Fix #1797 and #1798 #1804
Fix #1797 and #1798 #1804
Conversation
| @@ -3663,6 +3663,32 @@ def _combine_zeroshot_topics(self, | |||
| self.__dict__.clear() | |||
| self.__dict__.update(merged_model.__dict__) | |||
There was a problem hiding this comment.
I spoke too soon! merged_model.__dict__ replaces self.outliers with 0, so the outliers are never repositioned a couple of lines down.
There was a problem hiding this comment.
Maybe it's just best to set outliers = self._outliers just before merging on line 3627, then using this on line 3668 instead?
Will also have to set self._outliers = 1 inside the if statement
There was a problem hiding this comment.
I tested this out and it looks good to me
There was a problem hiding this comment.
Awesome, thanks for checking this out! I just pushed a commit which should finally resolve this issue. If you could do a final check, I can merge this PR.
|
@timdhu Thanks for the help fixing this issue. Your dedication to finding the underlying cause as well as testing/suggesting the needed changes are greatly appreciated! |
A fix for #1797 and #1798. As mentioned in these links, there were 2 issues:
.merge_modelsThe embeddings should now be ordered correctly and the outlier topic is not at the 0th position when using zero-shot topic modeling. This should make it possible to use zero-shot topic modeling with outlier reduction and other techniques.