if all elements in the input tensor to true.

If keepdim is True, the output tensor is of the same size as input except in the dimension(s) dim where it is of size 1. Otherwise, dim is squeezed oneflow.squeeze(), resulting in the output tensor having 1 (or len(dim)) fewer dimension(s).

Parameters
• input (oneflow.Tensor) – the Input Tensor

• dim (int, optional) – the dimension to reduce. Default: None

• keepdim (bool, optional) – whether the output tensor has dim retained or not. Default: False

For example:

>>> import oneflow as flow

>>> input = flow.Tensor([[1, 2, 3], [4, 5, 6]]) < 4
>>> input
tensor([[ True,  True,  True],
[False, False, False]], dtype=oneflow.bool)
>>> flow.all(input)
tensor(False, dtype=oneflow.bool)
>>> flow.all(input, 1)
tensor([ True, False], dtype=oneflow.bool)
>>> flow.all(input, 1, True)
tensor([[ True],
[False]], dtype=oneflow.bool)

oneflow.amax(input, dim=None, keepdim=False)Tensor

This function is equivalent to PyTorch’s amax function. It returns the maximum along a dimension.

Parameters
• input (oneflow.Tensor) – the input Tensor.

• dim (int or List of int, optional) – the dimension or the dimensions to reduce. Dim is None by default.

• keepdim (bool, optional) – whether to retain the dimension. keepdim is False by default.

Returns

Maximum of the input tensor

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow

>>> x = flow.tensor([[[0,1],[2,3]],[[4,5],[6,7]]])
>>> flow.amax(x, 1)
tensor([[2, 3],
[6, 7]], dtype=oneflow.int64)
>>> flow.amax(x, 0)
tensor([[4, 5],
[6, 7]], dtype=oneflow.int64)
>>> flow.amax(x)
tensor(7, dtype=oneflow.int64)
>>> flow.amax(x, 0, True)
tensor([[[4, 5],
[6, 7]]], dtype=oneflow.int64)

oneflow.amin(input, dim, keepdim=False)Tensor

This function is equivalent to PyTorch’s amin function. The documentation is referenced from: https://pytorch.org/docs/stable/generated/torch.amin.html.

Returns the minimum value of each slice of the input tensor in the given dimension(s) dim.

If keepdim is True, the output tensor is of the same size as input except in the dimension(s) dim where it is of size 1. Otherwise, dim is squeezed (see oneflow.squeeze()), resulting in the output tensor having 1 (or len(dim)) fewer dimension(s).

Parameters
• input (oneflow.Tensor) – the input Tensor.

• dim (int, Tuple[int]) – the dimension or dimensions to reduce.

• keepdim (bool) – whether the output tensor has dim retained or not.

Example:

>>> import oneflow as flow

>>> x = flow.tensor([[[0,1],[2,3]],[[4,5],[6,7]]])
>>> flow.amin(x, 1)
tensor([[0, 1],
[4, 5]], dtype=oneflow.int64)
>>> flow.amin(x, 0)
tensor([[0, 1],
[2, 3]], dtype=oneflow.int64)
>>> flow.amin(x)
tensor(0, dtype=oneflow.int64)
>>> flow.amin(x, 0, True)
tensor([[[0, 1],
[2, 3]]], dtype=oneflow.int64)

oneflow.any(input, dim=None, keepdim=False)Tensor

For each row of input in the given dimension dim, returns True if any element in the row evaluate to True and False otherwise. If the dimension is None, compute if any elements in the input tensor to true.

If keepdim is True, the output tensor is of the same size as input except in the dimension(s) dim where it is of size 1. Otherwise, dim is squeezed oneflow.squeeze(), resulting in the output tensor having 1 (or len(dim)) fewer dimension(s).

Parameters
• input (oneflow.Tensor) – the Input Tensor

• dim (int, optional) – the dimension to reduce. Default: None

• keepdim (bool, optional) – whether the output tensor has dim retained or not. Default: False

For example:

>>> import oneflow as flow

>>> input = flow.Tensor([[1, 2, 3], [4, 5, 6]]) < 4
>>> input
tensor([[ True,  True,  True],
[False, False, False]], dtype=oneflow.bool)
>>> flow.any(input)
tensor(True, dtype=oneflow.bool)
>>> flow.any(input, 0)
tensor([True, True, True], dtype=oneflow.bool)
>>> flow.any(input, 0, True)
tensor([[True, True, True]], dtype=oneflow.bool)

oneflow.arange(start: int = 0, end, step: int = 1, dtype: Optional[oneflow._oneflow_internal.dtype] = None, device: Optional[Union[oneflow._oneflow_internal.device, str]] = None, placement: Optional[oneflow._oneflow_internal.placement] = None, sbp: Optional[Union[oneflow._oneflow_internal.sbp.sbp, List[oneflow._oneflow_internal.sbp.sbp]]] = None, requires_grad: bool = False)

Returns a 1-D tensor of size $$\left\lfloor \frac{\text{end} - \text{start}}{\text{step}} \right\rfloor + 1$$ with values from start to end with step step. Step is the gap between two values in the tensor.

$\text{out}_{i+1} = \text{out}_i + \text{step}.$
Parameters
• start (int) – the starting value for the set of points. Default: 0.

• end (int) – the ending value for the set of points

• step (int) – the gap between each pair of adjacent points. Default: 1.

Keyword Arguments
• dtype (flow.dtype, optional) – If dtype is not given, infer the dtype from the other input arguments. If any of start, end, or step are floating-point, the dtype is inferred to be the floating-point data type. Otherwise, the dtype is inferred to be flow.int64.

• device (flow.device, optional) – the desired device of returned tensor. Default: if None, uses the current device for the default tensor.

• requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

For example:

>>> import oneflow as flow

>>> y = flow.arange(0, 5)
>>> y
tensor([0, 1, 2, 3, 4], dtype=oneflow.int64)

oneflow.arccos()

Returns a new tensor with the inverse cosine of the elements of input.

$\text{out}_{i} = \arccos(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> arr = np.array([0.5, 0.6, 0.7])
>>> input = flow.tensor(arr, dtype=flow.float32)
>>> output = flow.acos(input)
>>> output
tensor([1.0472, 0.9273, 0.7954], dtype=oneflow.float32)

oneflow.arccosh()

Returns a new tensor with the inverse hyperbolic cosine of the elements of input.

$\text{out}_{i} = \cosh^{-1}(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> x1 = flow.tensor(np.array([2, 3, 4]).astype(np.float32))
>>> out1 = flow.acosh(x1)
>>> out1
tensor([1.3170, 1.7627, 2.0634], dtype=oneflow.float32)
>>> x2 = flow.tensor(np.array([1.5, 2.6, 3.7]).astype(np.float32),device=flow.device('cuda'))
>>> out2 = flow.acosh(x2)
>>> out2
tensor([0.9624, 1.6094, 1.9827], device='cuda:0', dtype=oneflow.float32)

oneflow.arcsin()

Returns a new tensor with the arcsine of the elements of input.

$\text{out}_{i} = \sin^{-1}(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> input = flow.tensor(np.array([-0.5,  0.8, 1.0,  -0.8]), dtype=flow.float32)
>>> output = flow.asin(input)
>>> output.shape
oneflow.Size([4])
>>> output
tensor([-0.5236,  0.9273,  1.5708, -0.9273], dtype=oneflow.float32)
>>> input1 = flow.tensor(np.array([[0.8, 1.0], [-0.6, -1.0]]), dtype=flow.float32)
>>> output1 = input1.asin()
>>> output1.shape
oneflow.Size([2, 2])
>>> output1
tensor([[ 0.9273,  1.5708],
[-0.6435, -1.5708]], dtype=oneflow.float32)

oneflow.arcsinh()

Returns a new tensor with the inverse hyperbolic sine of the elements of input.

$\text{out}_{i} = \sinh^{-1}(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> input = flow.tensor(np.array([2, 3, 4]), dtype=flow.float32)
>>> output = flow.asinh(input)
>>> output.shape
oneflow.Size([3])
>>> output
tensor([1.4436, 1.8184, 2.0947], dtype=oneflow.float32)

>>> input1 = flow.tensor(np.array([[-1, 0, -0.4], [5, 7, 0.8]]), dtype=flow.float32)
>>> output1 = input1.asinh()
>>> output1.shape
oneflow.Size([2, 3])
>>> output1
tensor([[-0.8814,  0.0000, -0.3900],
[ 2.3124,  2.6441,  0.7327]], dtype=oneflow.float32)

oneflow.arctan()

Returns a new tensor with the arctangent of the elements of input.

$\text{out}_{i} = \tan^{-1}(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> input = flow.tensor(np.array([0.5, 0.6, 0.7]), dtype=flow.float32)
>>> output = flow.atan(input)
>>> output.shape
oneflow.Size([3])

oneflow.arctanh()

Returns a new tensor with the inverse hyperbolic tangent of the elements of input.

$\text{out}_{i} = \tanh^{-1}(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> np_arr = np.array([0.5, 0.6, 0.7]).astype(np.float32)
>>> input = flow.tensor(np_arr, dtype=flow.float32)
>>> output = flow.atanh(input)
>>> output
tensor([0.5493, 0.6931, 0.8673], dtype=oneflow.float32)

oneflow.argmax()

The op computes the index with the largest value of a Tensor at specified axis.

Parameters
• input (oneflow.Tensor) – Input Tensor

• dim (int, optional) – dimension to be calculated. Defaults to the last dim (-1)

• keepdim (bool optional) – whether the output tensor has dim retained or not. Ignored if dim=None.

Returns

A Tensor(dtype=int64) contains the index with the largest value of input

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow

>>> input = flow.tensor([[1, 3, 8, 7, 2],
...            [1, 9, 4, 3, 2]], dtype=flow.float32)
>>> output = flow.argmax(input)
>>> output
tensor(6, dtype=oneflow.int64)
>>> output = flow.argmax(input, dim=1)
>>> output
tensor([2, 1], dtype=oneflow.int64)

oneflow.argmin()

The op computes the index with the largest value of a Tensor at specified axis.

Parameters
• input (oneflow.Tensor) – Input Tensor

• dim (int, optional) – dimension to be calculated. Defaults to the last dim (-1)

• keepdim (bool optional) – whether the output tensor has dim retained or not. Ignored if dim=None.

Returns

A Tensor(dtype=int64) contains the index with the largest value of input

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow

>>> input = flow.tensor([[4, 3, 1, 0, 2],
...            [5, 9, 7, 6, 8]], dtype=flow.float32)
>>> output = flow.argmin(input)
>>> output
tensor(3, dtype=oneflow.int64)
>>> output = flow.argmin(input, dim=1)
>>> output
tensor([3, 0], dtype=oneflow.int64)

oneflow.argsort()Tensor

This operator sorts the input Tensor at specified dim and returns the indices of the sorted Tensor.

Parameters
• input (oneflow.Tensor) – the input Tensor.

• dim (int, optional) – the dimension to be sorted. Defaults to the last dim (-1).

• descending (bool, optional) – controls the sorting order (ascending or descending).

Returns

The indices of the sorted Tensor.

Return type

oneflow.Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow
>>> x = np.array([[10, 2, 9, 3, 7],
...               [1, 9, 4, 3, 2]]).astype("float32")
>>> input = flow.Tensor(x)
>>> output = flow.argsort(input)
>>> output
tensor([[1, 3, 4, 2, 0],
[0, 4, 3, 2, 1]], dtype=oneflow.int32)
>>> output = flow.argsort(input, descending=True)
>>> output
tensor([[0, 2, 4, 3, 1],
[1, 2, 3, 4, 0]], dtype=oneflow.int32)
>>> output = flow.argsort(input, dim=0)
>>> output
tensor([[1, 0, 1, 0, 1],
[0, 1, 0, 1, 0]], dtype=oneflow.int32)

oneflow.argwhere(input, dtype: Optional[oneflow._oneflow_internal.dtype] = oneflow.int32)

This operator finds the indices of input Tensor input elements that are non-zero.

It returns a list in which each element is a coordinate that points to a non-zero element in the condition.

Parameters
• input (oneflow.Tensor) – the input Tensor.

• dtype (Optional[flow.dtype], optional) – The data type of output. Defaults to None.

Returns

The result Tensor.

Return type

oneflow.Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow
>>> x = np.array([[0, 1, 0],
...            [2, 0, 2]]).astype(np.float32)

>>> input = flow.Tensor(x)
>>> output = flow.argwhere(input)
>>> output
tensor([[0, 1],
[1, 0],
[1, 2]], dtype=oneflow.int32)

oneflow.as_strided()

Create a view of an existing oneflow.Tensor input with specified size, stride and storage_offset. The documentation is referenced from: https://pytorch.org/docs/1.10/generated/torch.as_strided.html.

Parameters
• input (Tensor) – the input tensor.

• size (tuple or ints) – the shape of the output tensor.

• stride (tuple or ints) – the stride of the output tensor.

• storage_offset (int) – the offset in the underlying storage of the output tensor

Returns

the output Tensor.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow

>>> input = flow.rand(2,3,5)
>>> output = flow.as_strided(input, (2,3,3), (1,2,3), 1)
>>> output.size()
oneflow.Size([2, 3, 3])

oneflow.as_tensor(data, dtype=None, device=None)Tensor

The interface is consistent with PyTorch.

Converts data into a tensor, sharing data and preserving autograd history if possible.

If data is already a tensor with the requeseted dtype and device then data itself is returned, but if data is a tensor with a different dtype or device then it’s copied as if using data.to(dtype=dtype, device=device).

If data is a NumPy array (an ndarray) with the same dtype and device then a tensor is constructed using oneflow.from_numpy.

Parameters
• data (array_like) – Initial data for the tensor. Can be a list, tuple, NumPy ndarray, scalar, and other types.

• dtype (oneflow.dtype, optional) – the desired data type of returned tensor. Default: if None, infers data type from data.

• device (oneflow.device, optional) – the device of the constructed tensor. If None and data is a tensor then the device of data is used. If None and data is not a tensor then the result tensor is constructed on the CPU.

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> a = np.array([1, 2, 3])
>>> t = flow.as_tensor(a, device=flow.device('cuda'))
>>> t
tensor([1, 2, 3], device='cuda:0', dtype=oneflow.int64)
>>> t[0] = -1
>>> a
array([1, 2, 3])

oneflow.asin()

Returns a new tensor with the arcsine of the elements of input.

$\text{out}_{i} = \sin^{-1}(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> input = flow.tensor(np.array([-0.5,  0.8, 1.0,  -0.8]), dtype=flow.float32)
>>> output = flow.asin(input)
>>> output.shape
oneflow.Size([4])
>>> output
tensor([-0.5236,  0.9273,  1.5708, -0.9273], dtype=oneflow.float32)
>>> input1 = flow.tensor(np.array([[0.8, 1.0], [-0.6, -1.0]]), dtype=flow.float32)
>>> output1 = input1.asin()
>>> output1.shape
oneflow.Size([2, 2])
>>> output1
tensor([[ 0.9273,  1.5708],
[-0.6435, -1.5708]], dtype=oneflow.float32)

oneflow.asinh()

Returns a new tensor with the inverse hyperbolic sine of the elements of input.

$\text{out}_{i} = \sinh^{-1}(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> input = flow.tensor(np.array([2, 3, 4]), dtype=flow.float32)
>>> output = flow.asinh(input)
>>> output.shape
oneflow.Size([3])
>>> output
tensor([1.4436, 1.8184, 2.0947], dtype=oneflow.float32)

>>> input1 = flow.tensor(np.array([[-1, 0, -0.4], [5, 7, 0.8]]), dtype=flow.float32)
>>> output1 = input1.asinh()
>>> output1.shape
oneflow.Size([2, 3])
>>> output1
tensor([[-0.8814,  0.0000, -0.3900],
[ 2.3124,  2.6441,  0.7327]], dtype=oneflow.float32)

oneflow.atan()

Returns a new tensor with the arctangent of the elements of input.

$\text{out}_{i} = \tan^{-1}(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> input = flow.tensor(np.array([0.5, 0.6, 0.7]), dtype=flow.float32)
>>> output = flow.atan(input)
>>> output.shape
oneflow.Size([3])

oneflow.atan2()

Element-wise arctangent of input{i}/other{i} with consideration of the quadrant. Returns a new tensor with the signed angles in radians between vector (other{i},input{i}) and vector (1, 0).

The shapes of input and other must be broadcastable.

Parameters
• input (Tensor) – the first input tensor.

• other (Tensor) – the second input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> x1 = flow.Tensor(np.array([1,2,3]))
>>> y1 = flow.Tensor(np.array([3,2,1]))
>>> x2 = flow.Tensor(np.array([1.53123589,0.54242598,0.15117185]))
>>> y2 = flow.Tensor(np.array([-0.21906378,0.09467151,-0.75562878]))
>>> x3 = flow.Tensor(np.array([1,0,-1]))
>>> y3 = flow.Tensor(np.array([0,1,0]))

>>> flow.atan2(x1,y1).numpy()
array([0.32175055, 0.7853982 , 1.2490457 ], dtype=float32)
>>> flow.atan2(x2,y2).numpy()
array([1.7128955, 1.3980033, 2.9441385], dtype=float32)
>>> flow.atan2(x3,y3).numpy()
array([ 1.5707964,  0.       , -1.5707964], dtype=float32)

oneflow.atanh()

Returns a new tensor with the inverse hyperbolic tangent of the elements of input.

$\text{out}_{i} = \tanh^{-1}(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> np_arr = np.array([0.5, 0.6, 0.7]).astype(np.float32)
>>> input = flow.tensor(np_arr, dtype=flow.float32)
>>> output = flow.atanh(input)
>>> output
tensor([0.5493, 0.6931, 0.8673], dtype=oneflow.float32)

oneflow.batch_gather()

Gather the element in batch dims.

Parameters
• in (Tensor) – the input tensor.

• indices (Tensor) – the indices tensor, its dtype must be int32/64.

For example:

Example 1:

>>> import oneflow as flow
>>> import numpy as np

>>> x = flow.Tensor(np.array([[1, 2, 3],
...                           [4, 5, 6]]))
>>> indices = flow.tensor(np.array([1, 0]).astype(np.int64))
>>> out = flow.batch_gather(x, indices)

tensor([[4., 5., 6.],
[1., 2., 3.]], dtype=oneflow.float32)


Example 2:

>>> import oneflow as flow
>>> import numpy as np

>>> x = flow.Tensor(np.array([[[1, 2, 3], [4, 5, 6]],
...                           [[1, 2, 3], [4, 5, 6]]]))
>>> indices = flow.tensor(np.array([[1, 0],
...                                 [0, 1]]).astype(np.int64))
>>> out = flow.batch_gather(x, indices)

tensor([[[4., 5., 6.],
[1., 2., 3.]],
[[1., 2., 3.],
[4., 5., 6.]]], dtype=oneflow.float32)

oneflow.bernoulli(x, *, generator=None, out=None)

This operator returns a Tensor with binaray random numbers (0 / 1) from a Bernoulli distribution.

