Closed pli888 closed 3 years ago
Snorkel would be quite cumbersome for training image data. A user would have to write labelling functions rather than click on objects (see tutorial link below): https://www.snorkel.org/use-cases/visual-relation-tutorial
The BioImage Model Zoo, which is a database of pretrained Deep Learning models, is closer to what I have in mind to accompany the Digital Pathology Thematic Series. More details at the link below: https://bioimage.io/docs/#/
Do we still need this ticket then? @ChrisArmit you can create a backlog ticket for BioImage Model Zoo if you think thats a useful tool to consider integrating then we can close this ticket as we have no intention to implement snorkel.
User Story
@ScottBGI says:
Don't know if you've come across snorkel, but if one of the main downstream use cases for a lot of the imaging (medical and camera trap) and chemoinformatics and modelling data is for machine learning, it seems to be a new method for labeling functions (LFs) to label the training examples:
https://www.snorkel.org/features/
Is this something we want to adopt for obvious training datasets? Either as a curational step, or something we/the reviewers need to encourage the authors to do during the review/revision process?
The tool seems to be open source (Apache licensed):
https://github.com/snorkel-team/snorkel
There are some examples in this pre-print, including medical images:
https://arxiv.org/abs/1711.10160
The imaging series Chris A has been proposing potentially had this angle, so it could be a nice area to try to get an exemplar/showcase paper to start with.