scaffolding-pytorch-training-loop

What problem does it solve?

This Skill removes the boilerplate and failure-prone parts of building a serious PyTorch training loop, including reproducibility, checkpoint recovery, and training stability.

Core Features & Use Cases

• Deterministic Training: Seeds Python, NumPy, PyTorch CPU and CUDA, and DataLoader workers so runs are reproducible.
• Production Training Controls: Adds AMP, GradScaler, gradient clipping, learning-rate scheduling, early stopping, and telemetry logging.
• Reliable Resume Support: Saves and restores model, optimizer, scheduler, scaler, epoch, best metric, and RNG state so preempted jobs continue correctly.
• Use Case: Start a new vision, NLP, or tabular deep-learning project and generate a loop that can survive crashes, fit larger batches, and produce consistent results.

Quick Start

Use this skill to scaffold a production-grade PyTorch training loop for my model with deterministic seeding, mixed precision, checkpoint resume, early stopping, and W&B logging.