Issue with using callback to show intermediate results in iterative solvers

[ozanoktem]

Most solvers in ODL are iterative and my understanding is that setting a callback offers a way to pass instructions to such a solver that can be executed in each iterate. There now seems to be an issue with the way the callback-function in ODL interfaces with the show-function. More specifically, one can use callback to show intermediate reconstructions in each iterate by setting

callback_settings = (odl.solvers.CallbackPrintIteration() & odl.solvers.CallbackShow())

and then calling the ODL solver with callback=callback_settings. However, this yields an attribute error that allow the first iterate to be shown, but then iterates are terminated with the message

iter = 0
iter = 1
Traceback (most recent call last):

File ~/opt/anaconda3/envs/ODL/lib/python3.9/site-packages/spyder_kernels/py3compat.py:356 in compat_exec
exec(code, globals, locals)

File ~/Library/CloudStorage/Dropbox/Teaching/KTH/SF3582 Inverse Problems/Laborations/issue_callback.py:42
odl.solvers.iterative.iterative.landweber(fwd_op, reco, data, niter=10, callback=callback_settings)

File ~/odl/odl/solvers/iterative/iterative.py:120 in landweber
callback(x)

File ~/odl/odl/solvers/util/callback.py:133 in call
p(result)

File ~/odl/odl/solvers/util/callback.py:666 in call
self.fig = x.show(title, fig=self.fig,

File ~/odl/odl/discr/discr_space.py:1515 in show
return show_discrete_data(values, part, title=title, method=method,

File ~/odl/odl/util/graphics.py:439 in show_discrete_data
csub.colorbar.set_clim(minval, maxval)

AttributeError: 'Colorbar' object has no attribute 'set_clim'

Below is a small example demonstrating the issue.

import numpy as np
import odl
      
# Reconstruction space: discretized functions on the rectangle
# [-20, 20]x[-20,20] with 300 samples per dimension.
reco_space = odl.uniform_discr(
    min_pt=[-1, -1], max_pt=[1, 1], shape=[256, 256], dtype='float32')

# Create a discrete Shepp-Logan phantom (modified version)
phantom = odl.phantom.shepp_logan(reco_space, modified=True)

# Data space: Parallel-beam geometry with flat detector
# Angles: uniformly spaced, n = 1000, min = 0, max = pi
angle_partition = odl.uniform_partition(0, np.pi, 1000)
# Detector: uniformly sampled, n = 500, min = -30, max = 30
detector_partition = odl.uniform_partition(-1.5, 1.5, 300)
# Make a parallel beam geometry with flat detector
geometry = odl.tomo.Parallel2dGeometry(angle_partition, detector_partition)

# Forward operator: Ray transform 
fwd_op = odl.tomo.RayTransform(reco_space, geometry)

# Create projection data by calling the ray transform on the phantom
data = fwd_op(phantom)

# Optionally pass callback to the solver to display intermediate results
callback_settings = (odl.solvers.CallbackPrintIteration() & odl.solvers.CallbackShow())

# Built in Landweber reconstruction 
reco = fwd_op.domain.zero()
odl.solvers.iterative.iterative.landweber(fwd_op, reco, data, niter=10, callback= callback_settings)

[paulhausner]

I think this is just a problem with odl.utils.graphics usage of matplotlib. Rescaling the colorbar seems to not work anymore but it should be possible to instead rescale the plot (https://matplotlib.org/stable/api/_as_gen/matplotlib.pyplot.clim.html). See #1635

[leftaroundabout]

So, this particular issue is indeed easily fixed by

--- a/odl/util/graphics.py
+++ b/odl/util/graphics.py
@@ -436,14 +436,13 @@ def show_discrete_data(values, grid, title=None, method='',
                 plt.colorbar(mappable=csub, ticks=ticks, format=fmt)
             elif update_in_place:
                 # If it exists and we should update it
-                csub.colorbar.set_clim(minval, maxval)
+                csub.set_clim(minval, maxval)
                 csub.colorbar.set_ticks(ticks)
                 if '%' not in fmt:
                     labels = [fmt] * len(ticks)
                 else:

There is then another error a few lines down:

  File "/home/js/progwrit/python/odl/odl/util/graphics.py", line 446, in show_discrete_data
    csub.colorbar.draw_all()
AttributeError: 'Colorbar' object has no attribute 'draw_all'

Colorbar.draw_all has been deprecated since 3.6 and is gone in 3.8. The Matplotlib documentation recommends this replacement:

--- a/odl/util/graphics.py
+++ b/odl/util/graphics.py
@@ -443,7 +443,7 @@ def show_discrete_data(values, grid, title=None, method='',
                 else:
                     labels = [fmt % t for t in ticks]
                 csub.colorbar.set_ticklabels(labels)
-                csub.colorbar.draw_all()
+                fig.draw_without_rendering()
 
     # Set title of window
     if title is not None:

This does not work right though: all axis labels and also the colorbar itself vanish.

Honestly I don't see why this line is required at all: if I remove it completely the plot seems to update as desired.

Correction: actually the colorbar is not updated without the line.

[leftaroundabout]

fig.canvas.draw_idle() seems to do the trick: