search

Doodleverse / segmentation_gym

A neural gym for training deep learning models to carry out geoscientific image segmentation. Works best with labels generated using https://github.com/Doodleverse/dash_doodler
MIT License
45 stars 11 forks source link

New makedatasets #90

Closed dbuscombe-usgs closed 2 years ago

dbuscombe-usgs commented 2 years ago

Updates make_nd_datasets, train_model, and do_seg_images_in_folder to be compatible with NCLASSES>=2 in the config file

NCLASSES=2 for binary problems. This is a more intuitive solution for users making config files, since 'class' and 'noclass' are commonly considered 2 classes. It also makes the codebase smaller, because there are no checks for NCLASSES=1, and associated modifications to internal functions and workflows

Other minor bug fixes and code cleanups incorporated

Thoroughly tested with NCLASSES=2 and NCLASSES>2 on a handful of different datasets

These changes also test the latest doodleverse-utils changes to accommodate weighted class functions, float32 precision, etc. There will be a new doodleverse-utils release associated with this PR

Changes to do_seg also incorporate additional implementation strategies in 'seg_images_in_folder', reflected in the config arguments DO_CRF, OTSU_THRESHOLD, and TESTTIMEAUG. these need to be documented in the wiki

dbuscombe-usgs commented 2 years ago

Requires v0.0.10 https://pypi.org/project/doodleverse-utils/0.0.10/