Cheatsheet — CV Transfer Learning

• Freeze early layers first
• Match input resolution to backbone expectations
• Log class balance and hard negatives
• Prefer macro metrics when classes are skewed