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

Nov 20: new changes in Gym and `doodleverse-utils` version 0.0.12 #106

Closed dbuscombe-usgs closed 2 years ago

dbuscombe-usgs commented 2 years ago

There was an issue with the way MNDWI and NDWI datasets were being made, detailed here https://github.com/Doodleverse/doodleverse_utils/issues/14 and https://github.com/Doodleverse/doodleverse_utils/commit/6c8e1adc5443c6ed567f6988580542f1c2e0a6d0

now, all datasets are made without padding, so things are consistent across the board. this doesnt affect 'regular' models using ND data, only MNDWI and NDWI dataset creation (in utils/preprocess_data.py)

so, pip install -U doodleverse-utils

(should install version 0.0.12, https://pypi.org/manage/project/doodleverse-utils/release/0.0.12/)

I also did some general housekeeping in Gym, and added a new function batch_train_models.py which is handy for those who like to experiment a lot - it now prompts you for multiple datasets and corresponding config files, and trains multiple models sequentially

make_nd_datasets.py now uses image and label resizing and rescaling codes from doodleverse-utils

see https://github.com/Doodleverse/segmentation_gym/commit/de004f38508d47fbfd74e16a68436663033262af

so, git pull