Masterful: The Model Development Platform for Computer Vision

Intro

Masterful is the development platform for computer vision models. Masterful supports most types of classification, detection, and segmentation.

_images/arch.png

Users define the two inputs to Masterful:

  • Data

  • Model architecture

Masterful processes those inputs through four modules:

  • Regularization, which improves accuracy from the existing information.

  • Semi-supervised learning (SSL), which improves accuracy by learning from the information in raw, unlabeled images.

  • Optimization, which minimizes GPU-hours and wall-clock time of training.

  • Meta-learning, which minimizes developer “guessing and checking” and eliminates black box optimization.

Masterful returns a trained model (e.g. model weights).

Two ways to use Masterful: The CLI Trainer and The Python API

Masterful is available through two interfaces:

  1. Masterful CLI Trainer, a command line tool that takes YAML config files and training data organized in directories, trains a model, and saves the trained model and weights on disk. Internally, the Masterful CLI Trainer uses the Masterful Python API. The Masterful CLI Trainer emphasizes ease-of-use and simplicity.

  2. Masterful Python API, a Python API built on top of Tensorflow 2 offering developers more control.

Design Goals

Masterful is designed to achive three design goals:

  1. Accuracy, achieved via comprehensive regularization and semi-supervised learning (e.g. learning from raw, unlabeled images).

  2. Developer Productivity, achieved via high-speed metalearning. It’s an inefficient use of developer time to guess-and-check hyperparameters or do long runs with a black box optimizer.

  3. Speed, or minimizing GPU-hours and wall clock time, achieved via high-speed metalearning to discover ideal optimization hyperparameters.

Getting Started

Install with pip install --upgrade pip; pip install masterful.

For detailed installation instructions, visit the Installation Tutorial.

Then execute your first training run with Masterful with the Quickstart Tutorial.

Design

To learn about the concepts within the Masterful CV Platform, visit the technical report preprint.

arXiv_badge

If you use Masterful for academic research, you are encouraged to cite:

@article{wookeyhorikert2022masterful,
title={Masterful: A Training Platform for Computer Vision Models},
author={Wookey, Samuel and Ho, Yaoshiang and Rikert, Tom and Lopez, Juan David Gil and Beancur, Juan Manuel Mu{\~n}oz and Cortes, Santiago and Tawil, Ray and Sabin, Aaron and Lynch, Jack and Harper, Travis and others},
journal={arXiv preprint arXiv:2205.10469},
year={2022}
}