Parameters
• x (Tensor) – the input tensor of probability values for the Bernoulli distribution

• generator (Generator, optional) – a pseudorandom number generator for sampling

• out (Tensor, optional) – the output tensor.

Shape:
• Input: $$(*)$$. Input can be of any shape

• Output: $$(*)$$. Output is of the same shape as input

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> arr = np.array(
...    [
...        [1.0, 1.0, 1.0],
...        [1.0, 1.0, 1.0],
...        [1.0, 1.0, 1.0],
...    ]
... )
>>> x = flow.tensor(arr, dtype=flow.float32)
>>> y = flow.bernoulli(x)
>>> y
tensor([[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.]], dtype=oneflow.float32)

oneflow.bmm()

Performs a batch matrix-matrix product of matrices stored in input and mat2.

input and mat2 must be 3-D tensors each containing the same number of matrices.

If input is a (b x n x m) tensor, mat2 is a (b x m x p) tensor, out will be a (b x n x p) tensor.

Parameters
• input (oneflow.Tensor) – the first batch of matrices to be multiplied

• mat2 (oneflow.Tensor) – the second batch of matrices to be multiplied

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> input1 = flow.randn(10, 3, 4)
>>> input2 = flow.randn(10, 4, 5)
>>> of_out = flow.bmm(input1, input2)
>>> of_out.shape
oneflow.Size([10, 3, 5])

oneflow.broadcast_like()

Parameters
• x (Tensor) – The input Tensor.

• like_tensor (Tensor) – The like Tensor.

• broadcast_axes (Optional[Sequence], optional) – The axes you want to broadcast. Defaults to None.

Returns

Return type

[Tensor]

For example:

>>> import oneflow as flow

>>> x = flow.randn(3, 1, 1)
>>> like_tensor = flow.randn(3, 4, 5)
oneflow.Size([3, 4, 5])

oneflow.cast()

The operation takes input tensor x and casts it to the output with dtype

Parameters
• x (oneflow.Tensor) – A Tensor

• dtype (flow.dtype) – Data type of the output tensor

Returns

A Tensor with specific dtype.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> np_arr = np.random.randn(2, 3, 4, 5).astype(np.float32)
>>> input = flow.tensor(np_arr, dtype=flow.float32)
>>> output = flow.cast(input, flow.int8)
>>> np.array_equal(output.numpy(), np_arr.astype(np.int8))
True

oneflow.cat(tensors, dim=0)Tensor

Concatenate two or more Tensor s at specified dim.

Analogous to numpy.concatenate

Parameters
• inputs – a list of Tensor

• dim – a int.

Returns

A Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> input1 = flow.tensor(np.random.randn(2, 6, 5, 3), dtype=flow.float32)
>>> input2 = flow.tensor(np.random.randn(2, 6, 5, 3), dtype=flow.float32)
>>> input3 = flow.tensor(np.random.randn(2, 6, 5, 3), dtype=flow.float32)

>>> out = flow.cat([input1, input2, input3], dim=1) # equal to using flow.concat()
>>> out.shape
oneflow.Size([2, 18, 5, 3])

oneflow.ceil()

Returns a new tensor with the ceil of the elements of input, the smallest integer greater than or equal to each element.

The equation is:

$\text{out}_{i} = \left\lceil \text{input}_{i} \right\rceil = \left\lfloor \text{input}_{i} \right\rfloor + 1$
Parameters

input (oneflow.Tensor) – A Tensor.

Returns

The result Tensor

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> x = flow.tensor(np.array([0.1, -2, 3.4]).astype(np.float32))
>>> y = flow.ceil(x)
>>> y.shape
oneflow.Size([3])
>>> y
tensor([ 1., -2.,  4.], dtype=oneflow.float32)
>>> x = flow.tensor(np.array([[2.5, 4.6, 0.6],[7.8, 8.3, 9.2]]).astype(np.float32))
>>> y = x.ceil()
>>> y.shape
oneflow.Size([2, 3])
>>> y
tensor([[ 3.,  5.,  1.],
[ 8.,  9., 10.]], dtype=oneflow.float32)
>>> x = flow.tensor(np.array([[[2.2, 4.4, 6.5],[7.1, 8.2, 9.3]],[[10.6,11.2,12.2],[13.5,14.8,15.9]]]).astype(np.float32))
>>> y = flow.ceil(x)
>>> y.shape
oneflow.Size([2, 2, 3])
>>> y
tensor([[[ 3.,  5.,  7.],
[ 8.,  9., 10.]],

[[11., 12., 13.],
[14., 15., 16.]]], dtype=oneflow.float32)

oneflow.chunk()

Splits a tensor into a specific number of chunks. Each chunk is a view of the input tensor. Last chunk will be bigger if the tensor size along the given dimension dim is not divisible by chunks.

Parameters
• input (oneflow.Tensor) – The tensor to split.

• chunks (int) – Number of chunks to return.

• dim (int) – Dimension along which to split the tensor.

Returns

List of Tensors.

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> arr = np.random.randn(5, 3, 6, 9).astype(np.float32)
>>> input = flow.tensor(arr)
>>> output = []
>>> chunks = 3
>>> output = flow.chunk(input, chunks=chunks, dim=2)
>>> out_shape = []
>>> for i in range(0, chunks):
...     out_shape.append(output[i].numpy().shape)
>>> out_shape
[(5, 3, 2, 9), (5, 3, 2, 9), (5, 3, 2, 9)]

oneflow.clamp()

Clamp all elements in input into the range [ min, max ] and return a resulting tensor:

$\begin{split}y_i = \begin{cases} \text{min} & \text{if } x_i < \text{min} \\ x_i & \text{if } \text{min} \leq x_i \leq \text{max} \\ \text{max} & \text{if } x_i > \text{max} \end{cases}\end{split}$

If input is of type FloatTensor or DoubleTensor, args min and max must be real numbers, otherwise they should be integers.

Parameters
• input (Tensor) – the input tensor.

• min (Number) – lower-bound of the range to be clamped to. Defaults to None.

• max (Number) – upper-bound of the range to be clamped to. Defaults to None.

• out (Tensor, optional) – the output tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> arr = np.array([0.2, 0.6, -1.5, -0.3])
>>> input = flow.Tensor(arr)
>>> output = flow.clamp(input, min=-0.5, max=0.5)
>>> output
tensor([ 0.2000,  0.5000, -0.5000, -0.3000], dtype=oneflow.float32)

>>> arr = np.array([0.2, 0.6, -1.5, -0.3])
>>> input = flow.Tensor(arr)
>>> output = flow.clamp(input, min=None, max=0.5)
>>> output
tensor([ 0.2000,  0.5000, -1.5000, -0.3000], dtype=oneflow.float32)

>>> arr = np.array([0.2, 0.6, -1.5, -0.3])
>>> input = flow.Tensor(arr)
>>> output = flow.clamp(input, min=-0.5, max=None)
>>> output
tensor([ 0.2000,  0.6000, -0.5000, -0.3000], dtype=oneflow.float32)

oneflow.clip()

Alias for oneflow.clamp().

oneflow.concat()

cat(tensors, dim=0) -> Tensor

Concatenate two or more Tensor s at specified dim.

Analogous to numpy.concatenate

Parameters
• inputs – a list of Tensor

• dim – a int.

Returns

A Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> input1 = flow.tensor(np.random.randn(2, 6, 5, 3), dtype=flow.float32)
>>> input2 = flow.tensor(np.random.randn(2, 6, 5, 3), dtype=flow.float32)
>>> input3 = flow.tensor(np.random.randn(2, 6, 5, 3), dtype=flow.float32)

>>> out = flow.cat([input1, input2, input3], dim=1) # equal to using flow.concat()
>>> out.shape
oneflow.Size([2, 18, 5, 3])

oneflow.cos()

Returns a new tensor with the cosine of the elements of input.

$\text{out}_{i} = \cos(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> arr = np.array([1.4309,  1.2706, -0.8562,  0.9796])
>>> input = flow.tensor(arr, dtype=flow.float32)
>>> output = flow.cos(input).numpy()

oneflow.cosh()

Returns a new tensor with the hyperbolic cosine of the elements of input.

$\text{out}_{i} = \cosh(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> arr = np.array([ 0.1632,  1.1835, -0.6979, -0.7325])
>>> input = flow.tensor(arr, dtype=flow.float32)
>>> output = flow.cosh(input).numpy()
>>> output
array([1.0133467, 1.7859949, 1.2535787, 1.2804903], dtype=float32)

oneflow.cumprod(input, dim)Tensor

This operator computes the cumulative product of input elements in the given dimension.

The equation is:

$$y_{i}=x_{0}*x_{1}*…*x_{i}$$

Parameters
• input (Tensor) – the input tensor.

• dim (int) – the dimension to do cumsum whose valid range is [-N, N-1), and the N is tensor’s dimensions

Returns

The result tensor with cumprod result.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> input=flow.tensor([1, 2, 3])
>>> flow.cumprod(input, dim=0)
tensor([1, 2, 6], dtype=oneflow.int64)

oneflow.cumsum(input, dim)Tensor

This operator computes the cumulative sum of input elements in the given dimension.

The equation is:

$$y_{i}=x_{0}+x_{1}+…+x_{i}$$

Parameters
• input (Tensor) – the input ND tensor.

• dim (int) – the dimension to do cumsum, valid range is [-N, N-1), N is tensor’s dimensions

Returns

The result tensor with cumsum result.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> input = flow.ones(3, 3)
>>> dim = 1
>>> flow.cumsum(input, dim)
tensor([[1., 2., 3.],
[1., 2., 3.],
[1., 2., 3.]], dtype=oneflow.float32)

oneflow.decode_onerec(input: Tensor, key: str, dtype, shape, is_dynamic=False, reshape=None, batch_padding=None)Tensor

Decode a tensor from input which should be generated before by oneflow.nn.OneRecReader.

Parameters
• input – (Tensor): The tensor generated by oneflow.nn.OneRecReader before.

• key (str) – The field name of the tensor to be decode

• shape (bool) – The shape of the tensor to be decode

• is_dynamic (bool) – The tensor shape is dynamic or not

• reshape (tuple) – Set it if you want to reshape the tensor

For example:

import oneflow as flow
files = ['file01.onerec', 'file02.onerec']
# read onerec dataset form files

# decode
labels = flow.decode_onerec(readdata, key="labels", dtype=flow.int32, shape=(1,))
dense_fields = flow.decode_onerec(readdata, key="dense_fields", dtype=flow.float, shape=(13,))

oneflow.diag()

If input is a vector (1-D tensor), then returns a 2-D square tensor with the elements of input as the diagonal. If input is a matrix (2-D tensor), then returns a 1-D tensor with diagonal elements of input.

Parameters
• input (Tensor) – the input tensor.

• diagonal (Optional[int], 0) – The diagonal to consider. If diagonal = 0, it is the main diagonal. If diagonal > 0, it is above the main diagonal. If diagonal < 0, it is below the main diagonal. Defaults to 0.

Returns

the output Tensor.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> arr = np.array(
...     [
...        [1.0, 2.0, 3.0],
...        [4.0, 5.0, 6.0],
...        [7.0, 8.0, 9.0],
...     ]
... )

>>> input = flow.tensor(arr, dtype=flow.float32)
>>> flow.diag(input)
tensor([1., 5., 9.], dtype=oneflow.float32)

oneflow.diagonal(input, offset, dim1, dim2)Tensor

Returns a partial view of input with the its diagonal elements with respect to dim1 and dim2 appended as a dimension at the end of the shape.

Parameters
• input (Tensor) – the input tensor.Must be at least 2-dimensional.

• offset (Optional[int], 0) – which diagonal to consider. Default: 0 (main diagonal)

• dim1 (Optional[int], 0) – first dimension with respect to which to take diagonal. Default: 0

• dim2 (Optional[int], 1) – second dimension with respect to which to take diagonal. Default: 1

Returns

the output Tensor.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow

>>> input = flow.randn(2,  3,  4)
>>> output = flow.diagonal(input, offset=1, dim1=1, dim2=0)
>>> output.shape
oneflow.Size([4, 1])

oneflow.div()

Computes the division of input by other for each element, scalar and broadcast promotation are supported. The formula is:

$out = \frac{input}{other}$
Parameters

For example:

>>> import numpy as np
>>> import oneflow as flow

# element-wise divide
>>> input = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> other = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> out = flow.div(input,other).numpy()
>>> out.shape
(2, 3)

# scalar divide
>>> input = 5
>>> other = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> out = flow.div(input,other).numpy()
>>> out.shape
(2, 3)

>>> input = flow.tensor(np.random.randn(1,1), dtype=flow.float32)
>>> other = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> out = flow.div(input,other).numpy()
>>> out.shape
(2, 3)

oneflow.dot()

This operator computes the dot product of tensor input and other.

The equation is:

$$\sum_{i=1}^{n}(x[i] * y[i])$$

Parameters
• input (Tensor) – first tensor in the dot product.

• other (Tensor) – second tensor in the dot product.

Shape:
• input: Input must be 1D.

• other: Other must be 1D.

For example:

>>> import oneflow as flow
>>> flow.dot(flow.Tensor([2, 3]), flow.Tensor([2, 1]))
tensor(7., dtype=oneflow.float32)

oneflow.einsum(equation, *operands)oneflow.Tensor

Sums the product of the elements of the input operands along dimensions specified using a notation based on the Einstein summation convention.

Einsum allows computing many common multi-dimensional linear algebraic array operations by representing them in a short-hand format based on the Einstein summation convention, given by equation. The details of this format are described below, but the general idea is to label every dimension of the input operands with some subscript and define which subscripts are part of the output. The output is then computed by summing the product of the elements of the operands along the dimensions whose subscripts are not part of the output. For example, matrix multiplication can be computed using einsum as flow.einsum(“ij,jk->ik”, A, B). Here, j is the summation subscript and i and k the output subscripts (see section below for more details on why).

Equation:

The equation string specifies the subscripts (letters in [a-zA-Z]) for each dimension of the input operands in the same order as the dimensions, separating subcripts for each operand by a comma (‘,’), e.g. ‘ij,jk’ specify subscripts for two 2D operands. The dimensions labeled with the same subscript must be broadcastable, that is, their size must either match or be 1. The exception is if a subscript is repeated for the same input operand, in which case the dimensions labeled with this subscript for this operand must match in size and the operand will be replaced by its diagonal along these dimensions. The subscripts that appear exactly once in the equation will be part of the output, sorted in increasing alphabetical order. The output is computed by multiplying the input operands element-wise, with their dimensions aligned based on the subscripts, and then summing out the dimensions whose subscripts are not part of the output.

Optionally, the output subscripts can be explicitly defined by adding an arrow (‘->’) at the end of the equation followed by the subscripts for the output. For instance, the following equation computes the transpose of a matrix multiplication: ‘ij,jk->ki’. The output subscripts must appear at least once for some input operand and at most once for the output.

Ellipsis (‘…’) can be used in place of subscripts to broadcast the dimensions covered by the ellipsis. Each input operand may contain at most one ellipsis which will cover the dimensions not covered by subscripts, e.g. for an input operand with 5 dimensions, the ellipsis in the equation ‘ab…c’ cover the third and fourth dimensions. The ellipsis does not need to cover the same number of dimensions across the operands but the ‘shape’ of the ellipsis (the size of the dimensions covered by them) must broadcast together. If the output is not explicitly defined with the arrow (‘->’) notation, the ellipsis will come first in the output (left-most dimensions), before the subscript labels that appear exactly once for the input operands. e.g. the following equation implements batch matrix multiplication ‘…ij,…jk’.

A few final notes: the equation may contain whitespaces between the different elements (subscripts, ellipsis, arrow and comma) but something like ‘…’ is not valid. An empty string ‘’ is valid for scalar operands.

Note

flow.einsum handles ellipsis (‘…’) differently from NumPy in that it allows dimensions covered by the ellipsis to be summed over, that is, ellipsis are not required to be part of the output.

Note

This function does not optimize the given expression, so a different formula for the same computation may run faster or consume less memory. Projects like opt_einsum (https://optimized-einsum.readthedocs.io/en/stable/) can optimize the formula for you.

Parameters
• equation (String) – The subscripts for the Einstein summation.

• *operands (oneflow.Tensor) – The tensors to compute the Einstein summation of.

For example:

>>> import oneflow as flow

# trace
>>> flow.einsum('ii', flow.arange(4*4).reshape(4,4).to(flow.float32))
tensor(30., dtype=oneflow.float32)

# diagonal
>>> flow.einsum('ii->i', flow.arange(4*4).reshape(4,4).to(flow.float32))
tensor([ 0.,  5., 10., 15.], dtype=oneflow.float32)

# outer product
>>> x = flow.arange(5).to(flow.float32)
>>> y = flow.arange(4).to(flow.float32)
>>> flow.einsum('i,j->ij', x, y)
tensor([[ 0.,  0.,  0.,  0.],
[ 0.,  1.,  2.,  3.],
[ 0.,  2.,  4.,  6.],
[ 0.,  3.,  6.,  9.],
[ 0.,  4.,  8., 12.]], dtype=oneflow.float32)

# batch matrix multiplication
>>> As = flow.arange(3*2*5).reshape(3,2,5).to(flow.float32)
>>> Bs = flow.arange(3*5*4).reshape(3,5,4).to(flow.float32)
>>> flow.einsum('bij,bjk->bik', As, Bs).shape
oneflow.Size([3, 2, 4])

# batch permute
>>> A = flow.randn(2, 3, 4, 5)
>>> flow.einsum('...ij->...ji', A).shape
oneflow.Size([2, 3, 5, 4])

# bilinear
>>> A = flow.randn(3,5,4)
>>> l = flow.randn(2,5)
>>> r = flow.randn(2,4)
>>> flow.einsum('bn,anm,bm->ba', l, A, r).shape
oneflow.Size([2, 3])

oneflow.empty(*size, dtype: Optional[oneflow._oneflow_internal.dtype] = None, device: Optional[Union[oneflow._oneflow_internal.device, str]] = None, placement: Optional[oneflow._oneflow_internal.placement] = None, sbp: Optional[Union[oneflow._oneflow_internal.sbp.sbp, List[oneflow._oneflow_internal.sbp.sbp]]] = None, requires_grad: bool = False, pin_memory: bool = False)

Returns a tensor filled with uninitialized data. The shape of the tensor is defined by the variable argument size.

Parameters
• size (int... or oneflow.Size) – Defining the shape of the output tensor. Can be a variable number of arguments or a collection like a list or tuple or oneflow.Size.

• dtype (flow.dtype, optional) – The desired data type of returned tensor. Default: flow.float32.

