necatiincekara/gradient-check

Gradient Check

Use this skill when learning stalls or quantum parameters look inactive.

Workflow

1. Inspect:
   • src/model.py
   • src/trainable_quantum_model.py
   • src/enhanced_training.py
   • any active training script
2. Verify where gradients should flow and where dtype conversions happen.
3. Compare quantum vs classical gradient magnitudes.
4. Flag:
   • vanishing gradients
   • exploding gradients
   • barren plateau-like behavior
   • AMP or dtype boundary issues
5. Give concrete fixes in priority order:
   • residual paths
   • gradient scaling
   • optimizer separation
   • circuit depth or initialization changes
   • precision boundaries around the quantum layer

Repository note: the newer trainable path already introduced float32 boundaries and stabilization logic; check that path before assuming the base model behavior still applies.