torchrelay

Overview

Modules
`extra`	Extended variants of functions from the PyTorch functional API.
`itertools`	Same as the itertools python package, but in PyTorch.
`multivers`	Backward-compatible PyTorch API.
`jit`	TorchScript utilities.
Backends
`to`	Move/convert to a common dtype or device.
`to_max_backend`	Move to a common dtype or device.
`to_max_device`	Move to a common device.
`to_max_dtype`	Move to a common dtype.
`get_backend`	Return the backend (dtype and device) of a tensor.
`max_backend`	Get the (max) dtype and device.
`max_device`	Find a common device for all inputs.
`max_dtype`	Find the maximum data type from a series of inputs.
`as_torch_dtype`	Convert a numpy data type (or a data type name) to a torch data dtype.
`as_numpy_dtype`	Convert a torch data type (or a data type name) to a numpy data dtype.
Indexing
`moveelem`	Move elements in a tensor.
`slice_tensor_along`	Index a tensor along one dimensions.
`slice_tensor`	Index a tensor along one or several dimensions.
Shapes
`movedims`	Moves the position of one or more dimensions.
`movedims_front2back`	Move the first N dimensions to the back.
`movedims_back2front`	Move the last N dimensions to the front.
`shiftdim`	Shift the dimensions of x by n.

torch_version

torch_version(mode, version)

Check torch version

Parameters:

Name	Type	Description	Default
`mode`	`('<', '<=', '>', '>=')`		`'<'`
`version`	`tuple[int]`		required

Returns:

Type	Description
`True if "torch.version <mode> version"`

to

to(*args, dtype=None, device=None)

Move/convert to a common dtype or device.

Parameters:

Name	Type	Description	Default
`*args`	`tensor_like`	Input tensors or tensor-like objects	`()`
`dtype`	`str or dtype`	Target data type	`None`
`device`	`str or device`	Target device	`None`

Returns:

Name	Type	Description
`*args`	`tensor_like`	Converted tensors

to_max_backend

to_max_backend(*args, force_float=False, dtype=None, device=None)

Move to a common dtype and device.

See max_dtype and max_device.

Parameters:

Name	Type	Description	Default
`*args`	`tensor_like`		`()`
`force_float`	`bool`		`False`

Returns:

Name	Type	Description
`*args_to`	`tensor`

to_max_device

to_max_device(*args)

Move to a common device.

See max_device.

Parameters:

Name	Type	Description	Default
`*args`	`tensor_like`		`()`

Returns:

Name	Type	Description
`*args_to`	`tensor`

to_max_dtype

to_max_dtype(*args)

Move to a common data type.

See max_dtype.

Parameters:

Name	Type	Description	Default
`*args`	`tensor_like`		`()`

Returns:

Name	Type	Description
`*args_to`	`tensor`

get_backend

get_backend(x)

Return the backend (dtype and device) of a tensor

Parameters:

Name	Type	Description	Default
`x`	`tensor`		required

Returns:

Type	Description
`dict with keys 'dtype' and 'device'`

max_backend

max_backend(*args, dtype=None, device=None)

Get the (max) dtype and device.

Parameters:

Name	Type	Description	Default
`args`	`tensors`		`()`

Returns:

Type	Description
`dict with keys 'dtype' and 'device'`

max_device

max_device(*args)

Find a common device for all inputs.

If at least one input object is on a CUDA device:

if all cuda object are on the same cuda device, return it
if some objects are on different cuda devices, return device('cuda') without an index.

Else, return device('cpu') or None.

Parameters:

Name	Type	Description	Default
`*args`	`tensor_like or device_like`		`()`

Returns:

Name	Type	Description
`device`	`device`

max_dtype

max_dtype(*args, force_float=False)

Find the maximum data type from a series of inputs.

The returned dtype is the best one to use for upcasting the objects.

Tensors and arrays have priority python objects.
Tensors and arrays with non-null dimensionality have priority over scalars.
If any of the torch/numpy objects have a floating point type a floating point type is returned.
If any of the objects is complex, a complex type is returned.
If all torch/numpy objects have an integer type and there is an integer type that avoids overflowing, it is returned.
If no integer type that ensures underflowing exists, the default floating point data type is returned.
If force_float is True, a floating point data type is returned even if all input objects have an integer data type.

Parameters:

Name	Type	Description	Default
`*args`	`tensor_like or type_like`		`()`
`force_float`	`bool`		`False`

Returns:

Name	Type	Description
`dtype`	`dtype`

as_torch_dtype

as_torch_dtype(dtype, byteswap=True, upcast=False)

Convert a numpy data type (or a data type name) to a torch data dtype.

Warning

Builtin Python data types int and float are mapped to torch.int32 and torch.float32, to match torch.as_tensor's behavior. It differs from as_numpy_dtype, which maps these types to np.int64 and np.float64.

