For my NER model I use Weights & Biases sweeps for hyperparameter search. I do a grid search with about 100 runs and there are some really meaningful graphs. However, I can't figure out how to create a graph that shows about the best 10 runs in terms of f-score. Does anyone know how to do this?
Asked
Active
Viewed 116 times
0
-
Hey, can you share your sweep page? – ayush thakur Jul 21 '22 at 08:16
-
Here it is: https://wandb.ai/benelanger/jr-skilldetection-service?workspace=user- – EustassX Jul 21 '22 at 08:31
-
I have set the access back to private because the question was answered. – EustassX Jul 21 '22 at 10:16
1 Answers
3
In the sweep view, you can filter runs by certain criteria by clicking this button:
There, you can add a filter to only show runs with an f1 score, or an accuracy or whatever metric you have logged higher than a certain value:
Of course, this won't filter for the 10 best runs, but for all runs with an accuracy of 0.9 and higher (example in picture).
chefhose
- 2,399
- 1
- 21
- 32
-
I tried that too, but I didn't find "accuracy". But your solution works for me with "score" (which I think is accuracy) and "ents_f". So thanks a lot for your help :) But I wonder, can you actually filter by "accuracy"? – EustassX Jul 21 '22 at 08:58
-
1You should find all values you have logged in your runs. If you logged the f1-score as 'score', you can filter by score. – chefhose Jul 21 '22 at 09:04