snorkel
snorkel copied to clipboard
LabelModel with LFs dependencies
This pertains to Issue #1462 and #1513.
As of v0.9.5, LabelModel training assumes that LFs are conditionally independent given Y
.
Are there any plans to update the loss function to reflect the algorithm described in multi-task weak supervision paper?
![image](https://user-images.githubusercontent.com/4229782/65232240-6658ef00-db03-11e9-9381-56c5329ba4d5.png)
Hi @ajratner, @paroma,
A friendly ping to understand if there are any plans to drop the assumption of conditional independence between LFs while training the label model.
Hi @chaturv3di . The repo is currently in maintenance mode, so we are not currently planning to update the label model with this feature.
@paroma Wait, maintenance mode? You're no longer developing Snorkel?
I hope that @fredsala meant Snorkel Metal repository, but I would also like to hear a clarification on that. Does it mean that Snorkel is not going to evolve in the near future?
Going to leave this open as a feature request to consider it for future versions of Snorkel! Maintenance mode was in reference to the code associated with the multi-task weak supervision paper. Snorkel (the main repo here) is not in maintenance mode, though the core team is busy working on snorkel.ai right now!
I hope that @fredsala meant Snorkel Metal repository, but I would also like to hear a clarification on that. Does it mean that Snorkel is not going to evolve in the near future?
I believe what she meant is that at the moment the Snorkel team is busy on their product but they will make additions to Snorkel in the future.
Ed: oh, he said that :)