Parameters:

Name	Type	Description	Default
`dtype`	`str or dtype or dtype`	Input data type	required
`byteswap`	`bool`	If the data type is not implemented in PyTorch but its byteswapped version is, return the byteswapped version. If `False`, raise a `TypeError`.	`True`
`upcast`	`bool`	If the data type is not implemented in PyTorch, but its values can be represented by a larger data type that exists in PyTorch, return the larger data type. If `False`, raise a `TypeError`.	`False`

Returns:

Name	Type	Description
`dtype`	`dtype`	Torch data type

as_numpy_dtype

as_numpy_dtype(dtype, upcast=False)

Convert a torch data type (or a data type name) to a torch data dtype.

Warning

Builtin Python data types int and float are mapped to np.int64 and np.float64, to match np.asarray's behavior. It differs from as_torch_dtype, which maps these types to np.int32 and np.float32.

Parameters:

Name	Type	Description	Default
`dtype`	`str or dtype or dtype`	Input data type	required
`upcast`	`bool`	If the data type is not implemented in NumPy, but its values can be represented by a larger data type that exists in NumPy, return the larger data type. If `False`, raise a `TypeError`.	`False`

Returns:

Name	Type	Description
`dtype`	`dtype`	Numpy data type

moveelem

moveelem(input, source, destination, dim=-1)

Move elements in a tensor

Parameters:

Name	Type	Description	Default
`input`	`tensor`	Input tensor	required
`source`	`[sequence of] int`	Source indices of elements to move	required
`destination`	`[sequence of] int`	Target indices of moved elements	required
`dim`	`int`	Dimension along which to move elements	`-1`

Returns:

Name	Type	Description
`output`	`tensor`

slice_tensor_along

slice_tensor_along(x, index, dim=-1)

Index a tensor along one dimensions.

This function is relatively similar to torch.index_select, except that it uses the native indexing mechanism and can therefore returns a tensor that use the same storage as the input tensor.

It is faster but less versatile than slice_tensor.

Parameters:

Name	Type	Description	Default
`x`	`tensor`	Input tensor.	required
`index`	`int or list[int] or slice`	Indices to select along `dim`.	required
`dim`	`int`	Dimension to index.	`last`

Returns:

Name	Type	Description
`y`	`tensor`	Output tensor.

slice_tensor

slice_tensor(x, index, dim=None)

Index a tensor along one or several dimensions.

This function is relatively similar to torch.index_select, except that it uses the native indexing mechanism and can therefore returns a tensor that use the same storage as the input tensor.

Parameters:

Name	Type	Description	Default
`x`	`tensor`	Input tensor.	required
`index`	`index_like or tuple[index_like]`	Indices to select along each dimension in `dim`. If multiple dimensions are indexed, they must be held in a tuple (not a list). Each index can be a long, list of long, slice or tensor of long, but cannot be an ellipsis or tensor of bool.	required
`dim`	`int or sequence[int]`	Dimensions to index. If it is a list, `index` must be a tuple. By default, the last `n` dimensions (where `n` is the number of indices in `index`) are used.	`None`

Returns:

Name	Type	Description
`y`	`tensor`	Output tensor.

movedims

movedims(input, source, destination)

Moves the position of one or more dimensions

Other dimensions that are not explicitly moved remain in their original order and appear at the positions not specified in destination.

Parameters:

Name	Type	Description	Default
`input`	`tensor`	Input tensor	required
`source`	`int or sequence[int]`	Original positions of the dims to move. These must be unique.	required
`destination`	`int or sequence[int]`	Destination positions for each of the original dims. These must also be unique. If a single destination is provided if it is negative, the last source dimension is moved to `destination` and all other source dimensions are moved to its left. if it is positive, the first source dimension is moved to `destination` and all other source dimensions are moved to its right.	required

Returns:

Name	Type	Description
`output`	`tensor`	Tensor with moved dimensions.

movedims_front2back

movedims_front2back(tensor, dim)

Move the first N dimensions to the back

movedims_back2front

movedims_back2front(tensor, dim)

Move the last N dimensions to the front

shiftdim

shiftdim(x, n=None)

Shift the dimensions of x by n.

Parameters:

Name	Type	Description	Default
`x`	`tensor`	Input tensor.	required
`n`	`int`	Shift. When `n` is positive, `shiftdim` shifts the dimensions to the left and wraps the N leading dimensions to the end. When `n` is negative, `shiftdim` shifts the dimensions to the right and pads with singletons. When `n` is None, `shiftdim` removes all leading singleton dimensions. The number of removed dimensions is returned as well.	`None`

Returns:

Name	Type	Description
`x`	`tensor`	Output tensor.
`n`	`int, if n is None`	Number of removed dimensions