Refactored data augmentation, changed loss function, cleaned notebooks and other improvements #226
Conversation
* Notebook implementation of AxonDeepSeg(mini version) * Notebook implementation of AxonDeepSeg(mini version) * Incorpoation of Myelin and Axond dice coefficient and Cleaning of notebook
Updated Requirements.txt for Keras framework
@vs74 I'll do one final pass tomorrow (Thursday), and retest the functionality hands on. If there's only minor issues, I'll open them as new separate issues so we can work on them individually. |
…n the config file
There was a problem hiding this comment.
@vs74 Finished doing a final review. Did a lot of minor changes. Tested all the notebooks manually, and after some modifications they ran fine. The only thing I didn't thoroughly test is retraining an entire model (all epochs), however we trained the default models recently, so I don't suspect that any of the minor changes would impact it significantly (and the training is tested in the unit test for general errors anyways).
Everything looks good, ready to merge. This finalizes the new major release v3.0, which is reflected in the CHANGELOG.
Thanks so much for this major contribution @vs74 !
This major PR aims to handle the improvement in the performance of model as well as an improvised version of data augmentation.
DONE
Implemented data augmentation (Albumentation library) similar to previous version of ADS
Changed from
Cross Entropyas a loss function toDice coefficientto improve model performance as indicated in issue Try Dice instead of cross entropy in optimization #19 .Interpolation changed from
lineartonearest neighbourNotebooks are cleaned and removed irrelevant notebooks as indicated in Should we keep all Jupyter notebooks? #148
Migrated models to OSF storage to prevent bloating of the repository
Fixes #148, Fixes #19, Fixes #241, Fixes #278, Fixes #240 Fixes #273