Fix the PPLCNetV3 detection backbone channel count

[kozlov721]

Purpose

PPLCNetV3 crashes when a configuration sets use_detection_backbone
to true. The detection branch has never worked. This PR corrects the
channel count and adds the tests that the branch never had.

The PR also unblocks the codecov check on #443. That PR renames
detecion_blocks to detection_blocks and corrects an "initalized"
typo in EMACallback. Both renamed lines sit in untested code, so
codecov counts them as missed and fails codecov/patch. The new tests
cover them.

Specification

• Use the block widths directly in the PPLCNetV3 detection backbone.
  LCNetV3Layer scales its out_channels in the constructor. The
  backbone scaled the same value a second time:

  blocks_out_channels = [
      scale_up(blocks[i].out_channels, self.scale) for i in range(1, 5)
  ]

  The rec-light variant uses the scale 0.95. The second scale turns
  the true block widths [64, 128, 240, 480] into
  [64, 128, 224, 464]. The detection convolutions then expect 224
  channels, but the blocks give 240:

  RuntimeError: Given groups=1, weight of size [53, 224, 1, 1],
  expected input[2, 240, 6, 40] to have 224 channels,
  but got 240 channels instead
  

• Add tests/unittests/test_pplcnet_v3.py. It covers the two backbone
  modes:

  • The recognition mode gives 5 outputs. The last output is pooled to
    (1, max_text_len).
  • The detection mode gives 4 outputs with the widths
    [15, 22, 53, 456].

• Add test_ema_before_fit_start. It covers the EMACallback.ema
  guard. The guard raises before on_fit_start builds the model.

No variant sets use_detection_backbone to true. No test covered
the branch. The defect stayed hidden for these two reasons.

Dependencies & Potential Impact

• Not breaking. The detection branch raised a RuntimeError before
  this change. No configuration can depend on the old widths.
• The change does not touch the recognition path. The
  ocr_recognition predefined model uses that path.
• The tests read only the forward outputs. They do not name the
  detecion_blocks attribute. Pre-docs cleanup: fix the typo'd names and tighten the lint gate #443 can thus rename the attribute
  without a test change.
• No dependency changes.

Deployment Plan

None / not applicable — the changes ship with the next release.

Testing & Validation

• test_detection_backbone fails on main with the RuntimeError
  above. It passes with this change.
• pytest tests/unittests: 508 passed. The run excludes
  tests/unittests/test_losses, which holds the known expired
  gcloud-ADC failures.
• The pre-commit hooks ran on both commits. All hooks pass.
• pyright --warnings: at parity with main. The one error is the
  pre-existing aimet_torch import from the optional extra.

AI Usage

Assisted-by: Claude:claude-opus-5

Submitted code was reviewed by a human: NO

The author is taking the responsibility for the contribution: YES

🤖 Generated with Claude Code

https://claude.ai/code/session_01QA7RcK3rsE5mHX7J6X1Vxh

Summary by CodeRabbit

• Bug Fixes

  • Corrected channel sizing in the PPLCNetV3 detection backbone to ensure detection features have the expected dimensions.
  • Improved EMA callback behavior when accessed before training initialization by providing a clear error.

• Tests

  • Added coverage for PPLCNetV3 recognition and detection outputs, including feature dimensions.
  • Added validation for EMA access before training begins.

[coderabbitai]

Warning

Review limit reached

Next included review available in 15 minutes.

Check out review usage here.

View limit details

Limit details: You’ve used the included review currently available.

You've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Team

Run ID: de908298-15b9-4c94-8662-6f98f57a1503

📥 Commits

Reviewing files that changed from the base of the PR and between 0fb0481 and 8033ee5.

📒 Files selected for processing (1)

• tests/unittests/test_pplcnet_v3.py

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Team

Run ID: 0499220b-afe3-4973-a688-31ef34aa303f

📥 Commits

Reviewing files that changed from the base of the PR and between 407c6a1 and 0fb0481.

📒 Files selected for processing (1)

• luxonis_train/nodes/backbones/pplcnet_v3/pplcnet_v3.py

Included review availability: Your plan provides up to 1 included review per hour; 0 remain after this review.

---

📝 Walkthrough

Walkthrough

The change removes redundant PPLCNetV3 channel scaling and adds tests for recognition, detection, and EMA initialization behavior.

Changes

PPLCNetV3 backbone validation

Layer / File(s)	Summary
Detection channel construction luxonis_train/nodes/backbones/pplcnet_v3/pplcnet_v3.py	blocks_out_channels now uses block output channels without applying a second scale.
Backbone output tests tests/unittests/test_pplcnet_v3.py	Tests validate recognition output dimensions and detection feature channel widths for the rec-light backbone.

EMA callback initialization

Layer / File(s)	Summary
EMA initialization test tests/unittests/test_callbacks/test_ema.py	The test verifies that accessing ema before on_fit_start raises ValueError with the expected message.

Estimated code review effort: 2 (Simple) | ~10 minutes

Merge Risk: ⚪ Minimal · up to 0fb04

The detection backbone now uses the actual layer output widths, preventing channel mismatches during detection model construction. Recognition behavior remains covered, and the change is ready to merge.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 14.29% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 7 functions across 3 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main change: fixing the PPLCNetV3 detection backbone channel count.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches 💡 1

📝 Generate docstrings 💡

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/pplcnet-detection-backbone

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94.61%. Comparing base (3c668bd) to head (8033ee5).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff            @@
##           main     #446       +/-   ##
=========================================
+ Coverage      0   94.61%   +94.61%     
=========================================
  Files         0      290      +290     
  Lines         0    15868    +15868     
=========================================
+ Hits          0    15013    +15013     
- Misses        0      855      +855     

Flag	Coverage Δ
pytest	94.61% <100.00%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.