Pinky is a simple deep learning toolbox for earthquake localization and detection, for an example study have a look at our application of a CNN with 3 convolutional layers to a West Bohemia earthquake swarm.

To get started, have a look at our small synthetic example generator in the 'example' subdirectory.

• pyrocko
• tensorflow
• scikit-optimize (optional for hyperparameter optimization)

You can use pip to install dependencies:

pip install -r requirements.txt

Note that this will install tensorflow without GPU support. Checkout the tensorflow documentation to install with GPU support

You need to provide pinky with a configuration file.

pinky --config <config_filename>

A good starting point to see if data is properly loaded an preprocessed is to have look at a couple of examples.

pinky --config <config_filename> --show-data 9

This will generate a figure with 9 panels show the first 9 preprocessed data labels and features from your dataset.

To start training:

pinky --config <config_filename> --train

You can dump your examples to TFRecordDatasets to accelerate io operations:

pinky --config <config_filename> --write <new_config_filename>

and use the newly created config file to run --train

Invoke pinky with --debug to enable keep track of weight matrices in tensorboard.