Welcome to the TorchTNT documentation!

TNT is a library for PyTorch training tools and utilities. It has two main components, which are the top-level modules of the repo:

1. torchtnt.framework: contains a lightweight training framework to simplify maintaining training, evaluation, and prediction loops.
2. torchtnt.utils: contains a grab-bag of various independent, training-related utilities, including data related abstractions and wrappers around different publishers to simplify logging metrics.

Installation

TNT can be installed with pip. To do so, run:

pip install torchtnt

If you run into issues, make sure that Pytorch is installed first.

You can also install the latest version from master. Just run:

pip install git+https://github.com/pytorch/tnt.git@master

To update to the latest version from master:

pip install --upgrade git+https://github.com/pytorch/tnt.git@master

Documentation

Examples

• Examples

Core Concepts

• Distributed training
• Checkpointing

Framework

• Unit
• AutoUnit
• Train
• Evaluate
• Predict
• Fit
• State
• Callbacks

Utils

• Utils
  • Data Utils
  • Device Utils
  • Distributed Utils
  • Early Stop Checker
  • Environment Utils
  • Flops Utils
  • Filesystem Spec Utils
  • Logger Utils
  • Memory Utils
  • Module Summary Utils
  • OOM Utils
  • Optimizer Utils
  • Precision Utils
  • Prepare Module Utils
  • Progress Utils
  • Rank Zero Log Utils
  • Stateful
  • Test Utils
  • Timer Utils
  • TQDM Utils
  • Version Utils
  • Misc Utils