This repository has been archived by the owner on May 12, 2020. It is now read-only.
Releases: neurologists/ionic
Releases · neurologists/ionic
Alpha 0.4
Changelog
- Fixed an issue that caused incorrect variable initialization
- Added crop and stretch augmentations to increase training set by 7x - 28,000 images if no validation split is used
- IoU calculation now uses metric specified by Kaggle, using tf.metrics functions instead of custom implementation from Kaggle user
- Loss function now uses tf.keras implementation instead of custom implementation from Kaggle user
- Groundwork laid for depth integration - requires more testing
- Learning rate now decays over time as specified in the Tiramisu paper
- Major file structure changes and code refactoring to prepare for hyperparameter optimizer
- General bug fixes
Known issues
- None as of 10/18/2018
Alpha 0.1
Comments
Working release with proper loss and IoU functions, data augmentation with flips, input tiling and other functional updates. Last checked with a score of 0.75 IoU.
Known issues
- Using a validation split of 0.0 causes the accuracy visualization to display 0.0000