Fix scheduler stepping logic and checkpoint state persistence in PR#11

[Copilot]

Code review of PR#11 revealed three critical correctness issues in the optimizer/scheduler factory implementation that would cause incorrect training behavior.

Issues Fixed

1. LinearWarmupWrapper dual counter desynchronization

The wrapper maintained both _step_count and inherited last_epoch, causing warmup phase to be off by one step. Now uses only last_epoch from parent _LRScheduler.

Before:

def __init__(self, ...):
    self._step_count = 0  # Custom counter
    super().__init__(optimizer, last_epoch)  # Parent has last_epoch=-1
    
def get_lr(self):
    if self._step_count < self.warmup_steps:  # Uses custom counter
        alpha = self._step_count / self.warmup_steps

After:

def __init__(self, ...):
    # Removed self._step_count
    super().__init__(optimizer, last_epoch)
    
def get_lr(self):
    current_step = self.last_epoch + 1  # Uses parent counter consistently
    if current_step < self.warmup_steps:
        alpha = current_step / self.warmup_steps

2. EMA shadow weights lost on checkpoint resume

EMA.state_dict() existed but was never called. Training resume would reinitialize EMA from current weights, losing accumulated shadow.

• Added scheduler and ema parameters to save_checkpoint_bundle() and load_checkpoint_into()
• Reorganized initialization: EMA and scheduler created before checkpoint resume
• Both save_checkpoint_bundle() call sites updated

3. Scheduler state not persisted

Similar to EMA - scheduler would restart from step 0 on resume, breaking learning rate schedule continuity.

Testing

• All 12 unit tests pass
• Ruff linting clean
• CodeQL: 0 vulnerabilities

Files Modified

• rtdetr_pose/rtdetr_pose/sched_factory.py - Fixed LinearWarmupWrapper logic
• rtdetr_pose/tools/train_minimal.py - Added checkpoint persistence for scheduler/EMA state

Original prompt

review PR#11

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.

[Copilot]

Pull request overview

This PR attempts to fix three critical issues identified in PR#11's optimizer and scheduler factory implementation: dual counter desynchronization in LinearWarmupWrapper, missing EMA state persistence in checkpoints, and missing scheduler state persistence in checkpoints.

Changes:

• Added checkpoint persistence for scheduler and EMA state by extending save_checkpoint_bundle() and load_checkpoint_into() functions
• Reorganized initialization order to create scheduler and EMA before checkpoint resume to enable state loading
• Removed dual counter from LinearWarmupWrapper to use only the inherited last_epoch counter

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 6 comments.

File	Description
rtdetr_pose/rtdetr_pose/sched_factory.py	New scheduler factory module with LinearWarmupWrapper, build_scheduler(), and EMA class implementation
rtdetr_pose/tools/train_minimal.py	Added scheduler/EMA parameters to checkpoint functions, integrated factory scheduler/EMA into training loop, added CLI arguments for scheduler/EMA configuration
SECURITY_SUMMARY.md	Added security review summary documenting CodeQL scan results and security considerations for PR#11
PR11_REVIEW_FIXES.md	Documentation describing the three issues found in PR#11 and the fixes applied

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.