• device (oneflow.device, optional) – The desired device of returned local tensor. If None, uses the current device.

• placement (flow.placement, optional) – The desired device of returned global tensor. If None, will construct local tensor.

• sbp (flow.sbp or List[flow.sbp], optional) – The desired sbp of returned global tensor.

• requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

• pin_memory (bool, optional) – False.

For example:

>>> import oneflow as flow
>>> y = flow.empty(4, 5)  # construct local empty tensor
>>> y.shape
oneflow.Size([4, 5])
>>> y.is_global
False
>>> placement = flow.placement("cpu", ranks=[0])
>>> y = flow.empty(4, 5, placement=placement, sbp=flow.sbp.broadcast)  # construct consistent empty tensor
>>> y.is_global
True

class oneflow.enable_grad

This context manager is thread local; it will not affect computation in other threads.

Also functions as a decorator. (Make sure to instantiate with parenthesis.)

>>> import oneflow as flow
>>> x = flow.ones(2, 3, requires_grad=True)
...         y = x * x
True
...     return x * x
True

oneflow.eq(input, other)Tensor

Computes element-wise equality. The second argument can be a number or a tensor whose shape is broadcastable with the first argument.

Parameters
Returns

• A boolean tensor that is True where input is equal to other and False elsewhere

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> input = flow.tensor(np.array([2, 3, 4, 5]), dtype=flow.float32)
>>> other = flow.tensor(np.array([2, 3, 4, 1]), dtype=flow.float32)

>>> y = flow.eq(input, other)
>>> y
tensor([ True,  True,  True, False], dtype=oneflow.bool)

oneflow.equal()

eq(input, other) -> Tensor

Computes element-wise equality. The second argument can be a number or a tensor whose shape is broadcastable with the first argument.

Parameters
Returns

• A boolean tensor that is True where input is equal to other and False elsewhere

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> input = flow.tensor(np.array([2, 3, 4, 5]), dtype=flow.float32)
>>> other = flow.tensor(np.array([2, 3, 4, 1]), dtype=flow.float32)

>>> y = flow.eq(input, other)
>>> y
tensor([ True,  True,  True, False], dtype=oneflow.bool)

oneflow.erf()

Computes the error function of each element. The error function is defined as follows:

$\operatorname{erf}(x)=\frac{2}{\sqrt{\pi}} \int_{0}^{x} e^{-t^{2}} d t$
Parameters

x (oneflow.Tensor) – A Tensor

Returns

The result Tensor

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> x = flow.tensor(np.array([0, -1., 10.]), dtype=flow.float32)
>>> out = flow.erf(x)
>>> out.shape
oneflow.Size([3])
>>> out.numpy()
array([ 0.       , -0.8427008,  1.       ], dtype=float32)

>>> x = flow.tensor(np.array([[0, -1., 10.], [5, 7, 0.8]]), dtype=flow.float32)
>>> out = flow.erf(x)
>>> out.shape
oneflow.Size([2, 3])
>>> out.numpy()
array([[ 0.        , -0.8427008 ,  1.        ],
[ 1.        ,  1.        ,  0.74210095]], dtype=float32)

>>> x = flow.tensor(np.array([[0, -1., 10.], [5, 7, 0.8], [2, 3, 4]]), dtype=flow.float32)
>>> out = x.erf()
>>> out.shape
oneflow.Size([3, 3])
>>> out.numpy()
array([[ 0.        , -0.8427008 ,  1.        ],
[ 1.        ,  1.        ,  0.74210095],
[ 0.9953223 ,  0.9999779 ,  1.        ]], dtype=float32)

oneflow.erfc()

Computes the complementary error function of each element of input. The complementary error function is defined as follows:

$\operatorname{erfc}(x)=1-\frac{2}{\sqrt{\pi}} \int_{0}^{x} e^{-t^{2}} d t$
Parameters

x (oneflow.Tensor) – A Tensor

Returns

The result Tensor

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> x = flow.tensor(np.array([0, -1., 10.]), dtype=flow.float32)
>>> out = flow.erfc(x)
>>> out
tensor([1.0000e+00, 1.8427e+00, 2.8026e-45], dtype=oneflow.float32)

>>> x = flow.tensor(np.array([[0, -1., 10.], [5, 7, 0.8]]), dtype=flow.float32)
>>> out = flow.erfc(x)
>>> out
tensor([[1.0000e+00, 1.8427e+00, 2.8026e-45],
[1.5375e-12, 4.1838e-23, 2.5790e-01]], dtype=oneflow.float32)

oneflow.erfinv()

Computes the inverse error function of input. The inverse error function is defined in the range $$(-1, 1)$$ as:

$\mathrm{erfinv}(\mathrm{erf}(x)) = x$
Parameters

input (oneflow.Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> input=flow.tensor(np.random.randn(3,3).astype(np.float32))
>>> of_out=flow.erfinv(input)
>>> of_out.shape
oneflow.Size([3, 3])

oneflow.exp()

This operator computes the exponential of Tensor.

The equation is:

$out = e^x$
Parameters

x (oneflow.Tensor) – A Tensor

Returns

The result Tensor

Return type

oneflow.Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> x = flow.tensor(np.array([1, 2, 3]).astype(np.float32), dtype=flow.float32)
>>> y = flow.exp(x)
>>> y
tensor([ 2.7183,  7.3891, 20.0855], dtype=oneflow.float32)

oneflow.expand(input, *sizes)Tensor

This operator expand the input tensor to a larger size.

Passing -1 as the size for a dimension means not changing the size of that dimension.

Tensor can be also expanded to a larger number of dimensions and the new ones will be appended at the front.

For the new dimensions, the size cannot be set to -1.

Parameters
• input (oneflow.Tensor) – the input Tensor.

• *sizes (oneflow.Size or int) – The desired expanded size.

Returns

The result Tensor.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> x = np.array([[[[0, 1]],
...               [[2, 3]],
...               [[4, 5]]]]).astype(np.int32)
>>> input = flow.Tensor(x)
>>> input.shape
oneflow.Size([1, 3, 1, 2])
>>> out = input.expand(1, 3, 2, 2)
>>> out.shape
oneflow.Size([1, 3, 2, 2])

oneflow.expm1()

Returns a new tensor with the exponential of the elements minus 1 of input.

The equation is:

$y_{i} = e^{x_{i}} - 1$
Parameters

input (oneflow.Tensor) – A Tensor.

Returns

The result Tensor

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> x = flow.tensor(np.array([1, 2, 3]).astype(np.float32))
>>> y = flow.expm1(x)
>>> y.shape
oneflow.Size([3])
>>> y
tensor([ 1.7183,  6.3891, 19.0855], dtype=oneflow.float32)

>>> x = flow.tensor(np.array([[[2, 4, 6],[7, 8, 9]],[[10,11,12],[13,14,15]]]).astype(np.float32))
>>> y = flow.expm1(x)
>>> print(y.shape)
oneflow.Size([2, 2, 3])
>>> print(y.numpy())
[[[6.3890562e+00 5.3598152e+01 4.0242880e+02]
[1.0956332e+03 2.9799580e+03 8.1020840e+03]]

[[2.2025465e+04 5.9873141e+04 1.6275380e+05]
[4.4241238e+05 1.2026032e+06 3.2690165e+06]]]

oneflow.eye(n, m, *, device=None, requires_grad=False, placement=None, sbp)Tensor

This operator creates a 2-D Tensor with ones on the diagonal and zeros elsewhere.

Parameters
• n (int) – the number of rows.

• m (int, optional) – the number of colums with default being n. Defaults to None.

Keyword Arguments
• device (Union[flow.device, str], optional) – the desired device of returned tensor. Default: if None, uses the current device for the default tensor.

• requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

• placement (oneflow._oneflow_internal.placement, optional) – The placement attribute allows you to specify which physical device the tensor is stored on.

• sbp (Union[oneflow._oneflow_internal.sbp.sbp, List[oneflow._oneflow_internal.sbp.sbp]], optional) – When creating a global tensor, specify the SBP of the tensor.

Returns

The result tensor with ones on the diagonal and zeros elsewhere.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> out = flow.eye(3, 3)
>>> out
tensor([[1., 0., 0.],
[0., 1., 0.],
[0., 0., 1.]], dtype=oneflow.float32)
>>> out = flow.eye(3, 3, device="cuda")
>>> out
tensor([[1., 0., 0.],
[0., 1., 0.],
[0., 0., 1.]], device='cuda:0', dtype=oneflow.float32)

oneflow.flatten()

Flattens a contiguous range of dims into a tensor.

Parameters
• start_dim – first dim to flatten (default = 0).

• end_dim – last dim to flatten (default = -1).

For example:

>>> import numpy as np
>>> import oneflow as flow
>>> input = flow.randn(32, 1, 5, 5)
>>> output = flow.flatten(input, start_dim=1)
>>> output.shape
oneflow.Size([32, 25])

oneflow.flip(input, dims)Tensor

Reverse the order of a n-D tensor along given axis in dims.

Note

flow.flip makes a copy of input’s data. This is different from NumPy’s np.flip, which returns a view in constant time. Since copying a tensor’s data is more work than viewing that data, flow.flip is expected to be slower than np.flip.

Parameters
• input (Tensor) – the input tensor

• dims (a list or tuple) – axis to flip on

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> np_arr = np.arange(0, 8).reshape((2, 2, 2)).astype(np.float32)
>>> input = flow.Tensor(np_arr)
>>> input.shape
oneflow.Size([2, 2, 2])
>>> out = flow.flip(input, [0, 1])
>>> out
tensor([[[6., 7.],
[4., 5.]],

[[2., 3.],
[0., 1.]]], dtype=oneflow.float32)

oneflow.floor()

Returns a new tensor with the arcsine of the elements of input.

$\text{out}_{i} = \lfloor \text{input}_{i} \rfloor$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> input = flow.tensor(np.array([-0.5,  1.5, 0,  0.8]), dtype=flow.float32)
>>> output = flow.floor(input)
>>> output.shape
oneflow.Size([4])
>>> output.numpy()
array([-1.,  1.,  0.,  0.], dtype=float32)

>>> input1 = flow.tensor(np.array([[0.8, 1.0], [-0.6, 2.5]]), dtype=flow.float32)
>>> output1 = input1.floor()
>>> output1.shape
oneflow.Size([2, 2])
>>> output1.numpy()
array([[ 0.,  1.],
[-1.,  2.]], dtype=float32)

oneflow.floor_()

In-place version of oneflow.floor()

oneflow.fmod(input, other, *, out=None)Tensor

Computes the element-wise remainder of division.

The dividend and divisor may contain both for integer and floating point numbers. The remainder has the same sign as the dividend input.

Supports broadcasting to a common shape, integer and float inputs.

Parameters
• input (Tensor) – the dividend

• other (Tensor or Scalar) – the divisor

Keyword Arguments

out (Tensor, optional) – the output tensor.

Example:

>>> import oneflow as flow
>>> flow.fmod(flow.tensor([-3., -2, -1, 1, 2, 3], dtype=flow.float32), 2.)
tensor([-1., -0., -1.,  1.,  0.,  1.], dtype=oneflow.float32)
>>> flow.fmod(flow.tensor([1, 2, 3, 4, 5.], dtype=flow.float32), 1.5)
tensor([1.0000, 0.5000, 0.0000, 1.0000, 0.5000], dtype=oneflow.float32)
>>> flow.fmod(flow.tensor([1, 2, 3, 4., -5]), flow.tensor([4, 2, 1, 3., 1]))
tensor([1., 0., 0., 1., -0.], dtype=oneflow.float32)

oneflow.from_numpy()

Creates a Tensor from a numpy.ndarray.

The returned tensor and ndarray share the same memory. Modifications to the tensor will be reflected in the ndarray and vice versa.

It currently accepts ndarray with dtypes of numpy.float64, numpy.float32, numpy.float16, numpy.int64, numpy.int32, numpy.int8, numpy.uint8.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> np_arr = np.arange(6).reshape(2, 3)
>>> t = flow.from_numpy(np_arr)
>>> t
tensor([[0, 1, 2],
[3, 4, 5]], dtype=oneflow.int64)
>>> np_arr[0, 0] = -1
>>> t
tensor([[-1,  1,  2],
[ 3,  4,  5]], dtype=oneflow.int64)

oneflow.full(size: Union[int, Tuple[int, ], oneflow.Size], value: Union[float, int], dtype: Optional[oneflow._oneflow_internal.dtype] = None, device: Optional[Union[str, oneflow._oneflow_internal.device]] = None, placement: Optional[oneflow._oneflow_internal.placement] = None, sbp: Optional[oneflow._oneflow_internal.sbp.sbp] = None, requires_grad: bool = False)

Creates a tensor of size size filled with fill_value. The tensor’s dtype is inferred from value.

Parameters
• size (int...) – a list, tuple, or oneflow.Size of integers defining the shape of the output tensor.

• fill_value (Scalar) – the value to fill the output tensor with.

• dtype (flow.dtype, optional) – the desired data type of returned tensor.

• device (flow.device, optional) – the desired device of returned tensor. Default: if None, uses the current device for the default tensor type

• placement (flow.placement, optional) – the desired placement of returned global tensor. Default: if None, the returned tensor is local one using the argument device.

• sbp (flow.sbp.sbp or tuple of flow.sbp.sbp, optional) – the desired sbp descriptor of returned global tensor. Default: if None, the returned tensor is local one using the argument device.

• requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

For example:

>>> import oneflow as flow
>>> y = flow.full((5,),5)
>>> y
tensor([5, 5, 5, 5, 5], dtype=oneflow.int64)
>>> y = flow.full((2,3),5.0) # construct local tensor
>>> y
tensor([[5., 5., 5.],
[5., 5., 5.]], dtype=oneflow.float32)
>>> placement = flow.placement("cpu", ranks=[0])
>>> y = flow.full((2,3),5.0, placement=placement, sbp=flow.sbp.broadcast)  # construct global tensor
>>> y.is_global
True

oneflow.gather(input, dim, index, sparse_grad=False)Tensor

Gathers values along an axis specified by dim.

For a 3-D tensor the output is specified by:

out[i][j][k] = input[index[i][j][k]][j][k]  # if dim == 0
out[i][j][k] = input[i][index[i][j][k]][k]  # if dim == 1
out[i][j][k] = input[i][j][index[i][j][k]]  # if dim == 2


input and index must have the same number of dimensions. It is also required that index.size(d) <= input.size(d) for all dimensions d != dim. out will have the same shape as index. Note that input and index do not broadcast against each other.

Parameters
• input (Tensor) – the source tensor

• dim (int) – the axis along which to index

• index (LongTensor) – the indices of elements to gather

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> input = np.random.randn(3, 4, 3, 5)
>>> index = np.random.choice(np.arange(3), size=180, replace=True).reshape((3, 4, 3, 5))
>>> output = flow.gather(flow.Tensor(input), 1, flow.tensor(index, dtype=flow.int64))
>>> output.shape
oneflow.Size([3, 4, 3, 5])

oneflow.gather_nd(input, index)Tensor

This operator is a high-dimensional extension of gather, index is a K-dimensional tensor, which is regarded as a index of input Tensor input.

Each element defines a slice of input:

$output[i_{0},i_{1},...,i_{K-2}] = input[index(i_{0},i_{1},...,i_{K-2})]$
Parameters
• input – The input Tensor.

• index – The slice indices.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> input = flow.tensor(np.array([[1, 2,3], [4, 5,6],[7,8,9]]), dtype=flow.float)
>>> index_1 = flow.tensor(np.array([[0], [2]]), dtype=flow.int)
>>> out_1 = flow.gather_nd(input,index_1)
>>> print(out_1.shape)
oneflow.Size([2, 3])
>>> out_1
tensor([[1., 2., 3.],
[7., 8., 9.]], dtype=oneflow.float32)
>>> index_2 = flow.tensor(np.array([[0,2], [2,1]]), dtype=flow.int)
>>> out_2 = flow.gather_nd(input,index_2)
>>> out_2
tensor([3., 8.], dtype=oneflow.float32)

oneflow.gelu(x: Tensor)Tensor

The equation is:

$out = 0.5 * x * (1 + tanh(\sqrt{\frac{2}{\pi}} * (x + 0.044715x^{3})))$

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> x = np.array([-0.5, 0, 0.5]).astype(np.float32)
>>> input = flow.tensor(x)

>>> out = flow.gelu(input)
>>> out
tensor([-0.1543,  0.0000,  0.3457], dtype=oneflow.float32)


See GELU for more details.

oneflow.get_rng_state()

The documentation is referenced from: https://pytorch.org/docs/1.10/generated/torch.get_rng_state.html.

Sets the random number generator state.

Parameters

new_state (oneflow.ByteTensor) – The desired state

oneflow.gt()

Returns the truth value of $$input > other$$ element-wise.

Parameters
Returns

A Tensor with bool type.

Return type

oneflow.Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> input1 = flow.tensor(np.random.randn(2, 6, 5, 3), dtype=flow.float32)
>>> input2 = flow.tensor(np.random.randn(2, 6, 5, 3), dtype=flow.float32)

>>> out = flow.gt(input1, input2).shape
>>> out
oneflow.Size([2, 6, 5, 3])

oneflow.hsplit()

Splits input, a tensor with one or more dimensions, into multiple tensors horizontally according to indices_or_sections. Each split is a view of input. If input is one dimensional this is equivalent to calling oneflow.tensor_split(input, indices_or_sections, dim=0) (the split dimension is zero), and if input has two or more dimensions it’s equivalent to calling oneflow.tensor_split(input, indices_or_sections, dim=1) (the split dimension is 1), except that if indices_or_sections is an integer it must evenly divide the split dimension or a runtime error will be thrown. The documentation is referenced from: https://pytorch.org/docs/1.10/generated/torch.hsplit.html.

Parameters
• input (Tensor) – the input tensor.

• indices_or_sections (int or a list) – If indices_or_sections is an integer n , input is split into n sections along dimension dim.If input is divisible by n along dimension dim, each section will be of equal size, input.size (dim) / n. If input is not divisible by n, the sizes of the first int(input.size(dim) % n). sections will have size int(input.size(dim) / n) + 1, and the rest will have size int(input.size(dim) / n). If indices_or_sections is a list or tuple of ints, then input is split along dimension dim at each of the indices in the list, tuple or tensor. For instance, indices_or_sections=[2, 3] and dim=0 would result in the tensors input[:2], input[2:3], and input[3:].If indices_or_sections is a tensor, it must be a zero-dimensional or one-dimensional long tensor on the CPU.

Returns

the output TensorTuple.

Return type

oneflow.TensorTuple

For example:

>>> import oneflow as flow

>>> input = flow.rand(3,4,5,6)
>>> output = flow.hsplit(input,(1,3))
>>> output[0].size()
oneflow.Size([3, 1, 5, 6])
>>> output[1].size()
oneflow.Size([3, 2, 5, 6])
>>> output[2].size()
oneflow.Size([3, 1, 5, 6])

oneflow.in_top_k(targets, predictions, k)Tensor

Says whether the targets are in the top K predictions.

