Training Seisbench models on custom data

Our platform supports training and selection of the best performing Seisbench model on custom dataset. 

Training models

Before you start, make sure that you have an account on the platform. To create an account please contact the EPOS AI developers team.

Please download the repository and follow the steps described in  README to learn how to:

  • Convert your data to Seisbench format
  • Train multiple Seisbench models to select the model, which performs best at picking P and S waves. The code will perform the hyperparameters search to identify the parameters that lead to the best performance.

Monitoring training

The models training process can be monitored online on W&B platform. The metadata of each experiment and metrics of each model are automatically saved in your project space.

We highly recommend to utilise W&B functionality to analyse the experiments results e.g.: 

  • Go to Runs view and sort the run by selected metrics e.g. loss to check which model performs best 
  • Group the models by `model_name` to compare performance of each model type
  • Create a report to save your observations, summarise experiments results and ideas for future work. 

Summarising project results will help you to analyse whether e.g. it is worth to repeat the experiments with a modified search space for the hyperparameters that matches better your data. Check the documentation for more information. 

Models registry

Once you are satisfied with the performance of your model, we suggest to save it in W&B Models registry. For more information check the documentation

Back to top

Related Documents

  • No labels