Config file description?

[maihao14]

Hi @yetinam , Compared to tutorial training examples, this training.py is more comprehensive and helpful for training new models based on new data. When I try to rewrite the config JSON file, I meet some difficulties in rewriting my demand.

For example, I saw the README.md mention user could use the local datasets in Seisbench format, but when I replace
"data": "SCEDC" to "data": "./seisbench/datasets/ethz" (which is the path of the downloaded ETHZ dataset) it will raise errors.
Moreover, how to use data.filter in JSON to change the complete built-in dataset? E.g., reduce the samples amount, filter magnitude range? Is it possible to work by rewrite the json file?

Another question is could I load pre_trained weights in this training.py? I.E., how to express sbm.PhaseNet.from_pretrained("stead") in json file (or rewrite some lines of training.py)?

Tutorial 03a_training_phasenet.ipynb is a really user-friendly notebook for training. However, it didn't contain the following part, e.g., export the trained model (weights, logging) to a local folder for later reused, generate performance matrix based on validation set, etc. pick-benchmark have these scripts, so I'd like to create a notebook to extend 03a_training_phasenet.ipynb. I would appreciate it if you could offer some advice.

[yetinam]

Hi @maihao14 ,

thanks for your questions. Regarding the documentation of the config parameters, there is some to be found in the docstrings but I admit it's not very complete. Unfortunately, I won't have the time to comprehensively document the parameters in the foreseeable future. I think most parameters should be rather easy to reverse-engineer though. A good starting point for this is here.

Regarding the dataset, it seems that the README is wrong. You'll have to do a code change to accomodate local datasets. The additional functionalities you mention, i.e., data filtering and pretrained weights are not built into the benchmark yet. Both are no major changes, but you'll need to modify the code to incorporate this.

Regarding the tutorial, 03a_training_phasenet.ipynb is intended as a very minimalistic example. It would be great if you wanted to create a more extensive tutorial complementing it.