• Search by Module
• Search by Words
• Search Projects

• Java
• Python
• JavaScript
• TypeScript
• C++
• Scala
• Blog

Project: skeltorch (GitHub Link)

• skeltorch-master
  • examples
    • mnist_classifier
      • config.schema.json
      • experiments
        • .gitkeep
      • data
        • .gitkeep
      • mnist_classifier
        • data.py
        • model.py
        • __main__.py
        • __init__.py
        • runner.py
      • README.md
      • requirements.txt
      • config.default.json
    • glow
      • config.schema.json
      • experiments
        • .gitkeep
      • data
        • .gitkeep
      • README.md
      • requirements.txt
      • config.default.json
      • glow
        • data.py
        • model.py
        • __main__.py
        • __init__.py
        • runner.py
  • LICENSE
  • setup.py
  • README.md
  • skeltorch
    • data.py
    • __main__.py
    • experiment.py
    • __init__.py
    • configuration.py
    • runner.py
    • execution.py
  • requirements.txt
  • .gitignore
  • docs
    • main
      • first-steps.md
      • introduction.md
    • Makefile
    • make.bat
    • examples
      • glow.md
      • mnist.md
    • api
      • skeltorch.rst
      • execution.rst
      • experiment.rst
      • runner.rst
      • data.rst
      • configuration.rst
    • tutorials
      • creating-pipelines.md
      • file-structure-experiment.md
      • dealing-configuration-files.md
      • running-pipelines.md
    • requirements.txt
    • index.rst
    • conf.py

Skeltorch: light-weight framework for PyTorch projects

What is Skeltorch?

Skeltorch is a light-weight framework that helps researchers to prototype faster using PyTorch. To do so, Skeltorch provides developers with a set of predefined pipelines to organize projects and train/test their models.

Skeltorch is an experiment-based framework. What that means is that every possible variation of your model will be represented by a different experiment. Every experiment is uniquely identified by its name and contains:

• A set of immutable configuration parameters, specified during its creation.
• A copy of the data object, also created during the creation of the experiment.
• The checkpoints of the model associated with the experiment.
• A set of TensorBoard files with a graphical evolution of the losses and other data that may be logged.
• A textual log of the actions performed on the experiment.

Features

• Easy creation and loading of experiments.
• Automatic restoration of interrupted training.
• Readable JSON configuration files with the option to validate them using a schema.
• Visual logging using TensorBoard.
• Automatic logging using the native Python logging package.
• Automatic handling of random seeds, specified during the creation of an experiment.
• Easy implementation of custom pipelines.
• (NEW) Automatic handling of multi-GPU training.

Installing Skeltorch

Use pip to install Skeltorch in your virtual environment:

pip install skeltorch

Where should I start?

Skeltorch has been designed to be easy to use. We provide you with a lot of material to take your first steps with the framework:

1. Start by reading our first steps tutorial, where we give you a high-level overview of how to organize a project.

2. Take a look to one of our examples. If you are totally new to the framework, you might want to start with our MNIST Classifier example.

3. Read our tutorials to know everything you need to know about Skeltorch and how to customize default behavior.

4. For a deep understanding of the framework, we recommend you to take a look to our API Documentation.

Contributing

You are invited to submit your pull requests with new features or bug corrections.