SynID: Passport Synthetic Dataset for Presentation Attack Detection Paper • 2505.07540 • Published May 12, 2025 • 1 • 1
SynID: Passport Synthetic Dataset for Presentation Attack Detection Paper • 2505.07540 • Published May 12, 2025 • 1
view article Article Fixing Gradient Accumulation +4 lysandre, ArthurZ, muellerzr, ydshieh, BenjaminB, pcuenq • Oct 16, 2024 • 67
view reply The key takeaway is that gradient accumulation can make training with large effective batch sizes much more practical when GPU memory is limited. The numerical details around loss scaling are especially useful for avoiding subtle training bugs.