Training a Model on Google Colab, A Survival Guide

A practical guide to training ML models on Google Colab without losing progress to runtime disconnects. Covers how Colab actually works under the hood — borrowed VMs, FUSE-mounted Drive as a network filesystem, dynamic GPU quotas — and builds a resilient training workflow around these constraints. Key patterns include: always checkpoint to Drive (save locally first, then copy), extract datasets to /content before training to avoid 50,000 network round-trips per epoch, pin package versions to fight notebook rot, and use the official VS Code extension instead of banned SSH tunnels. Includes a complete runnable ResNet18 fine-tuning example that survives deliberate runtime kills and resumes cleanly from the last saved epoch.