Parameters
• targets (Tensor) – the target tensor of type int32 or int64.

• predictions (Tensor) – the predictions tensor of type float32 .

• k (int) – Number of top elements to look at for computing precision.

Returns

A Tensor of type bool. Computed Precision at k as a bool Tensor.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> targets1 = flow.tensor(np.array([3, 1]), dtype=flow.int32)
>>> predictions1 = flow.tensor(np.array([[0.0, 1.0, 2.0, 3.0], [3.0, 2.0, 1.0, 0.0],]), dtype=flow.float32)
>>> out1 = flow.in_top_k(targets1, predictions1, k=1)
>>> out1
tensor([ True, False], dtype=oneflow.bool)
>>> out2 = flow.in_top_k(targets1, predictions1, k=2)
>>> out2
tensor([True, True], dtype=oneflow.bool)
>>> targets2 = flow.tensor(np.array([3, 1]), dtype=flow.int32, device=flow.device('cuda'))
>>> predictions2 = flow.tensor(np.array([[0.0, 1.0, 2.0, 3.0], [3.0, 2.0, 1.0, 0.0],]), dtype=flow.float32, device=flow.device('cuda'))
>>> out3 = flow.in_top_k(targets2, predictions2, k=1)
>>> out3
tensor([ True, False], device='cuda:0', dtype=oneflow.bool)

oneflow.index_select(dim, index)Tensor

The interface is consistent with PyTorch. The documentation is referenced from: https://pytorch.org/docs/1.10/generated/torch.index_select.html.

Select values along an axis specified by dim.

index must be an Int32 Tensor with 1-D. dim must be in the range of input Dimensions. value of index must be in the range of the dim-th of input. Note that input and index do not broadcast against each other.

Parameters
• input (Tensor) – the source tensor

• dim (int) – the axis along which to index

• index (Tensor) – the 1-D tensor containing the indices to index

For example:

>>> import oneflow as flow
>>> input = flow.tensor([[1,2,3],[4,5,6]], dtype=flow.int32)
>>> input
tensor([[1, 2, 3],
[4, 5, 6]], dtype=oneflow.int32)
>>> index = flow.tensor([0,1], dtype=flow.int64)
>>> output = flow.index_select(input, 1, index)
>>> output
tensor([[1, 2],
[4, 5]], dtype=oneflow.int32)
>>> output = input.index_select(1, index)
>>> output
tensor([[1, 2],
[4, 5]], dtype=oneflow.int32)

class oneflow.inference_mode(mode=True)

Context-manager that enables or disables inference mode

InferenceMode is a new context manager analogous to no_grad to be used when you arecertain your operations will have no interactions with autograd (e.g., model training). Code run under this mode gets better performance by disabling view tracking and version counter bumps.

This context manager is thread local; it will not affect computation in other threads.

Also functions as a decorator. (Make sure to instantiate with parenthesis.)

Parameters

mode (bool) – Flag whether to enable or disable inference mode. (default: True)

>>> import oneflow as flow
>>> x = flow.ones(2, 3, requires_grad=True)
>>> with flow.inference_mode():
...     y = x * x
False
>>> @flow.inference_mode()
...     return x * x
False

oneflow.initial_seed()int

The documentation is referenced from: https://pytorch.org/docs/1.10/_modules/torch/random.html.

Returns the initial seed for generating random numbers as a Python long.

oneflow.is_floating_point(input)

Returns True if the data type of input is a floating point data type i.e., one of flow.float64, flow.float32, flow.float16.

Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow

>>> input = flow.tensor([1, 2, 3, 4, 5], dtype=flow.int)
>>> output = flow.is_floating_point(input)
>>> output
False

oneflow.is_grad_enabled()

Returns True if grad mode is currently enabled.

oneflow.is_nonzero(input) -> (bool)

Returns True if the input is a single element tensor which is not equal to zero after type conversions. i.e. not equal to flow.tensor([0.]) or flow.tensor([0]).

Throws a RuntimeError if input.shape.numel() != 1

For Example:

>>> import oneflow as flow
>>> flow.is_nonzero(flow.tensor([0.]))
False
>>> flow.is_nonzero(flow.tensor([1.5]))
True
>>> flow.is_nonzero(flow.tensor([3]))
True

oneflow.is_tensor(input) -> (bool)

Note that this function is simply doing isinstance(obj, Tensor). Using that isinstance check is better for typechecking with mypy, and more explicit - so it’s recommended to use that instead of is_tensor.

Parameters

obj (Object) – Object to test

For example:

>>> import oneflow as flow

>>> x=flow.tensor([1,2,3])
>>> flow.is_tensor(x)
True

oneflow.isinf()

Tests if each element of input is infinite (positive or negative infinity) or not.

Parameters

input (Tensor) – the input tensor.

Returns

A boolean tensor that is True where input is infinite and False elsewhere.

Example:

>>> import oneflow as flow
>>> flow.isinf(flow.tensor([1, float('inf'), 2, float('-inf'), float('nan')]))
tensor([False,  True, False,  True, False], dtype=oneflow.bool)

oneflow.isnan()

Returns a new tensor with boolean elements representing if each element of input is NaN or not.

Parameters

input (Tensor) – the input tensor.

Returns

A boolean tensor that is True where input is NaN and False elsewhere.

Example:

>>> import oneflow as flow
>>> flow.isnan(flow.tensor([1, float('nan'), 2]))
tensor([False,  True, False], dtype=oneflow.bool)

oneflow.le(input, other)Tensor

Returns the truth value of $$input <= other$$ element-wise.

Parameters
Returns

A Tensor with bool type.

Return type

oneflow.Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> input1 = flow.tensor(np.array([1, 2, 3]).astype(np.float32), dtype=flow.float32)
>>> input2 = flow.tensor(np.array([1, 1, 4]).astype(np.float32), dtype=flow.float32)

>>> out = flow.le(input1, input2)
>>> out
tensor([ True, False,  True], dtype=oneflow.bool)

oneflow.linspace(start: float, end: float, steps: int, dtype: oneflow._oneflow_internal.dtype = oneflow.float32, device: Optional[Union[oneflow._oneflow_internal.device, str]] = None, placement: Optional[oneflow._oneflow_internal.placement] = None, sbp: Optional[Union[oneflow._oneflow_internal.sbp.sbp, List[oneflow._oneflow_internal.sbp.sbp]]] = None, requires_grad: bool = False)

Creates a one-dimensional tensor of size steps whose values are evenly spaced from start to end, inclusive. That is, the value are:

$(\text{start}, \text{start} + \frac{\text{end} - \text{start}}{\text{steps} - 1}, \ldots, \text{start} + (\text{steps} - 2) * \frac{\text{end} - \text{start}}{\text{steps} - 1}, \text{end})$
Parameters
• start (float) – the starting value for the set of points

• end (float) – the ending value for the set of points

• steps (int) – size of the constructed tensor

Keyword Arguments
• dtype (flow.dtype, optional) – If dtype is not given, the dtype is inferred to be flow.float32.

• device (flow.device, optional) – the desired device of returned tensor. Default: if None, uses the current device for the default tensor.

• requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

For example:

>>> import oneflow as flow

>>> y = flow.linspace(3, 10, steps=5)
>>> y
tensor([ 3.0000,  4.7500,  6.5000,  8.2500, 10.0000], dtype=oneflow.float32)

oneflow.load(path: str, global_src_rank: Optional[int] = None)Any

Loads an object saved with oneflow.save() from a directory.

Parameters
• path (str) – The directory containing the object

• global_src_rank (int, optional) – The source rank for loading global tensors. When specified, only the process whose rank == global_src_rank will really read the files in path, and tensors in the loaded object will be consistent with placement = flow.placement(‘cuda’, [global_src_rank])

Returns

oneflow.log()

Returns a new tensor with the natural logarithm of the elements of input.

$y_{i} = \log_{e} (x_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> arr = np.random.randn(2, 3, 4, 5)
>>> input = flow.tensor(arr, dtype=flow.float32)
>>> output = flow.log(input)

oneflow.log1p()

Returns a new tensor with the natural logarithm of (1 + input).

$\text{out}_{i}=\log_e(1+\text{input}_{i})$

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> x = flow.tensor(np.array([1.3, 1.5, 2.7]), dtype=flow.float32)
>>> out = flow.log1p(x)
>>> out
tensor([0.8329, 0.9163, 1.3083], dtype=oneflow.float32)

oneflow.log2(input)Tensor

Returns a new tensor with the natural logarithm to the base 2 of the elements of input.

$y_{i} = \log2_{e} (x_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> arr = np.random.randn(2, 3, 4, 5)
>>> input = flow.tensor(arr, dtype=flow.float32)
>>> output = flow.log2(input)

oneflow.logical_and()

Computes the element-wise logical AND of the given input tensors. Zeros are treated as False and nonzeros are treated as True.

Parameters
Returns

The output Tensor

Return type

oneflow.Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> input1 = flow.tensor(np.array([1, 0, 1]).astype(np.float32), dtype=flow.float32)
>>> input2 = flow.tensor(np.array([1, 1, 0]).astype(np.float32), dtype=flow.float32)

>>> out = flow.logical_and(input1, input2)
>>> out
tensor([ True, False, False], dtype=oneflow.bool)

oneflow.logical_not()

Computes the element-wise logical NOT of the given input tensors. Zeros are treated as False and nonzeros are treated as True. :param input: The input Tensor :type input: oneflow.Tensor :param other: The Tensor to compute NOT with :type other: oneflow.Tensor

Returns

The output Tensor

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow

>>> input = flow.tensor([1, 0, -1], dtype=flow.float32)
>>> out = flow.logical_not(input)
>>> out
tensor([False,  True, False], dtype=oneflow.bool)

oneflow.logical_or()

Computes the element-wise logical OR of the given input tensors. Zeros are treated as False and nonzeros are treated as True.

Parameters
Returns

The output Tensor

Return type

oneflow.Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> input1 = flow.tensor(np.array([1, 0, 1]).astype(np.float32), dtype=flow.float32)
>>> input2 = flow.tensor(np.array([1, 0, 0]).astype(np.float32), dtype=flow.float32)

>>> out = flow.logical_or(input1, input2)
>>> out
tensor([ True, False,  True], dtype=oneflow.bool)

oneflow.logical_xor()

Computes the element-wise logical XOR of the given input tensors. Zeros are treated as False and nonzeros are treated as True.

Parameters
Returns

The output Tensor

Return type

oneflow.Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> input1 = flow.tensor(np.array([1, 0, 1]).astype(np.float32), dtype=flow.float32)
>>> input2 = flow.tensor(np.array([1, 0, 0]).astype(np.float32), dtype=flow.float32)
>>> out = flow.logical_xor(input1, input2)
>>> out
tensor([False, False,  True], dtype=oneflow.bool)

oneflow.lt(input, other)Tensor

Returns the truth value of $$input < other$$ element-wise.

Parameters
Returns

A Tensor with bool type.

Return type

oneflow.Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> input1 = flow.tensor(np.array([1, 2, 3]).astype(np.float32), dtype=flow.float32)
>>> input2 = flow.tensor(np.array([1, 2, 4]).astype(np.float32), dtype=flow.float32)

>>> out = flow.lt(input1, input2)
>>> out
tensor([False, False,  True], dtype=oneflow.bool)

oneflow.manual_seed(seed)

The documentation is referenced from: https://pytorch.org/docs/1.10/generated/torch.manual_seed.html.

Sets the seed for generating random numbers. Returns a oneflow.Generator object.

Parameters

seed (int) – The desired seed. Value must be within the inclusive range [-0x8000_0000_0000_0000, 0xffff_ffff_ffff_ffff]. Otherwise, a RuntimeError is raised. Negative inputs are remapped to positive values with the formula 0xffff_ffff_ffff_ffff + seed.

oneflow.masked_fill()

Fills elements of self tensor with value where mask is True. The shape of mask must be broadcastable with the shape of the underlying tensor.

Parameters

• value (float) – the value to fill in with

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> in_arr = np.array(
...     [[[-0.13169311,  0.97277078,  1.23305363,  1.56752789],
...     [-1.51954275,  1.87629473, -0.53301206,  0.53006478],
...     [-1.38244183, -2.63448052,  1.30845795, -0.67144869]],
...     [[ 0.41502161,  0.14452418,  0.38968   , -1.76905653],
...     [ 0.34675095, -0.7050969 , -0.7647731 , -0.73233418],
...     [-1.90089858,  0.01262963,  0.74693893,  0.57132389]]]
... )
>>> fill_value = 8.7654321 # random value e.g. -1e9 3.1415
>>> input = flow.tensor(in_arr, dtype=flow.float32)
>>> mask = flow.tensor((in_arr > 0).astype(np.int8), dtype=flow.int)

# tensor([[[-0.1317,  8.7654,  8.7654,  8.7654],
#  [-1.5195,  8.7654, -0.533 ,  8.7654],
#  [-1.3824, -2.6345,  8.7654, -0.6714]],

# [[ 8.7654,  8.7654,  8.7654, -1.7691],
#  [ 8.7654, -0.7051, -0.7648, -0.7323],
#  [-1.9009,  8.7654,  8.7654,  8.7654]]], dtype=oneflow.float32)

oneflow.masked_select(input, mask)

Returns a new 1-D tensor which indexes the input tensor according to the boolean mask mask which is a BoolTensor(In oneFlow BoolTensor is replaced by Int8Tensor).

The shapes of the mask tensor and the input tensor don’t need to match, but they must be broadcastable.

Parameters
• input (Tensor) – the input tensor.

• mask (Tensor) – the tensor containing the binary mask to index with

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> input = flow.tensor(np.array([[-0.4620, 0.3139], [0.3898, -0.7197], [0.0478, -0.1657]]), dtype=flow.float32)
>>> out
tensor([0.3139, 0.3898], dtype=oneflow.float32)

oneflow.matmul(input, other)Tensor

This operator applies matrix multiplication to two Tensor.

Parameters
Returns

The result Tensor

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> input1 = flow.tensor(np.random.randn(2, 6), dtype=flow.float32)
>>> input2 = flow.tensor(np.random.randn(6, 5), dtype=flow.float32)
>>> of_out = flow.matmul(input1, input2)
>>> of_out.shape
oneflow.Size([2, 5])

oneflow.max(input, dim=None, keepdim=False)

Computes the maximum value of all elements in the input tensor.

Parameters
• input (oneflow.Tensor) – the Input Tensor

• dim (int, optional) – the dimension to reduce. Default: None

• keepdim (bool, optional) – whether the output tensor has dim retained or not. Default: False

Returns

If dim is None, returns the maximum value of all elements in the input tensor. Otherwise, returns a tuple of Tensor (values, indices), where the values are the maximum value of all elements in the input tensor, the indices are the indices of the elements in the original input tensor.

Return type

Tensor or Tuple(oneflow.Tensor, oneflow.Tensor(dtype=int64))

For example:

>>> import oneflow as flow

>>> input = flow.Tensor([[4, 1, 5], [2, 6, 3]])
>>> flow.max(input)
tensor(6., dtype=oneflow.float32)
>>> (values, indices) = flow.max(input, dim=1)
>>> values
tensor([5., 6.], dtype=oneflow.float32)
>>> indices
tensor([2, 1], dtype=oneflow.int64)

oneflow.mean(input, dim=None, keepdim=False)Tensor

Computes the mean of row of elements in a tensor in the given dimension. If the dimension is None, mean of all elements will be caculated.

If keepdim is True, the output tensor is of the same size as input except in the dimension(s) dim where it is of size 1. Otherwise, dim is squeezed oneflow.squeeze(), resulting in the output tensor having 1 (or len(dim)) fewer dimension(s).

Parameters
• input (oneflow.Tensor) – the Input Tensor

• dim (int or tuple of ints, optional) – the dimension to reduce. Default: None

• keepdim (bool, optional) – whether the output tensor has dim retained or not. Default: False

For example:

>>> import oneflow as flow

>>> input = flow.Tensor([[1, 2, 3], [4, 5, 6]])
>>> flow.mean(input)
tensor(3.5000, dtype=oneflow.float32)
>>> flow.mean(input, dim=0)
tensor([2.5000, 3.5000, 4.5000], dtype=oneflow.float32)
>>> flow.mean(input, dim=1)
tensor([2., 5.], dtype=oneflow.float32)

oneflow.median(input)Tensor

Returns the median of the values in input. The documentation is referenced from: https://pytorch.org/docs/1.10/generated/torch.median.html#torch.median

Note

The median is not unique for input tensors with an even number of elements. In this case the lower of the two medians is returned.

Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> x = flow.tensor((1, 2, -1), dtype=flow.float32)
>>> flow.median(x)
tensor(1., dtype=oneflow.float32)

oneflow.median(input, dim=- 1, keepdim=False, *, out=None)

Returns a tuple (values, indices) where values contains the median of each row of input in the dimension dim, and indices contains the index of the median values found in the dimension dim.

By default, dim is the last dimension of the input tensor.

If keepdim is True, the output tensors are of the same size as input except in the dimension dim where they are of size 1. Otherwise, dim is squeezed (see flow.squeeze()), resulting in the outputs tensor having 1 fewer dimension than input.

Note

The median is not unique for input tensors with an even number of elements in the dimension dim. In this case the lower of the two medians is returned.

Parameters
• input (Tensor) – the input tensor.

• dim (int) – the dimension to reduce.

• keepdim (bool) – whether the output tensor has dim retained or not.

For example:

>>> import oneflow as flow
>>> a = flow.tensor([[ 0.2505, -0.3982, -0.9948,  0.3518, -1.3131],
...    [ 0.3180, -0.6993,  1.0436,  0.0438,  0.2270],
...    [-0.2751,  0.7303,  0.2192,  0.3321,  0.2488],
...    [ 1.0778, -1.9510,  0.7048,  0.4742, -0.7125]])
>>> flow.median(a, 1)
(tensor([-0.3982,  0.2270,  0.2488,  0.4742], dtype=oneflow.float32), tensor([1, 4, 4, 3], dtype=oneflow.int64))

oneflow.meshgrid(*tensors, indexing='ij')

The interface is consistent with PyTorch. The documentation is referenced from: https://pytorch.org/docs/1.10/_modules/torch/functional.html#meshgrid.

Take $$N$$ tensors, each of which can be either scalar or 1-dimensional vector, and create $$N$$ N-dimensional grids, where the $$i$$ th grid is defined by expanding the $$i$$ th input over dimensions defined by other inputs.

Parameters
• tensors (list of Tensor) – list of scalars or 1 dimensional tensors. Scalars will be treated as tensors of size $$(1,)$$ automatically.

