Add utility to configure PDHG for with explicit TGV or TV regularisation #1766
Conversation
|
Hey all, This still needs a few changes. Currently the Would you like me to do a PR or let you guys take care of it? |
paskino
changed the title
|
Hi @lauramurgatroyd - I like the style of the unit tests for explicit TV. They are comprehensive without being expensive. I have suggested a few changes to them and made some comments on the new utility file. |
| from cil.optimisation.functions import MixedL21Norm, BlockFunction, L2NormSquared, ScaledFunction | ||
| from cil.optimisation.operators import BlockOperator, IdentityOperator, GradientOperator, \ | ||
| SymmetrisedGradientOperator, ZeroOperator | ||
|
|
There was a problem hiding this comment.
Can we call this file something other than PDHG.py? e.g. set_up_PDHG.py or something similar?
| Forward operator. | ||
| data : AcquisitionData | ||
| alpha : float | ||
| Regularisation parameter. |
There was a problem hiding this comment.
Regularisation parameter for which part?
|
|
||
| def setup_explicit_TGV(A, data, alpha, delta=1.0, omega=1): | ||
| '''Function to setup LS + TGV problem for use with explicit PDHG | ||
|
|
There was a problem hiding this comment.
Need a TGV equation in here, defining alpha and beta and omega
| delta : float, default 1.0 | ||
| The Regularisation parameter for the symmetrised gradient, beta, can be controlled by delta | ||
| with beta = delta * alpha. | ||
| omega : float, default 1.0 |
There was a problem hiding this comment.
Least squares uses c instead of omega, perhaps we could follow suit?
| The constant in front of the data fitting term. Mathematicians like it to be 1/2 but it is 1 by default, | ||
| i.e. it is ignored if it is 1. | ||
|
|
||
| Returns: |
There was a problem hiding this comment.
Could we add a code snippet to explain how to use the returned K and F in PDHG? Also need to explain briefly what we mean by "explicit"
| function = F[0].function | ||
|
|
||
| np.testing.assert_equal(type(function), L2NormSquared) | ||
| np.testing.assert_array_equal(function.b.as_array(), ad.as_array()) |
There was a problem hiding this comment.
I would also be tempted to evaluate your functions on a small image, so reduce the size of the geometry to save computational cost, create a random test_data and then do
self.assertAlmostEqual(function(test_data), omega*LeastSquares(b=ad)(test_data))
| np.testing.assert_equal(type(K[1].operator), GradientOperator) | ||
| ig = ad.geometry.get_ImageGeometry() | ||
| expected_grad = alpha*GradientOperator(ig) | ||
| np.testing.assert_allclose(expected_grad.direct(ad)[1].as_array(), expected_grad.direct(ad)[1].as_array()) |
There was a problem hiding this comment.
Is that comparing the same thing?
| np.testing.assert_allclose(expected_grad.direct(ad)[1].as_array(), K[1].direct(ad)[1].as_array(), 10**(-4)) | ||
| np.testing.assert_allclose(expected_grad.direct(ad)[0].as_array(), K[1].direct(ad)[0].as_array(), 10**(-4)) |
There was a problem hiding this comment.
I think the expected_grad sould be operator on an image geometry not on acquisition data? Thus you should not be using ad
|
|
||
| # F[1] | ||
|
|
||
| np.testing.assert_equal(MixedL21Norm, type(F[1])) |
There was a problem hiding this comment.
Similarly, can you test this on an object?
casperdcl
force-pushed
the
master
branch
3 times, most recently
from
a910556
to
f8b513b
Compare
Description
Add utility to configure PDHG for with explicit TGV or TV regularisation with Least Squares.
Example Usage
An example is on CIL-Demos TGV_tomography.py
Changes
maxtoBlockDataContainerdotbetween aBlockDataContainerand aDataContainerTesting you performed
Related issues/links
Checklist
Contribution Notes
Please read and adhere to the developer guide and local patterns and conventions.