mentor

mentor is a PyTorch training framework that keeps a model’s weights, training history, provenance, and inference state together in a single .pt checkpoint file.

Contents

• Quickstart
  • Installation
  • Subclassing Mentee
  • Training loop
  • Resuming training
  • Inference-only loading
  • Batteries-included inference state
  • Gradient accumulation
• Curriculum Learning Support
  • Design principles
  • Layer inspection
  • Freezing and unfreezing
  • Per-layer learning rate coefficients
  • Combining freeze and LR coefficients
  • API reference
• Developing torch-mentor
  • Clone and install
  • Running the test suite
  • Building the docs locally
  • Code style
  • Versioning
  • Publishing to PyPI
• API Reference
  • mentor.mentee
  • mentor.trainers
  • mentor.reporting
• Examples
  • CIFAR-10 — reproducing He et al. (2016)
  • HuggingFace Transfer Learning with mentor
• Command-line tools
  • mtr_checkpoint
  • mtr_plot_file_hist

Indices and tables

• Index

• Module Index

• Search Page