pytorch-lightning

Overview

PyTorch Lightning is a lightweight wrapper for PyTorch that decouples the research code from the engineering code. It automates 40+ engineering details like epoch loops, optimization, and hardware acceleration, while allowing researchers to retain full control over the model logic.

When to Use

Use Lightning when you want to scale models to multi-GPU or multi-node environments without changing training code, or when you want to eliminate boilerplate for logging, checkpointing, and reproducibility.

Decision Tree

1. Are you testing code logic on a small subset?
   • YES: Use Trainer(fast_dev_run=True).
2. Do you need to scale to multiple GPUs?
   • YES: Set accelerator='gpu' and devices=N in the Trainer.
3. Do you have logic that is non-essential to the model (e.g., special logging)?
   • YES: Implement it as a Callback.

Workflows

1. Transitioning from Raw PyTorch to Lightning

   1. Define a class inheriting from L.LightningModule.
   2. Move model architecture into __init__ and logic into training_step.
   3. Implement configure_optimizers to return the optimizer and optional scheduler.
   4. Instantiate an L.Trainer and pass the model and DataLoader to trainer.fit().

2. Multi-GPU Training Scaling

   1. Initialize the Trainer with accelerator='gpu' and devices=N.
   2. Set strategy='ddp' or 'deepspeed_stage_2' for multi-node/large-scale runs.
   3. Optionally enable 16-bit precision using precision='16-mixed'.
   4. Run the script without code changes to standard logic.

3. Ensuring Training Reproducibility

   1. Call seed_everything(seed, workers=True) at the start of the script.
   2. Initialize the Trainer with deterministic=True.
   3. Avoid data-dependent logic in transforms that isn't handled by the derived seeds.

Non-Obvious Insights

• Overhead Analysis: The 'barebones' mode in the Trainer is specifically for overhead analysis and disables almost all logging and checkpointing for speed.
• Callback State Management: Custom callbacks must implement a state_key property if multiple instances of the same callback type are used in a single Trainer.
• Superior Debugging: The fast_dev_run flag is superior to limiting batches manually because it avoids all side effects like checkpointing or logging that might clutter the workspace during debugging.

Evidence

• "The Lightning Trainer automates 40+ tricks including: Epoch and batch iteration, optimizer.step(), loss.backward(), optimizer.zero_grad() calls." (https://lightning.ai/docs/pytorch/stable/starter/introduction.html)
• "By setting workers=True in seed_everything(), Lightning derives unique seeds across all dataloader workers." (https://lightning.ai/docs/pytorch/stable/common/trainer.html)

Scripts

• scripts/pytorch-lightning_tool.py: Template for a LightningModule and Trainer setup.
• scripts/pytorch-lightning_tool.js: Script to trigger Lightning training with CLI arguments.

Dependencies

• lightning
• torch

References

• PyTorch Lightning Reference