scdlrtpicker: deep-learning real-time phase picker for SeisComP
scdlrtpicker runs deep-learning phase-picking models (PhaseNet, EQTransformer, EQCCT, or your own) as a native SeisComP P and S picker. A model’s weights and architecture are exported once to a portable file, so this plugin never depends on the framework or language that trained it (PyTorch, TensorFlow, …).
Currently that format is ONNX (Open Neural Network Exchange), a standard, framework-neutral model representation every major training framework can export to; the only runtime requirement is ONNX Runtime, which loads and runs the .onnx file.
It plugs into SeisComP’s existing picker framework, so it works today in:
scautopick: automatic real-time P and S picks on the continuous stream;scolv: interactive re-picking (the repick key) using the same model.
It is a picker interface like the classical AIC / BK / GFZ P pickers or the L2 S picker, usable by any module that supports picker plugins, such as scautopick and scolv.
Check you the full documentation here:
Before running your model in real time with scautopick, I recommend testing it first in scolv to make sure it works correctly. For this, you can use the repicking functionality. However, scolv currently supports only P-phase picking. In the screenshot below, the DL3C (Deep Learning 3-Component) P picker is selected:
If you also want to test the S picker in scolv, you will need this patch. Its documentation is available here, which explains how it works. In the screenshot below, the SDL3C (S Deep Learning 3-Component) S picker is selected:
Also note that although some models can compute uncertainty and polarity information, scolv currently discards these values, so they are not displayed in the interface. You can still see them in the logs for debugging. However, they do work correctly in real-time operation with scautopick.