Add tests

[dpascualhe]

Add automatic tests using pytest + github actions. A good starting point: metrics.py.

[mansipatil12345]

Hello sir,

I noticed this issue regarding eval bugs and the accuracy metric. I’ve reviewed the problem and have a clear approach to address it:

Improving shape checks,
Handling division by zero safely, and
Refining per-class metric calculations.
Before I proceed, I’d like to get your guidance:

Does this approach align with your expectations?
Are there any specific considerations or requirements I should keep in mind while implementing the fixes?
I can also add tests to validate the changes to ensure everything works as expected. Please let me know your thoughts or suggestions before I move forward!

[dpascualhe]

Hello @mansipatil12345 👋

Thanks for your interest in the project! The current metrics implemented are working as intended, but adding safeguards like the ones proposed seems like a good idea. My only doubt, what do you mean by Refining per-class metric calculations?

If you plan to work on said safeguards, please open another issue as this one is specific for adding tests.

[mansipatil12345]

Thank you for reviewing my request and providing guidance. Based on your feedback, I will proceed with the following steps to address the issue:

Shape Checks:

Add robust checks to ensure predictions and ground truth tensors match in dimensions. This will prevent shape mismatches during metric computation.

Handling Division by Zero:

Introduce safeguards (e.g., conditional checks or np.where) to handle cases where a class has no ground truth samples (e.g., zero denominator). Metrics for such classes will return NaN to avoid errors and maintain clarity.

Refining Per-Class Metric Calculations:

Ensure that calculations for metrics like IoU and accuracy are correctly handled on a class-by-class basis.
This includes managing edge cases like rare or missing classes, where no pixels exist in the ground truth for a given class.

Validation Tests:

Write unit tests to validate the fixes, ensuring all edge cases (e.g., missing classes, shape mismatches) are covered.
I believe this approach aligns with the issue’s requirements, but I’d appreciate your confirmation or any additional suggestions before I proceed with the implementation.

Looking forward to your thoughts!