Improve confusion matrix plotting

[JonasVerbickas]

What does this PR do?

Current confusion matrix plotting tends to produce unusable results.
When I tried to create a normalized multilabel confusion matrix plot this cluttered mess was produced:

1. When normalizing converting tensor to float using val.item() produces numbers with too many decimal places.
2. Redundant "True class" and "Predicted class" use up a lot of space and make the graph unreadable
3. Longer labels overlap

Round floats to avoid floating point errors leading to UI overflow.

Rounding to two decimal places seems reasonable since it's difficult to fit more digits into multilabel confusion matrices.

ax.text(jj, ii, str(round(val.item(), 2)), ha="center", va="center", fontsize=15)

Rounding the val tensor itself does not work here. The issue is with the item() method converting incorrectly.

Changing this line converts this plot:

It might be worth considering using less aggressive rounding for simpler binary confusion matrices since they have room to display more digits without any overlap.

Removing redundant "True class" and "Predicted class"

Code reducing the number of times x and y labels are shown:

if i // cols == rows-1:  # bottom row only
    ax.set_xlabel("Predicted class", fontsize=15)
if i % cols == 0:  # leftmost column only
    ax.set_ylabel("True class", fontsize=15)

Produces a much cleaner plot without sacrificing any information:

Reduce overlap

By utilizing constrained_layout=True:

fig, axs = plt.subplots(nrows=rows, ncols=cols, constrained_layout=True) if ax is None else (ax.get_figure(), ax)

Reduces overlap between labels horizontally and vertically from {0, 1} ticks:

Even though this does not fix the top row, the longest names in this example are extreme.
Ensuring proper middle-row separation is good enough for now.

Before submitting

• Was this discussed/agreed via a Github issue? (no need for typos and docs improvements)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure to update the docs?
• Did you write any new necessary tests?

PR review

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

Did you have fun?

Make sure you had fun coding 🙃

---

📚 Documentation preview 📚: https://torchmetrics--2358.org.readthedocs.build/en/2358/

[Borda]

yes, the readability for binary cases is bad, just thinking about making the long labels as multi-line or smaller font? 🤔

[JonasVerbickas]

If I have the time, I will look into multiline labels this week.
That said, I think this commit is a good first step towards confusion matrix plot clarity.

[codecov]

Codecov Report

Merging #2358 (1337bcd) into master (b187bfd) will decrease coverage by 0%.
Report is 1 commits behind head on master.
The diff coverage is 0%.

Additional details and impacted files

@@          Coverage Diff           @@
##           master   #2358   +/-   ##
======================================
- Coverage      69%     69%   -0%     
======================================
  Files         303     303           
  Lines       17058   17060    +2     
======================================
  Hits        11760   11760           
- Misses       5298    5300    +2     

[SkafteNicki]

@JonasVerbickas thanks for the contribution. We are going to merge this PR such that it become part of the v1.3.1release. Feel free to send a new PR with further improvements to the plotting capabilities of torchmetrics.