the-lay / kinlin

Work paused! PyTorch framework for rapid experimentation

🔥 kinlin

derived from kindling (n., /ˈkɪnd.lɪŋ/) - material that can be readily ignited

---

🚧 WIP 🚧

Work is heavily in progress, nothing to see yet, move along ✋👮👉 Example: check out examples/target_example_usage.py

Idea

The idea is to have a framework on top of PyTorch that can make research experimentation easier:

• Reduce generic boilerplate and allow you to focus on the most important
• Make experiment code more descriptive and short, ideally so small that all can be in one file
• Track projects and experiments
  • Save projects' and associated experiments' properties
  • Save all hyperparameters
  • Save git revision and/or .patch of code that was used for the experiment
  • Save all actions that were used in the experiment process
• Handle training/validation process
  • Fully flexible (user-defined) training and validation
  • Fit for most of the CNN experimentations, and easily extended for other types of NN
  • Pretty and useful feedback
  • Support for metrics (associated with model) and callbacks (associated with training/validation process)
  • Automatically sync metrics with tensorboard for realtime visualization
• Increase productivity with network prototyping
  • Provide easily customizable building blocks for different models (Important: the project is being developed primarily for my own use (I focus on CNNs for 3D volumetric data), so not all features have same priority :) When the project will be stable enough to be used, PRs will be super welcome.)
  • Moonshot idea: plot model architecture for publication?
• Moonshot idea: generate reports of experiments automatically?
• Moonshot idea: have GUI, maybe as a jupyter notebook addon or a separate web-based interface?

---

Hierarchy overview

• Projects have Experiments
• Each Experiment should track everything that you did to your Model (a log: each event start and finish)
• Model is a combination of pytorch network and functions defining training, validation and testing
• You can train (and validate) with TrainingStrategy, for example SupervisedTraining
• Strategy combines Dataset, Model, PyTorch optimizer, Callbacks (like scheduler, checkpointer etc.) and model's metrics