• indexing ((string, optional) – the indexing mode, either “xy” or “ij”, defaults to “ij”. If “ij” is selected, the dimensions are in the same order as the cardinality of the inputs. If “xy” is selected, the first dimension corresponds to the cardinality of the second input and the second dimension corresponds to the cardinality of the first input.

Returns

If the input has $$k$$ tensors of size $$(N_1,), (N_2,), \ldots , (N_k,)$$, then the output would also have $$k$$ tensors, where all tensors are of size $$(N_1, N_2, \ldots , N_k)$$.

Return type

seq (sequence of Tensors)

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> input1 = flow.tensor(np.array([2, 2, 3]), dtype=flow.float32)
>>> input2 = flow.tensor(np.array([4, 5, 6]), dtype=flow.float32)
>>> of_x, of_y = flow.meshgrid(input1, input2)
>>> of_x
tensor([[2., 2., 2.],
[2., 2., 2.],
[3., 3., 3.]], dtype=oneflow.float32)
>>> of_y
tensor([[4., 5., 6.],
[4., 5., 6.],
[4., 5., 6.]], dtype=oneflow.float32)

oneflow.min(input, dim=None, keepdim=False)

Computes the minimum value of all elements in the input tensor.

Parameters
• input (oneflow.Tensor) – the Input Tensor

• dim (int, optional) – the dimension to reduce. Default: None

• keepdim (bool, optional) – whether the output tensor has dim retained or not. Default: False

Returns

If dim is None, returns the minimum value of all elements in the input tensor. Otherwise, returns a tuple of Tensor (values, indices), where the values are the minimum value of all elements in the input tensor, the indices are the indices of the elements in the original input tensor.

Return type

Tensor or Tuple(oneflow.Tensor, oneflow.Tensor(dtype=int64))

For example:

>>> import oneflow as flow

>>> input = flow.Tensor([[4, 1, 5], [2, 6, 3]])
>>> flow.min(input)
tensor(1., dtype=oneflow.float32)
>>> (values, indices) = flow.min(input, dim=1)
>>> values
tensor([1., 2.], dtype=oneflow.float32)
>>> indices
tensor([1, 0], dtype=oneflow.int64)

oneflow.mish(x: Tensor)Tensor

Applies the element-wise function:

$ext{mish}(x) = x * ext{tanh}( ext{softplus}(x))$

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> x = np.array([1, 2, 3]).astype(np.float32)
>>> input = flow.tensor(x)

>>> out = flow.mish(input)
>>> out
tensor([0.8651, 1.9440, 2.9865], dtype=oneflow.float32)


See Mish for more details.

oneflow.movedim()

Moves the dimension(s) of input at the position(s) in source to the position(s) in destination. Other dimensions of input that are not explicitly moved remain in their original order and appear at the positions not specified in destination. The documentation is referenced from: https://pytorch.org/docs/1.10/generated/torch.movedim.html.

Parameters
• input (Tensor) – the input tensor.

• source (int or a list) – Original positions of the dims to move. These must be unique.

• destination (int or a list) – Destination positions for each of the original dims. These must also be unique.

Returns

the output Tensor.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> input = flow.tensor(np.random.randn(2, 3, 4, 5), dtype=flow.float32)
>>> output = flow.movedim(input, 1, 0)
>>> output.shape
oneflow.Size([3, 2, 4, 5])
>>> output = flow.movedim(input, (1, 2), (0, 1))
>>> output.shape
oneflow.Size([3, 4, 2, 5])

oneflow.mul()

Computes the multiplication of input by other for each element, scalar and broadcast promotation are supported.

The formula is:

$\text{out}_i = \text{input}_i \times \text{other}_i$

For example:

>>> import numpy as np
>>> import oneflow as flow

# element-wise multiply
>>> input = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> other = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> out = flow.mul(input,other).numpy()
>>> out.shape
(2, 3)

# scalar mutiply
>>> input = 5
>>> other = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> out = flow.mul(input,other).numpy()
>>> out.shape
(2, 3)

>>> input = flow.tensor(np.random.randn(1,1), dtype=flow.float32)
>>> other = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> out = flow.mul(input,other).numpy()
>>> out.shape
(2, 3)

oneflow.mv(input, vec)Tensor

The documentation is referenced from: https://pytorch.org/docs/1.10/generated/torch.mv.html.

Performs a matrix-vector product of the matrix input and the vector vec.

If input is a $$(n \times m)$$ tensor, vec is a 1-D tensor of size m, out will be a 1-D tensor of size n.

Note

Parameters
Returns

the output Tensor

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> mat = flow.randn(2, 3)
>>> vec = flow.randn(3)
>>> out = flow.mv(mat, vec)
>>> out.shape
oneflow.Size([2])

oneflow.narrow(x, dim: int, start: int, length: int)Tensor

Returns a new tensor that is a narrowed version of input tensor. The dimension dim is input from start to start + length.

Parameters
• input – the tensor to narrow.

• dim – the dimension along which to narrow.

• start – the starting dimension.

• length – the distance to the ending dimension.

For example:

>>> import oneflow as flow
>>> input = flow.tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9]])
>>> flow.narrow(input, 0, 0, 2)
tensor([[1, 2, 3],
[4, 5, 6]], dtype=oneflow.int64)
>>> flow.narrow(input, 1, 1, 2)
tensor([[2, 3],
[5, 6],
[8, 9]], dtype=oneflow.int64)

oneflow.ne(input, other)Tensor

Computes element-wise not equality. The second argument can be a number or a tensor whose shape is broadcastable with the first argument.

Parameters
Returns

• A boolean tensor that is True where input is not equal to other and False elsewhere

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> input = flow.tensor(np.array([2, 3, 4, 5]), dtype=flow.float32)
>>> other = flow.tensor(np.array([2, 3, 4, 1]), dtype=flow.float32)

>>> y = flow.ne(input, other)
>>> y
tensor([False, False, False,  True], dtype=oneflow.bool)

oneflow.neg()

This operator computes the negative value of Tensor.

Parameters

input (oneflow.Tensor) – A Tensor

Returns

The result Tensor

Return type

oneflow.Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> input = flow.tensor(
...    np.array([1.0, -1.0, 2.3]).astype(np.float32), dtype=flow.float32
... )
>>> out = flow.negative(input)
>>> out
tensor([-1.0000,  1.0000, -2.3000], dtype=oneflow.float32)

oneflow.negative()

This operator computes the negative value of Tensor.

Parameters

input (oneflow.Tensor) – A Tensor

Returns

The result Tensor

Return type

oneflow.Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> input = flow.tensor(
...    np.array([1.0, -1.0, 2.3]).astype(np.float32), dtype=flow.float32
... )
>>> out = flow.negative(input)
>>> out
tensor([-1.0000,  1.0000, -2.3000], dtype=oneflow.float32)

oneflow.new_ones(x, size=None, dtype=None, device=None, placement=None, sbp=None, requires_grad=False)Tensor

Returns a Tensor of size size filled with 1. By default, the returned Tensor has the same oneflow.dtype and oneflow.device as this tensor.

Parameters
• size (int...) – a list, tuple, or flow.Size of integers defining the shape of the output tensor.

• dtype (flow.dtype, optional) – the desired type of returned tensor. Default: if None, same flow.dtype as this tensor.

• device (flow.device, optional) – the desired device of returned tensor. Default: if None, same flow.device as this tensor.

• placement (flow.placement, optional) – the desired placement of returned global tensor. Default: if None, the returned tensor is local one using the argument device.

• sbp (flow.sbp.sbp or tuple of flow.sbp.sbp, optional) – the desired sbp descriptor of returned global tensor. Default: if None, the returned tensor is local one using the argument device.

• requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> x = flow.Tensor(np.ones((1, 2, 3)))
>>> y = x.new_ones((2, 2))
>>> y
tensor([[1., 1.],
[1., 1.]], dtype=oneflow.float32)

oneflow.nms(boxes, scores, iou_threshold: float)

Performs non-maximum suppression (NMS) on the boxes according to their intersection-over-union (IoU).

NMS iteratively removes lower scoring boxes which have an IoU greater than iou_threshold with another (higher scoring) box.

Parameters
• boxes (Tensor[N, 4]) – boxes to perform NMS on. They are expected to be in (x1, y1, x2, y2) format with 0 <= x1 < x2 and 0 <= y1 < y2.

• scores (Tensor[N]) – scores for each one of the boxes

• iou_threshold (float) – discards all overlapping boxes with IoU > iou_threshold

Returns

int64 tensor with the indices of the elements that have been kept by NMS, sorted in decreasing order of scores

Return type

Tensor

class oneflow.no_grad

Disabling gradient calculation is useful for inference, when you are sure that you will not call Tensor.backward(). It will reduce memory consumption for computations that would otherwise have requires_grad=True.

In this mode, the result of every computation will have requires_grad=False, even when the inputs have requires_grad=True.

This context manager is thread local; it will not affect computation in other threads.

Also functions as a decorator. (Make sure to instantiate with parenthesis.)

>>> import oneflow as flow
>>> x = flow.ones(2, 3, requires_grad=True)
...     y = x * x
False
...     return x * x
False

oneflow.nonzero(input, *, out=None, as_tuple=False)Tensor or tuple of Tensors

Note

When as_tuple is False (default): returns a 2-D tensor where each row is the index for a nonzero value.

When as_tuple is True: returns a tuple of 1-D index tensors, allowing for advanced indexing, so x[x.nonzero(as_tuple=True)] gives all nonzero values of tensor x. Of the returned tuple, each index tensor contains nonzero indices for a certain dimension.

See below for more details on the two behaviors.

When as_tuple is False (default):

Returns a tensor containing the indices of all non-zero elements of input. Each row in the result contains the indices of a non-zero element in input. The result is sorted lexicographically, with the last index changing the fastest (C-style).

If input has $$n$$ dimensions, then the resulting indices tensor out is of size $$(z \times n)$$, where $$z$$ is the total number of non-zero elements in the input tensor.

When as_tuple is True:

Returns a tuple of 1-D tensors, one for each dimension in input, each containing the indices (in that dimension) of all non-zero elements of input .

If input has $$n$$ dimensions, then the resulting tuple contains $$n$$ tensors of size $$z$$, where $$z$$ is the total number of non-zero elements in the input tensor.

As a special case, when input has zero dimensions and a nonzero scalar value, it is treated as a one-dimensional tensor with one element.

Parameters

input (Tensor) – the input tensor.

Keyword Arguments

out (Tensor, optional) – the output tensor containing indices

Returns

If as_tuple is False, the output tensor containing indices. If as_tuple is True, one 1-D tensor for each dimension, containing the indices of each nonzero element along that dimension.

Return type

Tensor or tuple of Tensors

Example:

>>> import oneflow as flow
>>> flow.nonzero(flow.tensor([1, 1, 1, 0, 1]))
tensor([[0],
[1],
[2],
[4]], dtype=oneflow.int64)
>>> flow.nonzero(flow.tensor([[0.6, 0.0, 0.0, 0.0],
...                             [0.0, 0.4, 0.0, 0.0],
...                             [0.0, 0.0, 1.2, 0.0],
...                             [0.0, 0.0, 0.0,-0.4]]))
tensor([[0, 0],
[1, 1],
[2, 2],
[3, 3]], dtype=oneflow.int64)
>>> flow.nonzero(flow.tensor([1, 1, 1, 0, 1]), as_tuple=True)
(tensor([0, 1, 2, 4], dtype=oneflow.int64),)
>>> flow.nonzero(flow.tensor([[0.6, 0.0, 0.0, 0.0],
...                             [0.0, 0.4, 0.0, 0.0],
...                             [0.0, 0.0, 1.2, 0.0],
...                             [0.0, 0.0, 0.0,-0.4]]), as_tuple=True)
(tensor([0, 1, 2, 3], dtype=oneflow.int64), tensor([0, 1, 2, 3], dtype=oneflow.int64))
>>> flow.nonzero(flow.tensor(5), as_tuple=True)
(tensor([0], dtype=oneflow.int64),)

oneflow.normal(mean, std, *size: Union[int, Tuple[int, ], oneflow.Size, List[int]], out=None, placement: Optional[oneflow._oneflow_internal.placement] = None, sbp: Optional[oneflow._oneflow_internal.sbp.sbp] = None, generator=None, dtype: Optional[oneflow._oneflow_internal.dtype] = None, device: Optional[Union[str, oneflow._oneflow_internal.device]] = None, requires_grad: bool = False)

Returns a tensor of random numbers drawn from separate normal distributions whose mean and standard deviation are given.

Parameters
• mean (float) – the mean for all distributions

• std (float) – the standard deviation for all distributions

• size (int...) – a sequence of integers defining the shape of the output tensor.

Keyword Arguments
• out (Tensor, optional) – the output tensor.

• placement (flow.placement, optional) – The desired device of returned global tensor. If None, will construct local tensor.

• sbp (flow.sbp, optional) – The desired sbp of returned global tensor. It must be equal with the numbers of placement.

• generator (oneflow.Generator, optional) – a pseudorandom number generator for sampling

• dtype (oneflow.dtype, optional) – the desired data type of returned tensor. Default: oneflow.float32.

• device – the desired device of returned tensor. Default: cpu.

• requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

Example:

>>> import oneflow as flow
>>> generator = flow.Generator()
>>> generator.manual_seed(0)
>>> y = flow.normal(0, 1, 5, generator=generator)
>>> y
tensor([2.2122, 1.1631, 0.7740, 0.4838, 1.0434], dtype=oneflow.float32)

oneflow.numel(input)int

Returns the total number of elements in the input tensor.

Parameters

input (oneflow.Tensor) – Input Tensor

>>> import oneflow as flow

>>> a = flow.randn(1, 2, 3, 4, 5)
>>> flow.numel(a)
120
>>> a = flow.zeros(4,4)
>>> flow.numel(a)
16

oneflow.ones(*size: Union[int, Tuple[int, ], oneflow.Size, List[int]], dtype: Optional[oneflow._oneflow_internal.dtype] = None, device: Optional[Union[str, oneflow._oneflow_internal.device]] = None, placement: Optional[oneflow._oneflow_internal.placement] = None, sbp: Optional[oneflow._oneflow_internal.sbp.sbp] = None, requires_grad: bool = False)

Returns a tensor filled with the scalar value 1, with the shape defined by the variable argument size.

Parameters
• size (an integer or tuple of integer values) – defining the shape of the output tensor. Can be a variable number of arguments or a collection like a list or tuple.

• dtype (flow.dtype, optional) – the desired data type of returned tensor.

• device (flow.device, optional) – the desired device of returned tensor. Default: if None, uses the current device for the default tensor type

• placement (flow.placement, optional) – the desired placement of returned global tensor. Default: if None, the returned tensor is local one using the argument device.

• sbp (flow.sbp.sbp or tuple of flow.sbp.sbp, optional) – the desired sbp descriptor of returned global tensor. Default: if None, the returned tensor is local one using the argument device.

• requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

For example:

>>> import oneflow as flow
>>> y = flow.ones(5)
>>> y
tensor([1., 1., 1., 1., 1.], dtype=oneflow.float32)
>>> y = flow.ones(2,3) # construct local tensor
>>> y
tensor([[1., 1., 1.],
[1., 1., 1.]], dtype=oneflow.float32)
>>> placement = flow.placement("cpu", ranks=[0])
>>> y = flow.ones(4, 5, placement=placement, sbp=flow.sbp.broadcast) # construct global tensor
>>> y.is_global
True

oneflow.ones_like(input) is equivalent to flow.ones(input.shape, dtype=input.dtype)
Parameters

other (Tensor) – The size of input will determine size of the output tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> x = flow.tensor(np.random.rand(5), dtype=flow.float32)
>>> y = flow.ones_like(x)
>>> y
tensor([1., 1., 1., 1., 1.], dtype=oneflow.float32)

oneflow.permute(input, *dims)Tensor

Returns a view of the original tensor with its dimensions permuted.

Parameters

dims (tuple of ints) – The desired ordering of dimensions

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> input = flow.tensor(np.random.randn(2, 6, 5, 3), dtype=flow.float32)
>>> output = flow.permute(input, (1, 0, 2, 3)).shape
>>> output
oneflow.Size([6, 2, 5, 3])

oneflow.pow()
Takes the power of each element in input with exponent and returns a tensor with the result. Exponent can be either a single float number, a single int number, or a tensor with the same shape as input.

When exponent is a scalar value, the operation applied is:

$\begin{split}\\text{out}_i = x_i ^ \\text{exponent}\end{split}$
u200b

When exponent is a tensor, the operation applied is:

$\begin{split}\\text{out}_i = x_i ^ {\\text{exponent}_i}\end{split}$
Args:
• input (Tensor): the input tensor.

• exponent (int, float, Tensor): the exponent.

Returns:

Tensor: The result of variance on the specified axis of input Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> x = flow.tensor(np.array([1.0, 2.0, 3.0, 4.0, 5.0, 6.0]), dtype=flow.float32)
>>> out = flow.pow(x, 2)
>>> out
tensor([ 1.,  4.,  9., 16., 25., 36.], dtype=oneflow.float32)

>>> x = flow.tensor(np.array([1.0, 2.0, 3.0, 4.0]), dtype=flow.float32)
>>> y = flow.tensor(np.array([1.0, 2.0, 3.0, 4.0]), dtype=flow.float32)
>>> out = flow.pow(x, y)
>>> out
tensor([  1.,   4.,  27., 256.], dtype=oneflow.float32)

oneflow.prod(input, dim=None, keepdim=False)Tensor

Computes the product of row of elements in a tensor in the given dimension. If the dimension is None, product of all elements will be caculated.

If keepdim is True, the output tensor is of the same size as input except in the dimension(s) dim where it is of size 1. Otherwise, dim is squeezed oneflow.squeeze(), resulting in the output tensor having 1 (or len(dim)) fewer dimension(s).

Parameters
• input (oneflow.Tensor) – the Input Tensor

• dim (int or tuple of ints, optional) – the dimension to reduce. Default: None

• keepdim (bool, optional) – whether the output tensor has dim retained or not. Default: False

For example:

>>> import oneflow as flow

>>> input = flow.Tensor([[1, 2, 3], [4, 5, 6]])
>>> flow.prod(input)
tensor(720., dtype=oneflow.float32)
>>> flow.prod(input, dim=0)
tensor([ 4., 10., 18.], dtype=oneflow.float32)
>>> flow.prod(input, dim=1)
tensor([  6., 120.], dtype=oneflow.float32)

oneflow.rand(*size, out=None, generator=None, dtype: Optional[oneflow._oneflow_internal.dtype] = None, layout=None, device: Optional[Union[str, oneflow._oneflow_internal.device]] = None, placement: Optional[oneflow._oneflow_internal.placement] = None, sbp: Optional[oneflow._oneflow_internal.sbp.sbp] = None, requires_grad: bool = False)

Returns a tensor filled with random numbers from a uniform distribution on the interval [0, 1)

The shape of the tensor is defined by the variable argument size.

