Closed mvacaporale closed 3 years ago
Do we know why this issue did not show up for the 85 and 90 % sparse models? Is there anything to be weary of elsewhere in the code that the changed default will affect?
Good question: It didn't show up for those models because there were trained after sparsify_all_embeddings
was implemented. Thus, they had that param in their saved model configs (set to False), so they were loaded without issue.
The PR is for https://ntaresearch.atlassian.net/browse/RES-2309
The 80% sparse trifecta model was not overridden or lost. A small bug in the code simply prevented it from being loaded properly.