oneflow.Tensor

A oneflow.Tensor is a multi-dimensional matrix containing elements of a single data type.

Data types

OneFlow defines 8 Tensor types with CPU and GPU variants which are as follows:

Data type	dtype	CPU tensor	GPU tensor
Boolean	oneflow.bool	oneflow.BoolTensor	oneflow.cuda.BoolTensor
8-bit integer (unsigned)	oneflow.uint8	oneflow.ByteTensor	oneflow.cuda.ByteTensor
8-bit integer (signed)	oneflow.int8	oneflow.CharTensor	oneflow.cuda.CharTensor
64-bit floating point	oneflow.float64 or oneflow.double	oneflow.DoubleTensor	oneflow.cuda.DoubleTensor
32-bit floating point	oneflow.float32 or oneflow.float	oneflow.FloatTensor	oneflow.cuda.FloatTensor
16-bit floating point	oneflow.float16 or oneflow.half	oneflow.HalfTensor	oneflow.cuda.HalfTensor
32-bit integer (signed)	oneflow.int32 or oneflow.int	oneflow.IntTensor	oneflow.cuda.IntTensor
64-bit integer (signed)	oneflow.int64 or oneflow.long	oneflow.LongTensor	oneflow.cuda.LongTensor

Initializing and basic operations

A tensor can be constructed from a Python list or sequence using the oneflow.tensor() constructor:

>>> import oneflow
>>> import numpy as np
>>> oneflow.tensor([[1., -1.], [1., -1.]])
tensor([[ 1., -1.],
        [ 1., -1.]], dtype=oneflow.float32)
>>> oneflow.tensor(np.array([[1, 2, 3], [4, 5, 6]]))
tensor([[ 1, 2, 3],
        [ 4, 5, 6]], dtype=oneflow.int64)

Warning

oneflow.tensor() always copies data. If you have a Tensor data and just want to change its requires_grad flag, use requires_grad_() or detach() to avoid a copy. If you have a numpy array and want to avoid a copy, use oneflow.as_tensor().

>>> import oneflow
>>> oneflow.zeros([2, 4], dtype=oneflow.int32)
tensor([[ 0, 0, 0, 0],
        [ 0, 0, 0, 0]], dtype=oneflow.int32)
>>> cuda0 = oneflow.device('cuda:0')
>>> oneflow.ones([2, 4], dtype=oneflow.float64, device=cuda0)
tensor([[ 1., 1., 1., 1.],
        [ 1., 1., 1., 1.]], device='cuda:0', dtype=oneflow.float64)

For more information about building tensors, see Creation Ops

The contents of a tensor can be accessed and modified using Python’s indexing and slicing notation:

>>> import oneflow
>>> x = oneflow.tensor([[1, 2, 3], [4, 5, 6]])
>>> print(x[1][2])
tensor(6, dtype=oneflow.int64)
>>> x[0][1] = 8
>>> print(x)
tensor([[1, 8, 3],
        [4, 5, 6]], dtype=oneflow.int64)

Use oneflow.Tensor.item() to get a Python number from a tensor containing a single value:

>>> import oneflow
>>> x = oneflow.tensor([[1]])
>>> x
tensor([[1]], dtype=oneflow.int64)
>>> x.item()
1
>>> x = oneflow.tensor(2.5)
>>> x
tensor(2.5000, dtype=oneflow.float32)
>>> x.item()
2.5

For more information about indexing, see Indexing, Slicing, Joining, Mutating Ops

A tensor can be created with requires_grad=True so that oneflow.autograd records operations on them for automatic differentiation.

>>> import oneflow
>>> x = oneflow.tensor([[1., -1.], [1., 1.]], requires_grad=True)
>>> out = x.pow(2).sum()
>>> out.backward()
>>> x.grad
tensor([[ 2., -2.],
        [ 2.,  2.]], dtype=oneflow.float32)

Note

For more information on the oneflow.dtype, oneflow.device, and oneflow.layout attributes of a oneflow.Tensor, see Tensor Attributes.