Parameters
• size (int... or oneflow.Size) – Defining the shape of the output tensor. Can be a variable number of arguments or a collection like a list or tuple or oneflow.Size.

• out (optional) – The output tensor.

• dtype (flow.dtype, optional) – The desired data type of returned tensor. Default: flow.float32.

• layout (optional) – The desired layout of returned Tensor.

• generator (flow.Generator, optional) – a pseudorandom number generator for sampling

• device (flow.device, optional) – The desired device of returned local tensor. If None, uses the current device.

• placement (flow.placement, optional) – The desired device of returned global tensor. If None, will construct local tensor.

• sbp (flow.sbp, optional) – The desired sbp of returned global tensor. It must be equal with the numbers of placement.

• requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

For example:

>>> import oneflow as flow
>>> x = flow.rand(3,3) # construct local tensor
>>> x.shape
oneflow.Size([3, 3])
>>> x.is_global
False
>>> placement = flow.placement("cpu", ranks=[0])
>>> x = flow.rand(3, 3, placement=placement, sbp=sbp) # construct global tensor
>>> x.is_global
True

oneflow.randint(low: oneflow.int64, high: oneflow.int64, size: tuple, out=None, generator=None, dtype: Optional[oneflow._oneflow_internal.dtype] = None, layout=None, device: Optional[Union[str, oneflow._oneflow_internal.device]] = None, placement: Optional[oneflow._oneflow_internal.placement] = None, sbp: Optional[oneflow._oneflow_internal.sbp.sbp] = None, requires_grad: bool = False)

Returns a tensor filled with random integers generated uniformly between low (inclusive) and high (exclusive).

The shape of the tensor is defined by the variable argument size.

Parameters
• size (int... or oneflow.Size) – Defining the shape of the output tensor. Can be a variable number of arguments or a collection like a list or tuple or oneflow.Size.

• out (optional) – The output tensor.

• dtype (flow.dtype, optional) – The desired data type of returned tensor. Default: flow.int64.

• layout (optional) – The desired layout of returned Tensor.

• generator (flow.Generator, optional) –

• device (flow.device, optional) – The desired device of returned local tensor. If None, uses the current device.

• placement (flow.placement, optional) – The desired device of returned global tensor. If None, will construct local tensor.

• sbp (flow.sbp, optional) – The desired sbp of returned global tensor. It must be equal with the numbers of placement.

• requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

For example:

>>> import oneflow as flow
>>> generator = flow.Generator()
>>> generator.manual_seed(0)
>>> y = flow.randint(0, 5, (3,3), generator=generator) # construct local tensor
>>> y
tensor([[2, 2, 3],
[4, 3, 4],
[2, 4, 2]], dtype=oneflow.int64)
>>> y.is_global
False
>>> placement = flow.placement("cpu", ranks=[0])
>>> y = flow.randint(0, 5, (3,3), generator=generator, placement=placement, sbp=flow.sbp.broadcast) # construct global tensor
>>> y.is_global
True

oneflow.randn(*size, out=None, generator=None, dtype: Optional[oneflow._oneflow_internal.dtype] = None, layout=None, device: Optional[Union[str, oneflow._oneflow_internal.device]] = None, placement: Optional[oneflow._oneflow_internal.placement] = None, sbp: Optional[oneflow._oneflow_internal.sbp.sbp] = None, requires_grad: bool = False)

Returns a tensor filled with random numbers from a normal distribution with mean 0 and variance 1 (also called the standard normal distribution).

The shape of the tensor is defined by the variable argument size.

Parameters
• size (int... or oneflow.Size) – Defining the shape of the output tensor. Can be a variable number of arguments or a collection like a list or tuple or oneflow.Size.

• out (optional) – The output tensor.

• dtype (flow.dtype, optional) – The desired data type of returned tensor. Default: flow.float32.

• layout (optional) – The desired layout of returned Tensor.

• generator (flow.Generator, optional) – a pseudorandom number generator for sampling

• device (flow.device, optional) – The desired device of returned local tensor. If None, uses the current device.

• placement (flow.placement, optional) – The desired device of returned global tensor. If None, will construct local tensor.

• sbp (flow.sbp, optional) – The desired sbp of returned global tensor. It must be equal with the numbers of placement.

• requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

For example:

>>> import oneflow as flow
>>> x = flow.randn(3,3) # construct local tensor
>>> x.shape
oneflow.Size([3, 3])
>>> x.is_global
False
>>> placement = flow.placement("cpu", ranks=[0])
>>> x = flow.randn(3,3,placement=placement,sbp=sbp) # construct global tensor
>>> x.is_global
True

oneflow.randperm(n: oneflow.int64, generator: Optional[oneflow.framework.generator.create_generator] = None, out=None, dtype: Optional[oneflow._oneflow_internal.dtype] = None, layout=None, device: Optional[Union[str, oneflow._oneflow_internal.device]] = None, placement: Optional[oneflow._oneflow_internal.placement] = None, sbp: Optional[oneflow._oneflow_internal.sbp.sbp] = None, requires_grad: bool = False, pin_memory: bool = False)oneflow.Tensor

Returns a random permutation of integers from 0 to n - 1.

Parameters

n (int) – the upper bound (exclusive)

Keyword Arguments
• generator (oneflow.Generator, optional) – a pseudorandom number generator for sampling

• out (Tensor, optional) – output Tensor,not supported yet.

• dtype (oneflow.dtype, optional) – the desired data type of returned tensor. Default: oneflow.int64.

• layout – layout is not supported yet.

• device – the desired device of returned tensor. Default: cpu.

• placement – (flow.placement, optional): The desired device of returned global tensor. If None, will construct local tensor.

• sbp – (flow.sbp, optional): The desired sbp of returned global tensor. It must be equal with the numbers of placement.

• requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

• pin_memory (bool, optional) – pin_memory is not supported yet.

Example:

>>> import oneflow as flow
>>> generator = flow.Generator()
>>> generator.manual_seed(0)
>>> y = flow.randperm(5, generator=generator) # construct local tensor
>>> y
tensor([2, 4, 3, 0, 1], dtype=oneflow.int64)
>>> y.is_global
False
>>> placement = flow.placement("cpu", ranks=[0])
>>> y = flow.randperm(5, generator=generator, placement=placement, sbp=flow.sbp.broadcast) # construct global tensor
>>> y.is_global
True

oneflow.reciprocal()

Computes the safe reciprocal of x. If x is zero, the reciprocal will be also set to zero.

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> x = flow.tensor(np.array([[1, 2, 3], [4, 5, 6]]), dtype=flow.float32)
>>> out = flow.reciprocal(x)
>>> out.numpy()
array([[1.        , 0.5       , 0.33333334],
[0.25      , 0.2       , 0.16666667]], dtype=float32)

oneflow.repeat(input, sizes)Tensor

This operator repeat the input tensor to a larger size along the specified dimensions.

Parameters
• input (oneflow.Tensor) – the input Tensor.

• sizes (flow.Shape or List) – The number of times to repeat this tensor along each dimension.

Returns

The result Tensor.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> np_arr = np.random.randn(5, 3, 6, 9).astype(np.float32)
>>> input = flow.Tensor(np_arr)
>>> out = input.repeat(1, 1, 2, 2)
>>> out.shape
oneflow.Size([5, 3, 12, 18])
>>> out = input.repeat(2, 1, 1, 2, 2)
>>> out.shape
oneflow.Size([2, 5, 3, 12, 18])

oneflow.repeat_interleave(input, repeats, dim=None, *, output_size=None)Tensor

The documentation is referenced from: https://pytorch.org/docs/stable/generated/torch.repeat_interleave.html

Repeat elements of a tensor.

Warning

This is different from oneflow.Tensor.repeat() but similar to numpy.repeat.

Parameters
• input (oneflow.Tensor) – the input Tensor.

• repeats (Tensor or int) – The number of repetitions for each element. repeats is broadcasted to fit the shape of the given axis.

• dim (int, optional) – The dimension along which to repeat values. By default, use the flattened input array, and return a flat output array.

Keyword Arguments

output_size (int, optional) – Total output size for the given axis ( e.g. sum of repeats). If given, it will avoid stream syncronization needed to calculate output shape of the tensor.

Returns

Repeated tensor which has the same shape as input, except along the given axis.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> x = flow.tensor([1, 2, 3])
>>> y = flow.tensor([[1, 2], [3, 4]])
>>> flow.repeat_interleave(y, 2)
tensor([1, 1, 2, 2, 3, 3, 4, 4], dtype=oneflow.int64)
>>> flow.repeat_interleave(y, 3, dim=1)
tensor([[1, 1, 1, 2, 2, 2],
[3, 3, 3, 4, 4, 4]], dtype=oneflow.int64)
>>> flow.repeat_interleave(y, flow.tensor([1, 2]), dim=0)
tensor([[1, 2],
[3, 4],
[3, 4]], dtype=oneflow.int64)
>>> flow.repeat_interleave(y, flow.tensor([1, 2]), dim=0, output_size=3)
tensor([[1, 2],
[3, 4],
[3, 4]], dtype=oneflow.int64)

oneflow.reshape(input, shape: Optional[Sequence[int]] = None)

This operator reshapes a Tensor.

We can set one dimension in shape as -1, the operator will infer the complete shape.

Parameters
• x – A Tensor.

• shape – Shape of the output tensor.

Returns

A Tensor has the same type as x.

For example:

>>> import numpy as np
>>> import oneflow as flow
>>> x = np.array(
...    [[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16]]
... ).astype(np.float32)
>>> input = flow.Tensor(x)

>>> y = flow.reshape(input, shape=[2, 2, 2, -1]).shape
>>> y
oneflow.Size([2, 2, 2, 2])

oneflow.roc_auc_score()
oneflow.roll(input, shifts, dims=None)

Roll the tensor along the given dimension(s).

Elements that are shifted beyond the last position are re-introduced at the first position.

If a dimension is not specified, the tensor will be flattened before rolling and then restored to the original shape.

Parameters
• input (oneflow.Tensor) – the input Tensor.

• shifts (int or tuple of ints) – The number of places by which the elements of the tensor are shifted. If shifts is a tuple, dims must be a tuple of the same size, and each dimension will be rolled by the corresponding value.

• dims (int or tuple of ints) – Axis along which to roll.

Returns

The result Tensor.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> x = np.array([[1, 2],
...               [3, 4],
...               [5, 6],
...               [7, 8]])
>>> input = flow.Tensor(x)
>>> input.shape
oneflow.Size([4, 2])
>>> out = flow.roll(input, 1, 0)
>>> out
tensor([[7., 8.],
[1., 2.],
[3., 4.],
[5., 6.]], dtype=oneflow.float32)
>>> input.roll(-1, 1)
tensor([[2., 1.],
[4., 3.],
[6., 5.],
[8., 7.]], dtype=oneflow.float32)

oneflow.round()

This operator rounds the value of Blob to the nearest integer. :param input: A Tensor :type input: oneflow.Tensor

Returns

The result Tensor

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> x1 = flow.tensor(np.array([1.49999, 1.500001, 2.7]).astype(np.float32))
>>> out1 = flow.round(x1)
>>> out1.numpy()
array([1., 2., 3.], dtype=float32)
>>> x2 = flow.tensor(np.array([2.499999, 7.5000001, 5.3, 6.8]).astype(np.float32))
>>> out2 = flow.round(x2)
>>> out2.numpy()
array([2., 8., 5., 7.], dtype=float32)

oneflow.rsqrt()

Returns a new tensor with the reciprocal of the square-root of each of the elements of input.

$\text{out}_{i} = \frac{1}{\sqrt{\text{input}_{i}}}$
Parameters

input – the input tensor.

>>> import oneflow as flow
>>> import numpy as np

>>> a = flow.tensor(np.array([1.0, 2.0, 3.0]), dtype=flow.float32)
>>> out = flow.rsqrt(a).numpy()
>>> out
array([1.        , 0.70710677, 0.57735026], dtype=float32)

oneflow.save(obj: Any, path: Union[str, pathlib.Path], global_dst_rank: Optional[int] = None)None

Save an object to a directory.

Parameters
• obj – The object to be saved

• path (str) – The directory in which the object is saved

• global_dst_rank (int, optional) – The destination rank for saving global tensors. When specified, whole tensors will be saved by the process whose rank == global_src_rank, while other processes will not do any disk I/O.

oneflow.scatter(input, dim, index, src)

This operator writes the elements specified by index along with the axis dim from the src into the input.

Take a 3-D blob as example, the output is specified by:

input[index[i][j][k]][j][k] = src[i][j][k]  # if dim == 0
input[i][index[i][j][k]][k] = src[i][j][k]  # if dim == 1
input[i][j][index[i][j][k]] = src[i][j][k]  # if dim == 2


input, index and src (if it is a Tensor) should all have the same number of dimensions. It is also required that index.shape(d) <= src.shape(d) for all dimensions d, and that index.shape(d) <= self.shape(d) for all dimensions d != dim. Note that index and src do not broadcast.

Parameters
• input (Tensor) – The input blob.

• dim (int) – The axis along which to index

• index (Tensor) – The index blob of elements to scatter.

• src (Tensor or float) – The source blob whose elements will be scatterd and updated to output.

Returns

The scatterd Tensor.

Return type

Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> input = flow.ones((3,5))*2
>>> index = flow.tensor(np.array([[0,1,2],[0,1,4]], ), dtype=flow.int32)
>>> src = flow.Tensor(np.array([[0,10,20,30,40],[50,60,70,80,90]]))
>>> out = flow.scatter(input, 1, index, src)
>>> out
tensor([[ 0., 10., 20.,  2.,  2.],
[50., 60.,  2.,  2., 70.],
[ 2.,  2.,  2.,  2.,  2.]], dtype=oneflow.float32)

oneflow.scatter_add(input, dim, index, src)

This operator scatter the src with addition operation according to index along dim into the input.

Take a 3-D blob as example, the output is specified by:

input[index[i][j][k]][j][k] += src[i][j][k]  # if dim == 0
input[i][index[i][j][k]][k] += src[i][j][k]  # if dim == 1
input[i][j][index[i][j][k]] += src[i][j][k]  # if dim == 2

Parameters
• input (Tensor) – The input blob.

• dim (int) – The axis along which to index

• index (Tensor) – The index blob of elements to scatter.

• src (Tensor) – The source blob whose elements will be scatterd and added to output.

Returns

The scatterd Tensor.

Return type

Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> input = flow.ones((3,5))*2
>>> index = flow.tensor(np.array([[0,1,2],[0,1,4]], ), dtype=flow.int32)
>>> src = flow.Tensor(np.array([[0,10,20,30,40],[50,60,70,80,90]]))
>>> out = flow.scatter_add(input, 1, index, src)
>>> out
tensor([[ 2., 12., 22.,  2.,  2.],
[52., 62.,  2.,  2., 72.],
[ 2.,  2.,  2.,  2.,  2.]], dtype=oneflow.float32)

oneflow.scatter_nd(index, update, shape)

This operator inserts the elements in update according to the index and create a new Tensor.

Parameters
• index – The indices of update. Its type should be flow.int.

• update – The update Tensor.

• shape (Sequence[int]) – The constant tensor shape, the constant tensor elements are all zero.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> index = flow.tensor(np.array([[1], [6], [4]]), dtype=flow.int)
>>> update = flow.tensor(np.array([10.2, 5.1, 12.7]), dtype=flow.float)
>>> out = flow.scatter_nd(index, update, [8])
>>> out
tensor([ 0.0000, 10.2000,  0.0000,  0.0000, 12.7000,  0.0000,  5.1000,  0.0000],
dtype=oneflow.float32)

oneflow.searchsorted()oneflow.Tensor

The documentation is referenced from: https://pytorch.org/docs/1.10/generated/torch.searchsorted.html?highlight=searchsorted

Find the indices from the innermost dimension of sorted_sequence such that, if the corresponding values in values were inserted before the indices, the order of the corresponding innermost dimension within sorted_sequence would be preserved. Return a new tensor with the same size as values. If right is False (default), then the left boundary of sorted_sequence is closed. More formally, the returned index satisfies the following rules:

sorted_sequence

right

returned index satisfies

1-D

False

sorted_sequence[i-1] < values[m][n]…[l][x] <= sorted_sequence[i]

1-D

True

sorted_sequence[i-1] <= values[m][n]…[l][x] < sorted_sequence[i]

N-D

False

sorted_sequence[m][n]…[l][i-1] < values[m][n]…[l][x]

<= sorted_sequence[m][n]…[l][i]

N-D

True

sorted_sequence[m][n]…[l][i-1] <= values[m][n]…[l][x]

sorted_sequence[m][n]…[l][i]

Parameters
• sorted_sequence (Tensor) – N-D or 1-D tensor, containing monotonically increasing sequence on the innermost dimension.

• values (Tensor or Scalar) – N-D tensor or a Scalar containing the search value(s).

• out_int32 (bool optional) – indicate the output data type. torch.int32 if True, torch.int64 otherwise. Default value is False, i.e. default output data type is torch.int64.

• right (bool optional) – if False, return the first suitable location that is found. If True, return the last such index. If no suitable index found, return 0 for non-numerical value (eg. nan, inf) or the size of innermost dimension within sorted_sequence (one pass the last index of the innermost dimension). In other words, if False, gets the lower bound index for each value in values on the corresponding innermost dimension of the sorted_sequence. If True, gets the upper bound index instead. Default value is False.

For example:

>>> import oneflow as flow
>>> sorted_sequence = flow.tensor([[1, 3, 5, 7, 9], [2, 4, 6, 8, 10]])
>>> sorted_sequence
tensor([[ 1,  3,  5,  7,  9],
[ 2,  4,  6,  8, 10]], dtype=oneflow.int64)
>>> values = flow.tensor([[3, 6, 9], [3, 6, 9]])
>>> values
tensor([[3, 6, 9],
[3, 6, 9]], dtype=oneflow.int64)
>>> flow.searchsorted(sorted_sequence, values)
tensor([[1, 3, 4],
[1, 2, 4]], dtype=oneflow.int64)
>>> flow.searchsorted(sorted_sequence, values, right=True)
tensor([[2, 3, 5],
[1, 3, 4]], dtype=oneflow.int64)
>>> sorted_sequence_1d = flow.tensor([1, 3, 5, 7, 9])
>>> sorted_sequence_1d
tensor([1, 3, 5, 7, 9], dtype=oneflow.int64)
>>> flow.searchsorted(sorted_sequence_1d, values)
tensor([[1, 3, 4],
[1, 3, 4]], dtype=oneflow.int64)

oneflow.seed()int

The documentation is referenced from: https://pytorch.org/docs/1.10/generated/torch.seed.html.

