Controlling the Chaos: Reproducible Deep Learning Experiments Using Lightning and Hydra
summary
Shows a concrete pipeline that combines PyTorch Lightning for accelerators, checkpointing and Trainer multi-GPU support with Hydra for Hydra config groups and multirun sweeps; the team ships Quadra, a public repo of Quadra recipe folders that instantiate data transforms, backbones (timm, torch.hub), losses and callbacks from YAML, run tasks (classification, segmentation, self‑supervised, anomaly) and export TorchScript. Example runs (imagenet-like, Oxford seg, MNIST anomaly) produce separate experiment folders with resolved configs, checkpoints, pickles and git commit hashes so experiments are reproducible and sweeps (e.g. 2 trainer variants × 3 lrs = 6 runs) execute automatically.
AI-written summary · about
chapters (28)
- 0:00Introduction
- 0:36Happy Data Scientist Checklist
- 1:33PyTorch
- 2:23Image Classification
- 3:59PyTouch Lighting
- 4:23Refactor Code
- 4:51Training Step
- 5:41Running the Code
- 8:03Code Refactor
- 9:36YAML Files
- 10:56Checklist
- 11:08Repository
- 12:07Quadra
- 12:49Transforms
- 13:07Configuration Files
- 14:18Feature Extractor
- 14:57Lightning Module
- 15:42Hydra Module
- 16:26Hydra Task
- 17:00Hydra Data Mode
- 17:24Final View
- 18:15Classification
- 18:39Segmentation
- 19:10Anomaly Detection
- 19:50Report reproducibility
- 21:06Rerun experiment
- 21:29Shoutout
- 22:06QA Time
related talks
21mImplementing your AI strategy with PyTorch Lightning in the cloud
30mHow to Trust Your Deep Learning Code
18mLightning Talk: TorchRL - RLHF Support
26mThunder: Distribute and Optimize Your PyTorch Models With Zero...
25mHow to Train an LLM with Argo Workflows and Hera
14mLightning Talk: Large-Scale Distributed Training with Dynamo and...
25mInto Generative AI with PyTorch Lightning 2.0
17mScaling Reinforcement Learning