Note

Methods which mutate a tensor are marked with an underscore suffix. For example, oneflow.FloatTensor.add_() computes the absolute value in-place and returns the modified tensor, while oneflow.FloatTensor.add() computes the result in a new tensor.

Note

To change an existing tensor’s oneflow.device and/or oneflow.dtype, consider using to() method of Tensor object.

Warning

Current implementation of oneflow.Tensor introduces memory overhead, thus it might lead to unexpectedly high memory usage in the applications with many tiny tensors. If this is your case, consider using one large structure.

Tensor class reference

class oneflow.Tensor

There are a few main ways to create a tensor, depending on your use case.

• To create a tensor with pre-existing data, use oneflow.tensor().

• To create a tensor with specific size, use oneflow.* tensor creation ops (see Creation Ops).

• To create a tensor with the same size (and similar types) as another tensor, use oneflow.*_like tensor creation ops (see Creation Ops).

Tensor.new_empty	Returns a Tensor of size size filled with uninitialized data.
Tensor.new_ones	See oneflow.new_ones()
Tensor.new_zeros	Returns a Tensor of size size filled with 0.
Tensor.new_full	Returns a Tensor of size size filled with fill_value.
Tensor.new_tensor
Tensor.is_cuda	Is True if the Tensor is stored on the GPU, False otherwise.
Tensor.is_global	Return whether this Tensor is a global tensor.
Tensor.device	Is the oneflow.device where this Tensor is, which is invalid for global tensor.
Tensor.grad	Return the gradient calculated by autograd functions.
Tensor.ndim	See oneflow.Tensor.dim()
Tensor.abs	See oneflow.abs()
Tensor.acos	See oneflow.acos()
Tensor.acosh	See oneflow.acosh()
Tensor.add	See oneflow.add()
Tensor.add_	In-place version of oneflow.Tensor.add().
Tensor.addcdiv	See oneflow.addcdiv()
Tensor.addcdiv_	In-place version of oneflow.Tensor.addcdiv()
Tensor.addcmul	See oneflow.addcmul()
Tensor.addcmul_	In-place version of oneflow.Tensor.addcmul().
Tensor.addmm	See oneflow.addmm()
Tensor.all	See oneflow.all()
Tensor.amin	See oneflow.amin()
Tensor.amax	See oneflow.amax()
Tensor.any	See oneflow.any()
Tensor.arccos	See oneflow.arccos()
Tensor.arccosh	See oneflow.arccosh()
Tensor.arcsin	See oneflow.arcsin()
Tensor.arcsinh	See oneflow.arcsinh()
Tensor.arctan	See oneflow.arctan()
Tensor.arctanh	See oneflow.arctanh()
Tensor.argmax	See oneflow.argmax()
Tensor.argmin	See oneflow.argmin()
Tensor.argsort	See oneflow.argsort()
Tensor.argwhere	See oneflow.argwhere()
Tensor.asin	See oneflow.asin()
Tensor.asinh	See oneflow.asinh()
Tensor.atan	See oneflow.atan()
Tensor.atan2	See oneflow.atan2()
Tensor.atanh	See oneflow.atanh()
Tensor.backward	Computes the gradient of current tensor w.r.t. graph leaves.
Tensor.bmm	See oneflow.bmm()
Tensor.bool	Tensor.bool() is equivalent to Tensor.to(oneflow.bool).
Tensor.byte	self.byte() is equivalent to self.to(oneflow.uint8).
Tensor.cast	See oneflow.cast()
Tensor.ceil	See oneflow.ceil()
Tensor.ceil_	See oneflow.ceil_()
Tensor.chunk	See oneflow.chunk()
Tensor.clamp	See oneflow.clamp().
Tensor.clamp_	Inplace version of oneflow.Tensor.clamp().