Sets the seed for generating random numbers to a non-deterministic random number. Returns a 64 bit number used to seed the RNG.

oneflow.select()

Slices the self tensor along the selected dimension at the given index. This function returns a view of the original tensor with the given dimension removed.

Parameters
• input (Tensor) – the input tensor.

• dim (int) – the dimension to slice.

• select (int) – the index to select with.

Returns

the output Tensor.

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> input = flow.rand(3, 4, 5)
>>> out = flow.select(input, 0, 1)
>>> out.size()
oneflow.Size([4, 5])
>>> out = flow.select(input, 1, 1)
>>> out.size()
oneflow.Size([3, 5])

oneflow.selu(x: Tensor)Tensor

Applies element-wise function :math: ext{SELU}(x) = scale * (max(0,x) + min(0, lpha * (exp(x) - 1))), with $$lpha=1.6732632423543772848170429916717$$ and $$scale=1.0507009873554804934193349852946$$.

See SELU for more details.

For examples:

>>> import numpy as np
>>> import oneflow as flow

>>> x = np.array([1, 2, 3]).astype(np.float32)
>>> input = flow.tensor(x)
>>> out = flow.nn.functional.selu(input)
>>> out
tensor([1.0507, 2.1014, 3.1521], dtype=oneflow.float32)

class oneflow.set_grad_enabled(is_train=True)

This context manager is thread local; it will not affect computation in other threads.

Also functions as a decorator. (Make sure to instantiate with parenthesis.)

Parameters

mode (bool) – Flag whether to enable or disable gradient calculation. (default: True)

>>> import oneflow as flow
>>> x = flow.ones(2, 3, requires_grad=True)
...     y = x * x
True
...     return x * x
False

oneflow.set_printoptions(precision=None, threshold=None, edgeitems=None, linewidth=None, profile=None, sci_mode=None)

Set options for printing. Items shamelessly taken from NumPy

Parameters
• precision – Number of digits of precision for floating point output (default = 4).

• threshold – Total number of array elements which trigger summarization rather than full repr (default = 1000).

• edgeitems – Number of array items in summary at beginning and end of each dimension (default = 3).

• linewidth – The number of characters per line for the purpose of inserting line breaks (default = terminal_columns).

• profile – Sane defaults for pretty printing. Can override with any of the above options. (any one of default, short, full)

• sci_mode – Enable (True) or disable (False) scientific notation. If None (default) is specified, the value is defined by oneflow._tensor_str._Formatter. This value is automatically chosen by the framework.

Note

linewidth equals to terminal columns, manual setting will invalidate the default automatic setting.

oneflow.set_rng_state(state)

The documentation is referenced from: https://pytorch.org/docs/1.10/generated/torch.set_rng_state.html.

Returns the random number generator state as a oneflow.ByteTensor.

oneflow.sigmoid(input)Tensor

Applies the element-wise function $$\text{Sigmoid}(x) = \frac{1}{1 + \exp(-x)}$$

See Sigmoid for more details.

For examples:

>>> import numpy as np
>>> import oneflow as flow

>>> x = np.array([0.81733328, 0.43621480, 0.10351428])
>>> input = flow.tensor(x, dtype=flow.float32)
>>> out = flow.nn.functional.sigmoid(input)
>>> out
tensor([0.6937, 0.6074, 0.5259], dtype=oneflow.float32)

oneflow.sign()

Computes the sign of Tensor.

$\text{out}_{i} = \text{sgn}(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> x1 = flow.tensor(np.array([-2, 0, 2]).astype(np.float32))
>>> out1 = flow.sign(x1)
>>> out1.numpy()
array([-1.,  0.,  1.], dtype=float32)
>>> x2 = flow.tensor(np.array([-3.2, -4.5, 5.8]).astype(np.float32),device=flow.device('cuda'))
>>> out2 = flow.sign(x2)
>>> out2.numpy()
array([-1., -1.,  1.], dtype=float32)

oneflow.silu(x: Tensor)Tensor

The formula is:

$ext{silu}(x) = x * sigmoid(x)$

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> x = np.array([1, 2, 3]).astype(np.float32)
>>> input = flow.tensor(x)
>>> out = flow.silu(input)
>>> out
tensor([0.7311, 1.7616, 2.8577], dtype=oneflow.float32)


See SiLU for more details.

oneflow.sin()

Returns a new tensor with the sine of the elements of input.

sin(x: Tensor) -> Tensor

$\text{y}_{i} = \sin(\text{x}_{i})$
Parameters

x (Tensor) – the input tensor.

For example: .. code-block:: python

>>> import oneflow as flow
>>> import numpy as np
>>> x1 = flow.tensor(np.array([-0.5461,  0.1347, -2.7266, -0.2746]).astype(np.float32))
>>> y1 = flow.sin(x1)
>>> y1
tensor([-0.5194,  0.1343, -0.4032, -0.2712], dtype=oneflow.float32)

>>> x2 = flow.tensor(np.array([-1.4, 2.6, 3.7]).astype(np.float32), device=flow.device('cuda'))
>>> y2 = flow.sin(x2)
>>> y2
tensor([-0.9854,  0.5155, -0.5298], device='cuda:0', dtype=oneflow.float32)

oneflow.sin_()

In-place version of oneflow.sin()

oneflow.sinh()

Returns a new tensor with the hyperbolic sine of the elements of input.

$\text{out}_{i} = \sinh(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> x1 = flow.tensor(np.array([1, 2, 3]), dtype=flow.float32)
>>> x2 = flow.tensor(np.array([1.53123589,0.54242598,0.15117185]), dtype=flow.float32)
>>> x3 = flow.tensor(np.array([1,0,-1]), dtype=flow.float32)

>>> flow.sinh(x1).numpy()
array([ 1.1752012,  3.6268604, 10.017875 ], dtype=float32)
>>> flow.sinh(x2).numpy()
array([2.20381  , 0.5694193, 0.1517483], dtype=float32)
>>> flow.sinh(x3).numpy()
array([ 1.1752012,  0.       , -1.1752012], dtype=float32)

oneflow.slice(input, slice_tup_list: Sequence[Tuple[int, int, int]])

Extracts a slice from a tensor. The slice_tup_list assigns the slice indices in each dimension, the format is (start, stop, step). The operator will slice the tensor according to the slice_tup_list.

Parameters
• input – A Tensor.

• slice_tup_list – A list of slice tuple, indicate each dimension slice (start, stop, step).

For example:

>>> import numpy as np
>>> import oneflow as flow
>>> input = flow.Tensor(np.random.randn(3, 6, 9).astype(np.float32))
>>> tup_list = [[None, None, None], [0, 5, 2], [0, 6, 3]]
>>> y = flow.slice(input, slice_tup_list=tup_list)
>>> y.shape
oneflow.Size([3, 3, 2])

oneflow.slice_update(input, update, slice_tup_list: Sequence[Tuple[int, int, int]])

Update a slice of tensor x. Like x[start:stop:step] = update.

Parameters
• x – A Tensor, whose slice will be updated.

• update – A Tensor, indicate the update content.

• slice_tup_list – A list of slice tuple, indicate each dimension slice (start, stop, step).

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> input = flow.Tensor(np.array([1, 1, 1, 1, 1]).astype(np.float32))
>>> update = flow.Tensor(np.array([2, 3, 4]).astype(np.float32))
>>> flow.slice_update(input, update, slice_tup_list=[[1, 4, 1]])
tensor([1., 2., 3., 4., 1.], dtype=oneflow.float32)

oneflow.softmax(x: Tensor, dim: int)Tensor

Softmax is defined as:

$\begin{split}\text{Softmax}(x_{i}) = \frac{\\exp(x_i)}{\sum_j \exp(x_j)}\end{split}$

See Softmax for more details.

oneflow.softplus(x: Tensor, beta: double = 1, threshold: double = 20)Tensor

Applies the element-wise function:

$\text{Softplus}(x) = \frac{1}{\beta} * \log(1 + \exp(\beta * x))$

For numerical stability the implementation reverts to the linear function when $$input \times \beta > threshold$$.

See Softplus for more details.

oneflow.softsign(x: Tensor)Tensor

The formula is:

$softsign(x) = \frac{x}{1 + |x|}$

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> x = np.array([1, 2, 3]).astype(np.float32)
>>> input = flow.tensor(x)
>>> out = flow.nn.functional.softsign(input)
>>> out
tensor([0.5000, 0.6667, 0.7500], dtype=oneflow.float32)


See Softsign for more details.

oneflow.sort(input, dim: int = - 1, descending: bool = False)

Sorts the elements of the input tensor along a given dimension in ascending order by value.

Parameters
• input (oneflow.Tensor) – the input Tensor.

• dim (int, optional) – the dimension to be sorted. Defaults to the last dim (-1).

• descending (bool, optional) – controls the sorting order (ascending or descending).

Returns

A tuple of (values, indices), where where the values are the sorted values and the indices are the indices of the elements in the original input tensor.

Return type

Tuple(oneflow.Tensor, oneflow.Tensor(dtype=int32))

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> x = np.array([[1, 3, 8, 7, 2], [1, 9, 4, 3, 2]], dtype=np.float32)
>>> input = flow.Tensor(x)
>>> (values, indices) = flow.sort(input)
>>> values
tensor([[1., 2., 3., 7., 8.],
[1., 2., 3., 4., 9.]], dtype=oneflow.float32)
>>> indices
tensor([[0, 4, 1, 3, 2],
[0, 4, 3, 2, 1]], dtype=oneflow.int32)
>>> (values, indices) = flow.sort(input, descending=True)
>>> values
tensor([[8., 7., 3., 2., 1.],
[9., 4., 3., 2., 1.]], dtype=oneflow.float32)
>>> indices
tensor([[2, 3, 1, 4, 0],
[1, 2, 3, 4, 0]], dtype=oneflow.int32)
>>> (values, indices) = flow.sort(input, dim=0)
>>> values
tensor([[1., 3., 4., 3., 2.],
[1., 9., 8., 7., 2.]], dtype=oneflow.float32)
>>> indices
tensor([[0, 0, 1, 1, 0],
[1, 1, 0, 0, 1]], dtype=oneflow.int32)

oneflow.split()

Splits the tensor into chunks.

If split_size_or_sections is an integer type, then x will be split into equally sized chunks (if possible). Last chunk will be smaller if the tensor size along the given dimension dim is not divisible by split_size.

If split_size_or_sections is a list, then x will be split into len(split_size_or_sections) chunks with sizes in dim according to split_size_or_sections.

Parameters
• x – tensor to split.

• split_size_or_sections – size of a single chunk or list of sizes for each chunk.

• dim – dimension along which to split the tensor.

For example:

>>> import oneflow as flow
>>> a = flow.arange(10).view(5, 2)
>>> flow.split(a, 2)
(tensor([[0, 1],
[2, 3]], dtype=oneflow.int64), tensor([[4, 5],
[6, 7]], dtype=oneflow.int64), tensor([[8, 9]], dtype=oneflow.int64))
>>> flow.split(a, [1, 4])
(tensor([[0, 1]], dtype=oneflow.int64), tensor([[2, 3],
[4, 5],
[6, 7],
[8, 9]], dtype=oneflow.int64))

oneflow.sqrt()

Returns a new tensor with the square-root of the elements of input.

$\text{out}_{i} = \sqrt{\text{input}_{i}}$
Parameters

input – the input tensor.

>>> import oneflow as flow
>>> import numpy as np

>>> arr = np.array([1.0, 2.0, 3.0])
>>> input = flow.tensor(arr, dtype=flow.float32)
>>> output = flow.sqrt(input).numpy()
>>> output
array([1.       , 1.4142135, 1.7320508], dtype=float32)

oneflow.square()

Returns a new tensor with the square of the elements of input.

$\text{out}_{i} = \sqrt{\text{input}_{i}}$
Parameters

input – the input tensor.

>>> import oneflow as flow
>>> import numpy as np

>>> arr = np.array([1.0, 2.0, 3.0])
>>> input = flow.tensor(arr, dtype=flow.float32)
>>> output = flow.square(input).numpy()
>>> output
array([1., 4., 9.], dtype=float32)

oneflow.squeeze()

This operator removes the specified dimention which size is 1 of the input Tensor. If the dim is not specified, this operator will remove all the dimention which size is 1 of the input Tensor.

The amount of element in return value is the same as Tensor input.

Parameters
• input (oneflow.Tensor) – the input Tensor.

• dim (int, optinal) – Defaults to None, if given, the input will be squeezed only in this dimension.

Returns

The result Tensor.

Return type

Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> input = flow.tensor(np.array([[[[1, 1, 1]]]]).astype(np.int32))
>>> input.shape
oneflow.Size([1, 1, 1, 3])
>>> out = flow.squeeze(input, dim=[1, 2]).shape
>>> out
oneflow.Size([1, 3])

oneflow.stack()

Concatenates a sequence of tensors along a new dimension. The returned tensor shares the same underlying data with input tensors.

A dim value within the range [-input.ndimension() - 1, input.ndimension() + 1] can be used. Negative dim will correspond to stack() applied at dim = dim + input.ndimension() + 1.

Parameters
• inputs (List[oneflow.Tensor]) – the list of input tensors. Each tensor should have the same shape.

• dim (int) – the index at which to insert the concatenated dimension.

Returns

A Tensor

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> x1 = flow.tensor(np.random.rand(1, 3, 5))
>>> x2 = flow.tensor(np.random.rand(1, 3, 5))
>>> y = flow.stack([x1, x2], dim = -1)
>>> y.shape
oneflow.Size([1, 3, 5, 2])

oneflow.std()

Returns the standard-deviation of each row of the input tensor in the dimension dim. If dim is a list of dimensions, reduce over all of them.

If keepdim is True, the output tensor is of the same size as input except in the dimension(s) dim where it is of size 1. Otherwise, dim is squeezed, resulting in the output tensor having 1 (or len(dim)) fewer dimension(s).

If unbiased is False, then the standard-deviation will be calculated via the biased estimator. Otherwise, Bessel’s correction will be used.

Parameters
• input (Tensor) – the input tensor.

• dim (int or tuple of ints) – the dimension or dimensions to reduce.

• unbiased (bool) – whether to use the unbiased estimation or not

• keepdim (bool) – whether the output tensor has dim retained or not.

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> arr = np.array([1.0, 2.0, 3.0])
>>> input = flow.tensor(arr)
>>> output = flow.std(input, dim=0).numpy()
>>> output
array(1.)

oneflow.sub()

Computes the subtraction of input by other for each element, scalar and broadcast promotation are supported. The formula is:

$out = input - other$

For example:

>>> import numpy as np
>>> import oneflow as flow

# element-wise subtract
>>> input = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> other = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> out = flow.sub(input,other).numpy()
>>> out.shape
(2, 3)

# scalar subtract
>>> input = 5
>>> other = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> out = flow.sub(input,other).numpy()
>>> out.shape
(2, 3)

>>> input = flow.tensor(np.random.randn(1,1), dtype=flow.float32)
>>> other = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> out = flow.sub(input,other).numpy()
>>> out.shape
(2, 3)

oneflow.sum(input, dim=None, keepdim=False)Tensor

Computes the sum of row of elements in a tensor in the given dimension. If the dimension is None, sum of all elements will be caculated.

If keepdim is True, the output tensor is of the same size as input except in the dimension(s) dim where it is of size 1. Otherwise, dim is squeezed oneflow.squeeze(), resulting in the output tensor having 1 (or len(dim)) fewer dimension(s).

Parameters
• input (oneflow.Tensor) – the Input Tensor

• dim (int or tuple of ints, optional) – the dimension to reduce. Default: None

• keepdim (bool, optional) – whether the output tensor has dim retained or not. Default: False

For example:

>>> import oneflow as flow

>>> input = flow.Tensor([[1, 2, 3], [4, 5, 6]])
>>> flow.sum(input)
tensor(21., dtype=oneflow.float32)
>>> flow.sum(input, dim=0)
tensor([5., 7., 9.], dtype=oneflow.float32)
>>> flow.sum(input, dim=1)
tensor([ 6., 15.], dtype=oneflow.float32)

oneflow.swapaxes(input, axis0, axis1)Tensor

This function is equivalent to NumPy’s swapaxes function.

For example:

>>> import oneflow as flow

>>> x = flow.tensor([[[0,1],[2,3]],[[4,5],[6,7]]])
>>> x.shape
oneflow.Size([2, 2, 2])
>>> flow.swapaxes(x, 0, 1).shape
oneflow.Size([2, 2, 2])
>>> flow.swapaxes(x, 0, 2).shape
oneflow.Size([2, 2, 2])

oneflow.swapdims(input, dim0, dim1)Tensor

This function is equivalent to torch’s swapdims function.

For example:

>>> import oneflow as flow

>>> x = flow.tensor([[[0,1],[2,3]],[[4,5],[6,7]]])
>>> x
tensor([[[0, 1],
[2, 3]],

[[4, 5],
[6, 7]]], dtype=oneflow.int64)
>>> flow.swapdims(x, 0, 1)
tensor([[[0, 1],
[4, 5]],

[[2, 3],
[6, 7]]], dtype=oneflow.int64)
>>> flow.swapdims(x, 0, 2)
tensor([[[0, 4],
[2, 6]],

[[1, 5],
[3, 7]]], dtype=oneflow.int64)

oneflow.t()

oneflow.t(input) → Tensor.

Expects input to be <= 2-D tensor and transposes dimensions 0 and 1.

0-D and 1-D tensors are returned as is. When input is a 2-D tensor this is equivalent to transpose(input, 0, 1).

Parameters

input (oneflow.Tensor) – An input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> x = flow.tensor(np.random.randn(), dtype=flow.float32)
>>> flow.t(x).shape
oneflow.Size([])
>>> x = flow.tensor(np.random.randn(3), dtype=flow.float32)
>>> flow.t(x).shape
oneflow.Size([3])
>>> x = flow.tensor(np.random.randn(2,3), dtype=flow.float32)
>>> flow.t(x).shape
oneflow.Size([3, 2])

oneflow.tan()

Returns the tan value of the elements of input.

$\text{out}_{i} = \tan(\text{input}_{i})$
Parameters

input (Tensor) – the input tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> np_arr = np.array([-1/4*np.pi, 0, 1/4*np.pi]).astype(np.float32)
>>> input = flow.tensor(np_arr, dtype=flow.float32)
>>> output = flow.tan(input)
>>> output
tensor([-1.,  0.,  1.], dtype=oneflow.float32)

oneflow.tanh(x: Tensor)Tensor

The equation is:

$out = \frac{e^x-e^{-x}}{e^x+e^{-x}}$

See Tanh for more details.

oneflow.tensor()
Constructs a tensor with data, return a global tensor if placement and sbp are in kwargs,

otherwise return a local tensor.

Parameters

data – Initial data for the tensor. Can be a list, tuple, NumPy ndarray, scalar or tensor.

Keyword Arguments
• dtype (oneflow.dtype, optional) – Default: if None, infers data type from data.

