[Fix] Fix indentation bug in hd_loss that mishandles ignore_index

[Mr-Neutr0n]

Motivation

There is an indentation bug in the hd_loss() function in mmseg/models/losses/huasdorff_distance_loss.py. The class_weight multiplication and total_loss accumulation are placed outside the if i != ignore_index guard block.

This causes two problems:

1. When a class index i equals ignore_index, no loss is computed for that iteration, but the previous iteration's hd_loss value is re-added to total_loss (and potentially re-weighted with the wrong class_weight[i]). This silently produces incorrect loss values.
2. If ignore_index happens to equal 1 (the first value in range(1, num_class)), hd_loss is never assigned, causing a NameError.

Modification

Move the class_weight check and total_loss += lines inside the if i != ignore_index block so that only computed (non-ignored) class losses contribute to the total.

Before (buggy)

for i in range(1, num_class):
    if i != ignore_index:
        ...
        hd_loss = multiplied.mean()
    if class_weight is not None:    # <-- outside the guard
        hd_loss *= class_weight[i]
    total_loss += hd_loss            # <-- outside the guard

After (fixed)

for i in range(1, num_class):
    if i != ignore_index:
        ...
        hd_loss = multiplied.mean()
        if class_weight is not None:    # <-- inside the guard
            hd_loss *= class_weight[i]
        total_loss += hd_loss            # <-- inside the guard

BC-breaking (No)

This is a pure bugfix. Users who were not using HuasdorffDisstanceLoss are unaffected. Users who were using it will get correct loss values.

[CLAassistant]

All committers have signed the CLA.

[Mr-Neutr0n]

I have read the CLA Document and I hereby sign the CLA

[Mr-Neutr0n]

Re-ran the hausdorff loss logic locally with the fix applied (PyTorch 2.x, Python 3.12) — hd_loss now correctly skips class_weight multiplication and total_loss accumulation when i == ignore_index, as the test setup expects. The indentation fix is correct.

The pr_stage_test failure was on the minimum_version_cpu job (PyTorch 1.8.1, Python 3.7), which is the most restrictive env in the matrix. That job wasn't retried recently (and the CircleCI artifacts aren't browsable from my side without re-running). The indentation change is type- and behavior-equivalent — it can't have introduced a new failure on PyTorch 1.8.1 because:

1. It only moves two existing lines (hd_loss *= class_weight[i] and total_loss += hd_loss) inside an existing if i != ignore_index: block
2. The semantics when i != ignore_index are unchanged
3. The semantics when i == ignore_index are now correct (was a real bug)

My best guess: the 1.8.1/3.7 failure was a pre-existing flake on that combo (the matrix is quite old) or an unrelated dependency-resolution issue. Worth re-running just that job with gh workflow run or pushing an empty commit to retrigger — happy to do either.

[Mr-Neutr0n]

Pushed a no-op commit (097f1d4) to retrigger pr_stage_test on the current branch. The actual code change is still the same indentation fix in hd_loss; this is just to get a fresh run on the flaky old-matrix job.