Tensor.clip	Alias for oneflow.Tensor.clamp().
Tensor.clip_	Alias for oneflow.Tensor.clamp_().
Tensor.clone	See oneflow.clone()
Tensor.contiguous
Tensor.copy_	Copies the elements from src into self tensor and returns self.
Tensor.cos	See oneflow.cos()
Tensor.cosh	See oneflow.cosh()
Tensor.cpu	Returns a copy of this object in CPU memory.
Tensor.cuda	Returns a copy of this object in CUDA memory.
Tensor.cumprod	See oneflow.cumprod()
Tensor.cumsum	See oneflow.cumsum()
Tensor.data
Tensor.dot	See oneflow.dot()
Tensor.detach
Tensor.placement	Is the oneflow.placement where this Tensor is, which is invalid for local tensor.
Tensor.sbp	Is the oneflow.sbp representing that how the data of the global tensor is distributed, which is invalid for local tensor.
Tensor.diag	See oneflow.diag()
Tensor.diagonal	See oneflow.diagonal()
Tensor.dim	Tensor.dim() → int
Tensor.div	See oneflow.div()
Tensor.div_	In-place version of oneflow.Tensor.div().
Tensor.double	Tensor.double() is equivalent to Tensor.to(flow.float64).
Tensor.dtype
Tensor.digamma	See oneflow.digamma()
Tensor.element_size	Tensor.element_size() → int
Tensor.eq	See oneflow.eq()
Tensor.equal	See oneflow.equal()
Tensor.erf	See oneflow.erf()
Tensor.erfc	See oneflow.erfc()
Tensor.erfinv	See oneflow.erfinv()
Tensor.erfinv_	Inplace version of oneflow.erfinv()
Tensor.exp	See oneflow.exp()
Tensor.exp2	See oneflow.exp2()
Tensor.expand	See oneflow.expand()
Tensor.expand_as	Expand this tensor to the same size as other.
Tensor.expm1	See oneflow.expm1()
Tensor.fill_	Tensor.fill_(value) → Tensor
Tensor.flatten	See oneflow.flatten()
Tensor.flip	See oneflow.flip()
Tensor.float	Tensor.float() is equivalent to Tensor.to(flow.float32).
Tensor.floor	See oneflow.floor()
Tensor.floor_	See oneflow.floor_()
Tensor.floor_divide
Tensor.fmod	See oneflow.fmod()
Tensor.gather	See oneflow.gather()
Tensor.ge	See oneflow.ge()
Tensor.get_device	For CUDA tensors, this function returns the device ordinal of the GPU on which the tensor resides.
Tensor.grad_fn	Return the function that created this tensor if it’s requires_grad is True.
Tensor.gt	See oneflow.gt()
Tensor.gt_	In-place version of oneflow.Tensor.gt().
Tensor.half	self.half() is equivalent to self.to(dtype=oneflow.float16).
Tensor.in_top_k	See oneflow.in_top_k()
Tensor.index_select	See oneflow.index_select()
Tensor.index_add
Tensor.index_add_	The interface is consistent with PyTorch.
Tensor.int	Tensor.int() is equivalent to Tensor.to(flow.int32).
Tensor.is_contiguous	Returns True if self tensor is contiguous in memory.
Tensor.is_floating_point	See oneflow.is_floating_point()
Tensor.is_lazy	Return whether this Tensor is a lazy tensor.
Tensor.is_leaf	All Tensors that have requires_grad which is False will be leaf Tensors by convention.
Tensor.isinf	See oneflow.isinf()
Tensor.isnan	See oneflow.isnan()
Tensor.item	Returns the value of this tensor as a standard Python number.
Tensor.le	See oneflow.le()
Tensor.lerp	See oneflow.lerp()
Tensor.lerp_	See oneflow.lerp_()
Tensor.log	See oneflow.log()
Tensor.log1p	See oneflow.log1p()
Tensor.log2	See oneflow.log2()
Tensor.log10	See oneflow.log10()
