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

Remove the nested file structure in the test dataset and update docs if necessary #98

Closed dbuscombe-usgs closed 1 year ago

dbuscombe-usgs commented 1 year ago

As pointed out in https://github.com/Doodleverse/segmentation_gym/issues/92, the test dataset is structured badly for new makedatasets workflows.

This is a reminder for me to remove the nested directories in 'images' and 'labels' in the test dataset, and reupload.

dbuscombe-usgs commented 1 year ago

I remade all test datasets, and remade all 3 models with the latest version of Gym (as of today). There is a new zenodo release: https://zenodo.org/record/7232051#.Y1HDKnbMIuU 🎉

This is zenodo release version 3.0

We should make a repo release to version this code.

dbuscombe-usgs commented 1 year ago

I updated the wiki

The new models are better. At least, the two residual UNet models are better

@ebgoldstein and I suspect that the learning rate may be too low for the "vanilla" unet model .... it is there for demonstration purposes only

closing this issue