v1.1.2: Dockerfile bugfix

v1.1.2

Bugfix to Dockerfile: need to install ncurses, otherwise can get a error while loading shared libraries: libncurses.so.6 error.

v1.1.1 hotfix to DataGenerator.py

v1.1.1 hotfix

There was a bug in DataGenerator.FileDataGenerator which failed to accurately identify chip IDs from image filenames that ended in '[chip_id]_image.tif' rather than '[chip_id].tif' - this has been corrected.

v1.1.0 adding file streaming

v1.1.0 enabling streaming of files

This version adds the ability to read directly from files without generating massive numpy arrays of image data. Changes were made in:

• DataGenerator.py: Adding a new generator, FileDataGenerator, which streams data from RGB 8-bit .tifs.
• bin/train_model.py: Adding and adjusting arguments to enable read-in directly from files, should still be compatible with v1.0.0 and v1.0.1 commands.
• bin/make_rgbs.py: Since the massive numpy arrays can be ungainly and challenging for users to generate, some may wish to skip that step. make_rgbs.py essentially does the first half of what make_np_arrays.py does, but skips the messy array conversion part.

Model training has been run using RGB 8-bit imagery instead of numpy arrays and the same results were obtained.

Happy modeling!

Patch v1.0.1

Patch v1.0.1

updates:

• Bugfixes to inference: test images previously weren't scaled the same way as training and validation images were.
• Added simplification to polygons in bin/make_predictions.py to reduce csv file size and speed eval.
• Tuned hyperparameters in bin/train_model.py to match previous runs.

v1.0.0 Functional release!

v1.0.0 is functional in terms of both scripts and package functionality as well as the docker container.