Tensor.logical_and	See oneflow.logical_and()
Tensor.logical_or	See oneflow.logical_or()
Tensor.logical_not	See oneflow.logical_not()
Tensor.logical_xor	See oneflow.logical_xor()
Tensor.long	Tensor.long() is equivalent to Tensor.to(flow.int64).
Tensor.lt	See oneflow.lt()
Tensor.masked_fill	See oneflow.masked_fill()
Tensor.masked_fill_	In-place version of oneflow.Tensor.masked_fill().
Tensor.masked_select	See oneflow.masked_select()
Tensor.matmul	See oneflow.matmul()
Tensor.mm	See oneflow.mm()
Tensor.mv	See oneflow.mv()
Tensor.max	See oneflow.max()
Tensor.maximum	See oneflow.maximum()
Tensor.median	See oneflow.median()
Tensor.mean	See oneflow.mean()
Tensor.min	See oneflow.min()
Tensor.minimum	See oneflow.minimum()
Tensor.mish	See oneflow.mish()
Tensor.mode	See oneflow.mode()
Tensor.mul	See oneflow.mul()
Tensor.mul_	In-place version of oneflow.Tensor.mul().
Tensor.frac	See oneflow.frac().
Tensor.frac_	In-place version of oneflow.Tensor.frac().
Tensor.nansum	See oneflow.nansum()
Tensor.narrow	See oneflow.narrow()
Tensor.ndimension
Tensor.ne	See oneflow.ne()
Tensor.neg	See oneflow.neg()
Tensor.negative	See oneflow.negative()
Tensor.nelement	Tensor.nelement() → int
Tensor.nonzero	See oneflow.nonzero()
Tensor.norm	See oneflow.norm()
Tensor.normal_	Fills self tensor with elements samples from the normal distribution parameterized by mean and std.
Tensor.numel	See oneflow.numel()
Tensor.numpy	Tensor.numpy() → numpy.ndarray
Tensor.offload	Transfer tensor data from GPU memory back to host (CPU) memory.
Tensor.load	Load tensor data stored on the host (CPU) back to GPU memory.
Tensor.is_offloaded	Determine whether the tensor has been moved to CPU memory and the CUDA device memory has been released.
Tensor.permute	See oneflow.permute()
Tensor.pow	See oneflow.pow()
Tensor.prod	See oneflow.prod()
Tensor.quantile	See oneflow.quantile()
Tensor.reciprocal	See oneflow.reciprocal()
Tensor.register_hook	Registers a backward hook.
Tensor.relu	See oneflow.relu()
Tensor.repeat	See oneflow.repeat()
Tensor.repeat_interleave	See oneflow.repeat_interleave()
Tensor.requires_grad	Is True if gradient need to be computed for this Tensor, False otherwise.
Tensor.requires_grad_	Sets this tensor’s requires_grad attribute in-place.
Tensor.reshape	See oneflow.reshape()
Tensor.reshape_as	Returns this tensor as the same shape as other.
Tensor.retain_grad	Enables this Tensor to have their grad populated during backward().
Tensor.roll	See oneflow.roll()
Tensor.round	See oneflow.round()
Tensor.round_	See oneflow.round_()
Tensor.rsqrt	See oneflow.rsqrt()
Tensor.selu	See oneflow.selu()
Tensor.shape
Tensor.sigmoid	See oneflow.sigmoid()
Tensor.sign	See oneflow.sign()
Tensor.silu	See oneflow.silu()
Tensor.sin	See oneflow.sin()
Tensor.sin_	See oneflow.sin_()
Tensor.sinh	See oneflow.sinh()
Tensor.size	Returns the size of the self tensor.
Tensor.softmax	See oneflow.softmax()
Tensor.softplus	See oneflow.softplus()
Tensor.softsign	See oneflow.softsign()
Tensor.sort	See oneflow.sort()
Tensor.split	See oneflow.split()
Tensor.sqrt	See oneflow.sqrt()
Tensor.square	See oneflow.square()
Tensor.squeeze	See oneflow.squeeze()
