Bugfix for PhaseNet label IDs #277
Conversation
seisbench/generate/labeling.py
Outdated
| label: [*model_labels].index(label) for label in labels | ||
| } # label ids for P and S | ||
| if noise_column: | ||
| label_ids["Noise"] = len(model_labels) - sum(list(label_ids.values())) |
There was a problem hiding this comment.
I'm not sure I understand this formula. I would expect it's just [*model_labels].index("n")
There was a problem hiding this comment.
It is quite similar to your idea but I wrote it the other way around and add "N" afterwards
|
Hi @JanisHe , I'm not sure I'd call it a bug but it's definitely good to add this functionality. However, that's why I'd prefer merging it into |
|
Hi Jannes, |
|
Hi @JanisHe , |
- Added another test - Fixed issue with noise label id - Extended documentation
|
I've merged the branch now. Thanks a lot for your contribution! |
Hi Jannes,
when using the original PhaseNet model or diting as base models for transfer learning, the label IDs are wrong. Default labels_ids are {"P": 0, "S": 1, "Noise": 2}, however, PhaseNet and diting have {"Noise": 0, "P": 1, "S": 2}. Therefore, transfer learning was not possible since the ids did not depend on the given labels by the loaded model. I rewrote the method "_colums_to_dict_and_labels" and added an argument model_labels to get the correct dictionary and label IDs.
Cheers,
Janis