tomosipo.links.torch module

This module adds support for torch arrays as astra.data3d backends

This module is not automatically imported by tomosipo, you must import it manually as follows:

>>> import tomosipo as ts
>>> import tomosipo.torch_support

Now, you may use torch tensors as you would numpy arrays:

>>> vg = ts.volume(shape=(10, 10, 10))
>>> pg = ts.parallel(angles=10, shape=10)
>>> A = ts.operator(vg, pg)
>>> x = torch.zeros(A.domain_shape)
>>> A(x).shape == A.range_shape
True

You can also directly apply the tomographic operator to data on the GPU:

>>> A(x.cuda()).is_cuda
True

class tomosipo.links.torch.TorchLink(shape, initial_value)

Bases: Link

Link implementation for torch arrays

clone()

context()

Context-manager to manage ASTRA interactions

This context-manager makes sure that the current CUDA stream is set to the CUDA device of the current linked data.

Returns:

Return type:

property data

Returns a shared array with the underlying data.

Changes to the return value will be reflected in the astra data.

If you want to avoid this, consider copying the data immediately, using x.data.clone() for instance.

NOTE: if the underlying object is an Astra projection data type, the order of the axes will be in (Y, num_angles, X) order.

Returns:

torch.Tensor

Return type:

torch.Tensor

property linked_data

Returns a numpy array or GPULink object

Returns:

Return type:

new_empty(shape)

new_full(shape, value)

new_zeros(shape)