/pyutility

Primary LanguagePythonMIT LicenseMIT

pyutils

A python/pytorch utility library

License: MIT

News

  • v0.0.2 available. Added new datasets and quantization!
  • v0.0.1 available. Feedbacks are highly welcomed!

Installation

conda install scopex/label/ScopeX::torchonn-pyutils
pip install torchonn-pyutils --no-build-isolation

or install from cloned codes from github if you would like to modify the code

git clone https://github.com/JeremieMelo/pyutility.git
cd pyutility
pip3 install --editable .

Usage

import pyutils

Features

  • Support pytorch training utility and datasets.

TODOs

  • Support lr_scheduler
  • Support trainer

Dependencies

  • Python >= 3.6
  • PyTorch >= 1.8.0
  • Tensorflow >= 2.5.0
  • Others are listed in requirements.txt

Files

File Description
datasets/ Defines different datasets and builder
loss/ Defines different loss functions/criterions
optimizer/ Defines different optimizers
lr_scheduler/ Defines different learning rate schedulers
quant/ Defines different weight/activation quantizers
activation.py Activation functions
compute.py functions related to computing
config.py Hierarchical yaml configuration file parser
distribution_sampler.py Sample from customized distributions
general.py Common helper functions
initializer.py Initialization methods for PyTorch Parameters
loss.py Loss functions for PyTorch model training
quantize.py Quantization functions
torch_train.py Helper functions for torch training
typing.py Defines common types

Contact

Jiaqi Gu (jqgu@utexas.edu)