• device (oneflow.device, optional) – the desired device of returned tensor. If placement and sbp is None, uses the current cpu for the default tensor type.

• placement (oneflow.placement, optional) – the desired placement of returned tensor.

• sbp (oneflow.sbp or tuple of oneflow.sbp, optional) – the desired sbp of returned tensor.

• requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False

• pin_memory (bool, optional) – If set, returned tensor would be allocated in the pinned memory. Works only for CPU tensors. Default: False.

Note

The Keyword Argument device is mutually exclusive with placement and sbp.

For example:

>>> import oneflow as flow

>>> x = flow.tensor([1,2,3])
>>> x
tensor([1, 2, 3], dtype=oneflow.int64)

oneflow.tensor_scatter_nd_update(tensor, indices, updates)

This operation creates a new tensor by applying sparse updates to the input tensor. This is similar to an index assignment.

This operator is very similar to scatter_nd(), except that the updates are scattered onto an existing tensor (as opposed to a zero-tensor).

Parameters
• tensor – The tensor will be scattered.

• indices – The indices of update. Its type should be flow.int.

• update – The update Tensor.

For example:

>>> import oneflow as flow
>>> tensor = flow.arange(8)
>>> indices = flow.tensor([[1], [3], [5]])
>>> updates = flow.tensor([-1, -2, -3])
tensor([ 0, -1,  2, -2,  4, -3,  6,  7], dtype=oneflow.int64)

oneflow.tensor_split()

Splits a tensor into multiple sub-tensors, all of which are views of input, along dimension dim according to the indices or number of sections specified by indices_or_sections . The documentation is referenced from: https://pytorch.org/docs/1.10/generated/torch.tensor_split.html.

Parameters
• input (Tensor) – the input tensor.

• indices_or_sections (int or a list) – If indices_or_sections is an integer n , input is split into n sections along dimension dim.If input is divisible by n along dimension dim, each section will be of equal size, input.size (dim) / n. If input is not divisible by n, the sizes of the first int(input.size(dim) % n). sections will have size int(input.size(dim) / n) + 1, and the rest will have size int(input.size(dim) / n). If indices_or_sections is a list or tuple of ints, then input is split along dimension dim at each of the indices in the list, tuple or tensor. For instance, indices_or_sections=[2, 3] and dim=0 would result in the tensors input[:2], input[2:3], and input[3:].If indices_or_sections is a tensor, it must be a zero-dimensional or one-dimensional long tensor on the CPU.

• dim (int) – dimension along which to split the tensor.

Returns

the output TensorTuple.

Return type

oneflow.TensorTuple

For example:

>>> import oneflow as flow

>>> input = flow.rand(3,4,5)
>>> output = flow.tensor_split(input,(2,3),2)
>>> output[0].size()
oneflow.Size([3, 4, 2])
>>> output[1].size()
oneflow.Size([3, 4, 1])
>>> output[2].size()
oneflow.Size([3, 4, 2])

oneflow.tensordot(a, b, dims=Union[int, Tensor, Tuple[List[int], List[int]], List[List[int]]], out=None)Tensor

Compute tensor dot along given dimensions.

Given two tensors a and b, and dims which represent two lists containing dim indices, tensordot traverses the two lists and calculate the tensor dot along every dim pair.

Parameters
• a (oneflow.Tensor) – The input tensor to compute tensordot

• b (oneflow.Tensor) – The input tensor to compute tensordot

• dims (int or list or tuple or oneflow.Tensor) – The dims to calculate tensordot. If it’s an integer or oneflow.Tensor with only one element, the last dims of tensor a and the first dims of tensor b will be calculated. If it’s a list or tuple or oneflow.Tensor with more than one element, it must contain two array-like object, which represent the dims of tensor a and tensor b to be calculated.

• out (oneflow.Tensor) – The tensor to save result (NOT IMPLEMENTED YET)

Returns

The result tensor

Return type

oneflow.Tensor

For example:

>>> import oneflow as flow
>>> a = flow.randn(3, 4, 5)
>>> b = flow.randn(4, 5, 6)
>>> flow.tensordot(a, b, dims=2).shape
oneflow.Size([3, 6])
>>> b = flow.randn(5, 6, 7)
>>> flow.tensordot(a, b, dims=1).shape
oneflow.Size([3, 4, 6, 7])
>>> b = flow.randn(3, 4, 7)
>>> flow.tensordot(a, b, dims=[[0, 1], [0, 1]]).shape
oneflow.Size([5, 7])


Note

Three common use cases are:

• dims = 0 : tensor product $$a \otimes b$$

• dims = 1 : tensor dot product $$a \cdot b$$

• dims = 2 : (default) tensor double contraction $$a : b$$

The part of documentation is referenced from https://numpy.org/doc/stable/reference/generated/numpy.tensordot.html.

Note

The operation is equivalent to the series of operations:

• Permute the dimensions of the tensor A that require tensordot to the end

• Permute the dimensions of the tensor B that require tensordot to the start

• Reshape the permuted tensor A into a 2-dimensional tensor, where the size of the 0th dimension is the product of the dimensions that do not require dot product, and the size of the 1st dimension is the product of the dimensions that require dot product

• Reshape the permuted tensor B into a 2-dimensional tensor, where the size of the 0th dimension is the product of the dimensions that require dot product, and the size of the 1st dimension is the product of the dimensions that do not require dot product

• Calculate the matrix multiplication of reshaped tensor A and reshaped tensor B

• Reshape the result of matrix multiplication, the target shape is the concatenation of the dimensions that do not require tensordot of tensor A and B

This series of operations can be equivalently represented by the following code:

>>> import oneflow as flow
>>> a = flow.randn(2, 4, 3)
>>> b = flow.randn(3, 4, 2)
>>> dims = [[0, 2], [2, 0]]
>>> permuted_a = a.permute(1, 0, 2) # 0, 2 are the dimensions requiring tensordot and are placed in the end in permuting
>>> permuted_b = b.permute(2, 0, 1) # 2, 0 are the dimensions requiring tensordot and are placed at the beginning in permuting
>>> reshaped_a = permuted_a.reshape(4, 2 * 3) # 4 is the dimensions of a that do not require tensordot
>>> reshaped_b = permuted_b.reshape(2 * 3, 4) # 4 is the dimensions of a that do not require tensordot
>>> matmul_result = flow.matmul(reshaped_a, reshaped_b)
>>> result = matmul_result.reshape(4, 4) # 4, 4 are the concatentation of dimensions that do not require tensordot of a and b
>>> flow.all(result == flow.tensordot(a, b, dims))
tensor(True, dtype=oneflow.bool)

oneflow.tile(input, dims)Tensor

The interface is consistent with PyTorch. The documentation is referenced from: https://pytorch.org/docs/1.10/generated/torch.tile.html.

Constructs a tensor by repeating the elements of input. The dims argument specifies the number of repetitions in each dimension.

If dims specifies fewer dimensions than input has, then ones are prepended to dims until all dimensions are specified. For example, if input has shape (8, 6, 4, 2) and dims is (2, 2), then dims is treated as (1, 1, 2, 2).

Analogously, if input has fewer dimensions than dims specifies, then input is treated as if it were unsqueezed at dimension zero until it has as many dimensions as dims specifies. For example, if input has shape (4, 2) and dims is (3, 3, 2, 2), then input is treated as if it had the shape (1, 1, 4, 2).

Note

This function is similar to NumPy’s tile function.

Parameters
• input (oneflow.Tensor) – the tensor whose elements to repeat.

• dims (tuple) – the number of repetitions per dimension.

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> np_arr = np.random.randn(5, 3, 6, 9).astype(np.float32)
>>> input = flow.Tensor(np_arr)
>>> out = input.tile(2,1,2,1)
>>> out.shape
oneflow.Size([10, 3, 12, 9])
>>> x = np.random.randn(5, 2, 1)
>>> input = flow.Tensor(x)
>>> out = input.tile(3,4)
>>> out.shape
oneflow.Size([5, 6, 4])

oneflow.topk(input, k, dim: Optional[int] = None, largest: bool = True, sorted: bool = True)

Finds the values and indices of the k largest entries at specified axis.

Parameters
• input (oneflow.Tensor) – Input Tensor

• k (int) – the k in “top-k”

• dim (int, optional) – the dimension to sort along. Defaults to the last dim (-1)

• largest (bool, optional) – controls whether to return largest or smallest elements

• sorted (bool, optional) – controls whether to return the elements in sorted order (Only Support True Now!)

Returns

A tuple of (values, indices), where the indices are the indices of the elements in the original input tensor.

Return type

Tuple(oneflow.Tensor, oneflow.Tensor(dtype=int32))

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> x = np.array([[1, 3, 8, 7, 2], [1, 9, 4, 3, 2]], dtype=np.float32)
>>> (values, indices) = flow.topk(flow.Tensor(x), k=3, dim=1)
>>> values
tensor([[8., 7., 3.],
[9., 4., 3.]], dtype=oneflow.float32)
>>> indices
tensor([[2, 3, 1],
[1, 2, 3]], dtype=oneflow.int64)
>>> values.shape
oneflow.Size([2, 3])
>>> indices.shape
oneflow.Size([2, 3])
>>> (values, indices) = flow.topk(flow.Tensor(x), k=2, dim=1, largest=False)
>>> values
tensor([[1., 2.],
[1., 2.]], dtype=oneflow.float32)
>>> indices
tensor([[0, 4],
[0, 4]], dtype=oneflow.int64)
>>> values.shape
oneflow.Size([2, 2])
>>> indices.shape
oneflow.Size([2, 2])

oneflow.transpose()

Returns a tensor that is a transposed version of input. The given dimensions dim0 and dim1 are swapped.

The resulting out tensor shares its underlying storage with the input tensor, so changing the content of one would change the content of the other.

Parameters
• input (oneflow.Tensor) – the input tensor.

• dim0 (int) – the first dimension to be transposed.

• dim1 (int) – the second dimension to be transposed.

Returns

A transposed tensor.

Return type

Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow
>>> input = flow.tensor(np.random.randn(2, 6, 5, 3), dtype=flow.float32)
>>> out = flow.transpose(input, 0, 1).shape
>>> out
oneflow.Size([6, 2, 5, 3])

oneflow.tril()

Returns the lower triangular part of a matrix (2-D tensor) or batch of matrices input along the specified diagonal, the other elements of the result tensor out are set to 0.

Note

• if diagonal = 0, the diagonal of the returned tensor will be the main diagonal,

• if diagonal > 0, the diagonal of the returned tensor will be above the main diagonal,

• if diagonal < 0, the diagonal of the returned tensor will be below the main diagonal.

Parameters
• input (Tensor) – the input tensor.

• diagonal (int, optional) – the diagonal to specify.

For example:

>>> import oneflow as flow
>>> import numpy as np

>>> x = flow.tensor(np.ones(shape=(3, 3)).astype(np.float32))
>>> flow.tril(x)
tensor([[1., 0., 0.],
[1., 1., 0.],
[1., 1., 1.]], dtype=oneflow.float32)

oneflow.unbind()

This function is equivalent to PyTorch’s unbind function. Removes a tensor dimension.

Returns a tuple of all slices along a given dimension, already without it.

Parameters
• x (Tensor) – the tensor to unbind

• dim (int) – dimension to remove

For example:

>>> import oneflow as flow

>>> x = flow.tensor(range(12)).reshape([3,4])
>>> flow.unbind(x)
(tensor([0, 1, 2, 3], dtype=oneflow.int64), tensor([4, 5, 6, 7], dtype=oneflow.int64), tensor([ 8,  9, 10, 11], dtype=oneflow.int64))
>>> flow.unbind(x, 1)
(tensor([0, 4, 8], dtype=oneflow.int64), tensor([1, 5, 9], dtype=oneflow.int64), tensor([ 2,  6, 10], dtype=oneflow.int64), tensor([ 3,  7, 11], dtype=oneflow.int64))

oneflow.unsqueeze(input, dim)Tensor

Returns a new tensor with a dimension of size one inserted at the specified position.

The returned tensor shares the same underlying data with this tensor.

A dim value within the range [-input.ndimension() - 1, input.ndimension() + 1) can be used. Negative dim will correspond to unsqueeze() applied at dim = dim + input.ndimension() + 1.

Parameters
• input (Tensor) – the input tensor.

• dim (int) – the index at which to insert the singleton dimension

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> x = flow.randn(2, 3, 4)
>>> y = x.unsqueeze(2)
>>> y.shape
oneflow.Size([2, 3, 1, 4])

oneflow.var()

Returns the variance of each row of the input tensor in the given dimension dim.

If keepdim is True, the output tensor is of the same size as input except in the dimension(s) dim where it is of size 1. Otherwise, dim is squeezed (see flow.squeeze()), resulting in the output tensor having 1 (or len(dim)) fewer dimension(s).

Parameters
• input (Tensor) – the input tensor.

• dim (int or tuple of ints) – the dimension or dimensions to reduce. Defaults to None.

• unbiased (bool, optional) – whether to use Bessel’s correction ($$\delta N = 1$$). Defaults to True.

• keepdim (bool, optional) – whether the output tensor has dim retained or not. Defaults to False.

Returns

The result of variance on the specified axis of input Tensor

Return type

Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow

>>> input = flow.tensor(np.random.randn(2, 3, 4, 5))
>>> output = flow.var(input, 1, True)

oneflow.vsplit()

Splits input, a tensor with two or more dimensions, into multiple tensors vertically according to indices_or_sections. Each split is a view of input. This is equivalent to calling oneflow.tensor_split(input, indices_or_sections, dim=0) (the split dimension is 0), except that if indices_or_sections is an integer it must evenly divide the split dimension or a runtime error will be thrown. The documentation is referenced from: https://pytorch.org/docs/1.10/generated/torch.vsplit.html.

Parameters
• input (Tensor) – the input tensor.

• indices_or_sections (int or a list) – If indices_or_sections is an integer n , input is split into n sections along dimension dim.If input is divisible by n along dimension dim, each section will be of equal size, input.size (dim) / n. If input is not divisible by n, the sizes of the first int(input.size(dim) % n). sections will have size int(input.size(dim) / n) + 1, and the rest will have size int(input.size(dim) / n). If indices_or_sections is a list or tuple of ints, then input is split along dimension dim at each of the indices in the list, tuple or tensor. For instance, indices_or_sections=[2, 3] and dim=0 would result in the tensors input[:2], input[2:3], and input[3:].If indices_or_sections is a tensor, it must be a zero-dimensional or one-dimensional long tensor on the CPU.

Returns

the output TensorTuple.

Return type

oneflow.TensorTuple

For example:

>>> import oneflow as flow

>>> input = flow.rand(4, 4, 5, 6)
>>> output = flow.vsplit(input, (1, 3))
>>> output[0].size()
oneflow.Size([1, 4, 5, 6])
>>> output[1].size()
oneflow.Size([2, 4, 5, 6])
>>> output[2].size()
oneflow.Size([1, 4, 5, 6])

oneflow.where(condition, x=None, y=None)

Return a tensor of elements selected from either x or y, depending on condition. If the element in condition is larger than 0,

it will take the x element, else it will take the y element

Note

If x is None and y is None, flow.where(condition) is identical to flow.nonzero(condition, as_tuple=True).

The tensors condition, x, y must be broadcastable.

Parameters
• condition (IntTensor) – When 1 (nonzero), yield x, otherwise yield y

• x (Tensor or Scalar) – value (if :attr:x is a scalar) or values selected at indices where condition is True

• y (Tensor or Scalar) – value (if :attr:x is a scalar) or values selected at indices where condition is False

Returns

A tensor of shape equal to the broadcasted shape of condition, x, y

Return type

Tensor

For example:

>>> import numpy as np
>>> import oneflow as flow
>>> x = flow.tensor(
...    np.array([[-0.4620, 0.3139], [0.3898, -0.7197], [0.0478, -0.1657]]),
...    dtype=flow.float32,
... )
>>> y = flow.tensor(np.ones(shape=(3, 2)), dtype=flow.float32)
>>> condition = flow.tensor(np.array([[0, 1], [1, 0], [1, 0]]), dtype=flow.int32)
>>> out = condition.where(x, y)
>>> out
tensor([[1.0000, 0.3139],
...
[0.0478, 1.0000]], dtype=oneflow.float32)

oneflow.zeros(*size: Union[int, Tuple[int, ], oneflow.Size, List[int]], dtype: Optional[oneflow._oneflow_internal.dtype] = None, device: Optional[Union[str, oneflow._oneflow_internal.device]] = None, placement: Optional[oneflow._oneflow_internal.placement] = None, sbp: Optional[oneflow._oneflow_internal.sbp.sbp] = None, requires_grad: bool = False)

Returns a tensor filled with the scalar value 0, with the shape defined by the variable argument size.

Parameters
• size (an integer or tuple of integer values) – a variable number of arguments or a collection like a list or tuple.

• dtype (flow.dtype, optional) – the desired data type of returned tensor.

• device (flow.device, optional) – the desired device of returned tensor. Default: if None, uses the current device for the default tensor type

• placement (flow.placement, optional) – the desired placement of returned global tensor. Default: if None, the returned tensor is local one using the argument device.

• sbp (flow.sbp.sbp or tuple of flow.sbp.sbp, optional) – the desired sbp descriptor of returned global tensor. Default: if None, the returned tensor is local one using the argument device.

• requires_grad (bool, optional) – If autograd should record operations on the returned tensor. Default: False.

For example:

>>> import oneflow as flow
>>> y = flow.zeros(5)
>>> y
tensor([0., 0., 0., 0., 0.], dtype=oneflow.float32)
>>> y = flow.zeros(2,3)
>>> y
tensor([[0., 0., 0.],
[0., 0., 0.]], dtype=oneflow.float32)

oneflow.zeros_like(input) is equivalent to flow.zeros(input.shape, dtype=input.dtype)
Parameters

other (Tensor) – The size of input will determine size of the output tensor.

For example:

>>> import oneflow as flow
>>> import numpy as np
>>> x = flow.tensor(np.random.rand(5), dtype=flow.float32)
>>> y = flow.zeros_like(x)
>>> y
tensor([0., 0., 0., 0., 0.], dtype=oneflow.float32)

oneflow.relu()

Applies the rectified linear unit function element-wise. See ReLU for more details.

Parameters

inplace – If set to True, will do this operation in-place. Default: False

For examples:

>>> import oneflow as flow
>>> import numpy as np

>>> ndarr = np.asarray([1, -2, 3])
>>> input = flow.Tensor(ndarr)
>>> output = flow.relu(input)
>>> output
tensor([1., 0., 3.], dtype=oneflow.float32)

oneflow.set_num_threads()

Sets the number of threads used for intraop parallelism on CPU.

Warning

To ensure that the correct number of threads is used, set_num_threads must be called before running eager, eager globe or ddp.