Tensor.squeeze_	In-place version of oneflow.Tensor.squeeze()
Tensor.std	See oneflow.std()
Tensor.storage_offset	Returns self tensor’s offset in the underlying storage in terms of number of storage elements (not bytes).
Tensor.stride
Tensor.logsumexp	See oneflow.logsumexp()
Tensor.sum	See oneflow.sum()
Tensor.swapaxes	See oneflow.swapaxes()
Tensor.swapdims	See oneflow.swapdims()
Tensor.sub	See oneflow.sub()
Tensor.sub_	In-place version of oneflow.Tensor.sub().
Tensor.tan	See oneflow.tan()
Tensor.tanh	See oneflow.tanh()
Tensor.tile	See oneflow.tile()
Tensor.to	Performs Tensor dtype and/or device conversion.
Tensor.local_to_global	Creates a global tensor from a local tensor.
Tensor.global_to_global	Performs Tensor placement and/or sbp conversion.
Tensor.to_global	Creates a global tensor if this tensor is a local tensor, otherwise performs Tensor placement and/or sbp conversion.
Tensor.to_local	Returns the local component of this global tensor in the current rank.
Tensor.to_consistent	This interface is no longer available, please use oneflow.Tensor.to_global() instead.
Tensor.tolist	Returns the tensor as a (nested) list.
Tensor.topk	See oneflow.topk()
Tensor.transpose	See oneflow.transpose()
Tensor.tril	See oneflow.tril()
Tensor.triu	See oneflow.triu()
Tensor.trunc	See oneflow.trunc()
Tensor.type_as	Returns this tensor cast to the type of the given tensor.
Tensor.type	Returns the type if dtype is not provided, else casts this object to the specified type.
Tensor.t	See oneflow.t()
Tensor.T	Is this Tensor with its dimensions reversed.
Tensor.unbind	See oneflow.unbind()
Tensor.unfold	Returns a view of the original tensor which contains all slices of size size from self tensor in the dimension dimension.
Tensor.uniform_	Tensor.uniform_(from=0, to=1) → Tensor
Tensor.unsqueeze	See oneflow.unsqueeze()
Tensor.unsqueeze_	In-place version of oneflow.Tensor.unsqueeze()
Tensor.as_strided	See oneflow.as_strided()
Tensor.as_strided_	In-place version of oneflow.Tensor.as_strided()
Tensor.var	See oneflow.var()
Tensor.view	Returns a new tensor with the same data as the self tensor but of a different shape.
Tensor.view_as	Expand this tensor to the same size as other.
Tensor.where	See oneflow.where()
Tensor.zero_	Fills self tensor with zeros.
Tensor.nms	See oneflow.nms()
Tensor.pin_memory	Copies the tensor to pinned memory, if it’s not already pinned.
Tensor.is_pinned	Returns true if this tensor resides in pinned memory.
Tensor.inverse	See oneflow.linalg.inv()
Tensor.cross	See oneflow.cross()
Tensor.scatter	See oneflow.scatter()
Tensor.scatter_	Inplace version of oneflow.Tensor.scatter()
Tensor.scatter_add	See oneflow.scatter_add()
Tensor.scatter_add_	Inplace version of oneflow.Tensor.scatter_add()
Tensor.bernoulli	See oneflow.bernoulli()
Tensor.bernoulli_	The inplace version of oneflow.Tensor.bernoulli_().
Tensor.bincount	See oneflow.bincount()
Tensor.isclose
Tensor.allclose
Tensor.broadcast_to	See oneflow.broadcast_to()
Tensor.unique	See oneflow.unique()
Tensor.bitwise_and	See oneflow.bitwise_and()
Tensor.bitwise_or	See oneflow.bitwise_or()
Tensor.bitwise_xor	See oneflow.bitwise_xor()
Tensor.baddbmm	See oneflow.baddbmm()