tomosipo.torch_support module

This module provides functions to interoperate with torch

class tomosipo.torch_support.AutogradOperator(operator, num_extra_dims=0, is_2d=False)

Bases: object

A linear tomographic projection operator with torch autograd support

This class has almost the same interface as a normal tomosipo operator, but it supports pytorch autograd, allowing for automatic differentiation. This can be used for example on the reconstruction algorithms in ts_algorithms. Additive operators, numpy inputs and outputs, and transformations are not supported on AutogradOperators.

property T

The transpose operator

This property returns the transpose (backprojection) operator.

property astra_compat_pg

property astra_compat_vg

property domain

The domain (volume geometry) of the operator

property domain_shape

The expected shape of the input (volume) data

property range

The range (projection geometry) of the operator

property range_shape

The expected shape of the output (projection) data

transpose()

Return backprojection operator

class tomosipo.torch_support.Backward(*args, **kwargs)

Bases: Function

static backward(ctx, grad_output)

static forward(ctx, input, vg, pg, projector)

class tomosipo.torch_support.Forward(*args, **kwargs)

Bases: Function

static backward(ctx, grad_output)

static forward(ctx, input, vg, pg, projector)

class tomosipo.torch_support.OperatorFunction(*args, **kwargs)

Bases: Function

static backward(ctx, grad_output)

static forward(ctx, input, operator, num_extra_dims=0, is_2d=False)

tomosipo.torch_support.autograd_operator(volume_geometry, projection_geometry, voxel_supersampling=1, detector_supersampling=1, num_extra_dims=0, is_2d=False)

Create a tomographic operator with PyTorch autograd support.

Parameters:

• volume_geometry (VolumeGeometry) – The domain of the operator.

• projection_geometry (ProjectionGeometry) – The range of the operator.

• voxel_supersampling (int (optional)) – Specifies the amount of voxel supersampling, i.e., how many (one dimension) subvoxels are generated from a single parent voxel. The default is 1.

• detector_supersampling (int (optional)) – Specifies the amount of detector supersampling, i.e., how many rays are cast per detector. The default is 1.

• num_extra_dims (int (optional)) – Number of extra dimensions to prepend to the input and output of this operator. This can be useful to add channel and batch dimensions when training neural networks. The default is 0.

• is_2d (bool (optional)) – Whether to remove the first dimension of the operator, resulting in a 2D operator. The default is False.

Returns:

A tomographic operator with PyTorch autograd support.

Return type:

AutogradOperator

tomosipo.torch_support.backward(input, vg, pg, projector=None)

tomosipo.torch_support.forward(input, vg, pg, projector=None)

tomosipo.torch_support.to_autograd(operator, num_extra_dims=0, is_2d=False)

Converts an operator to an autograd function

Example:

>>> import tomosipo as ts
>>> vg = ts.volume(shape=10)
>>> pg = ts.parallel(angles=10, shape=10)
>>> A = ts.operator(vg, pg)

>>> f = to_autograd(A)
>>> vd = torch.randn(*A.domain_shape, requires_grad=True)
>>> f(vd).sum().backward()
>>> vd.grad is not None
True

Likewise, you may use the transpose:

>>> g = to_autograd(A.T)
>>> pd = torch.randn(*A.T.domain_shape, requires_grad=True)
>>> g(pd).sum().backward()
>>> vd.grad is not None
True

Parameters:

• operator (Operator) – Tomosipo operator whose behavior will be mimicked

• num_extra_dims (int (optional)) – Number of extra dimensions to prepend to the input and output of this operator. Set this to 2 to add channel and batch dimensions when taining neural networks. The default is 0.

• is_2d (bool (optional)) – Whether to remove the first dimension of the operator, resulting in a 2D operator. The default is False.

Returns:

An autograd enabled function

Return type:

Function