WIP: 1d and non squared 2d input shapes

[dkrako]

Add support for 1d and non-squared 2d input shapes

[dkrako]

I did some first work today for 1d-support, mainly working on helpers/utils.py and adding some lazy fixtures for later use.
The channel-methods from the utils module now support 1d and non-squared 2d input and are fully tested.

For better clarity I renamed the channel first methods as can be seen here:
5e6c0a8

• get_channel_first -> infer_channel_first
• get_channel_first_batch -> make_channel_first
• get_channel_last_batch -> make_channel_last

I hope you don't mind the renaming.

[codecov-commenter]

Codecov Report

Merging #74 (a4e7cb2) into main (8170c25) will decrease coverage by 3.17%.
The diff coverage is 88.62%.

@@            Coverage Diff             @@
##             main      #74      +/-   ##
==========================================
- Coverage   97.36%   94.18%   -3.18%     
==========================================
  Files          16       16              
  Lines        2235     2458     +223     
==========================================
+ Hits         2176     2315     +139     
- Misses         59      143      +84     

Impacted Files	Coverage Δ
quantus/metrics/faithfulness_metrics.py	95.42% <ø> (-3.65%)	⬇️
quantus/metrics/localisation_metrics.py	92.52% <80.37%> (-5.07%)	⬇️
quantus/metrics/axiomatic_metrics.py	93.29% <83.07%> (-4.52%)	⬇️
quantus/metrics/complexity_metrics.py	94.57% <83.33%> (-5.43%)	⬇️
quantus/metrics/randomisation_metrics.py	93.98% <84.21%> (-4.41%)	⬇️
quantus/helpers/tf_model.py	93.87% <87.50%> (+0.39%)	⬆️
quantus/helpers/pytorch_model.py	91.30% <88.88%> (-1.56%)	⬇️
quantus/metrics/robustness_metrics.py	93.31% <90.82%> (-2.68%)	⬇️
quantus/helpers/utils.py	96.66% <94.56%> (-1.04%)	⬇️
quantus/helpers/perturb_func.py	96.42% <95.94%> (-2.40%)	⬇️
... and 6 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 8170c25...a4e7cb2. Read the comment docs.

[annahedstroem]

This PR is comprehensive but 1d-support is not implemented for the following perturbation functions:

• baseline_replacement_by_blur: uses conv2D_numpy, scipy.signal.convolve
• rotation, translation_x_direction, translation_y_direction: these use cv2 which expect images as input

Also:

• get_superpixels also doesn't support 1d, as skimage uses 2d only.
• IterativeRemovalOfFeatures therefore won't work for 1d yet.

[annahedstroem]

A summary of contribution:

• Comprehensive changes that allow for more flexibility in input shapes for Quantus - 1D and flexible img_sizes where img width != height.
• Additional pytests tests created for these purposes

1d-support is NOT implemented for the following functions/ metrics:

• baseline_replacement_by_blur: uses conv2D_numpy, scipy.signal.convolve
• rotation, translation_x_direction, translation_y_direction: these use cv2 which expect images as input
• get_superpixels also doesn't support 1d, as skimage` uses 2d only.
• IterativeRemovalOfFeatures therefore won't work for 1d yet.

Other notes:

• We disregard for now the decrease in code coverage by approx ~3% as it relates to the depreciation warnings added to each of the metrics
• Also an issue is created for making a tutorial to showcase this new functionality New tutorial: for how to use Quantus with 1D shapes #83 but does not necessarily have to be completed to merge this PR

OK - ACCEPTED.

[dkrako]

Done as discussed.

[annahedstroem]

OK