Juneberry improves the experience of machine learning experimentation by providing a framework for automating the training, evaluation, and comparison of multiple models against multiple datasets, thereby reducing errors and improving reproducibility.
This README describes how to use the Juneberry framework to execute machine learning tasks. Juneberry follows a (mostly) declarative programming model composed of sets of config files (dataset, model, and experiment configurations) and Python plugins for features such as model construction and transformation.
The Getting Started documentation explains how to install Juneberry. It also includes a simple test command you can use to verify the installation.
The Workspace and Experiment Overview documentation contains information about the structure of the Juneberry workspace and how to organize experiments.
The Juneberry Basic Tutorial describes how to create a model, train the model, and run an experiment.
During normal use of Juneberry, you may encounter warning messages. The Known Warnings in Juneberry documentation contains information about known warning messages and what (if anything) should be done about them.
The vignettes directory contains detailed walkthroughs of various Juneberry tasks. The vignettes provide helpful examples of how to construct various Juneberry configuration files, including datasets, models, and experiments. A good start is Replicating a Classic Machine Learning Result with Juneberry.
When this environment variable is set, the torch.cuda.memory_summary() will appear during training after
the model is loaded and again after the specified period of epochs starting at epoch one.
In other words, if JUNEBERRY_CUDA_MEMORY_SUMMARY_PERIOD is set to 10, the memory summary will be emitted
after model load and again after epochs 1, 11, 21, 31, etc.
Copyright 2021 Carnegie Mellon University. See LICENSE.txt file for license terms.