Improved MalenoV

The original MalenoV data, source code and instructions can be found at: https://github.com/bolgebrygg/MalenoV

The souce code was developed on Python3. First, ensure that GPU resources are available on the server. Second, run pip install -r requirements.txt to install all necessary packages. The project directory is structured as follows:

project
│   README.md
│   requirements.txt
|   train_demo.py
│   predict_demo.py
└───data
    │   F3_entire.segy
    │   *.pts

Dataset

All the data files should be placed under data directory, which are not included in this repository. The user can download example files can at: https://goo.gl/wb145Z Specifically, the .segy is the seismic file, and .pts are the labels.

Training models

To run the training job, run python train_demo.py. Inside train_demo.py, one can specify input parameters, and customize the CNN model. Speficially, on the bottom part of the code block (~line 1555), one can find filenames, which specifies the relative path to the segy data; cube_incr specifies the radius of the cubes.

In train_dict, files specifies the list of files that are labeled data points (please refer to data provided from https://github.com/bolgebrygg/MalenoV for labeled data format); num_tot_iterations specifies the number of times we draw samples from the seismic cube (typically leave it to 1); epochs specifies the number of epochs for training; num_train_ex specifies the total number of training and validation data (a training/validation split of 0.2 is used but can be easily changed); batch_size specifies the size of the mini-batch for each iteration; data_augmentation is set to False since it is not implemented at this moment; save_model specifies if we'd like to save the keras model into .h5 format; save_location specifies the path to the saved model if we set save_model to True; test_size specifies the number of test data; sample_step specifies the step size used in sparse sampling (set to 1 if one wish to do continous sampling).

Predicting models

To run the predicting job, run python predict_demo.py. This file is very similar to train_demo.py; we will eventually merge the two files in the future. Again, on the bottom of the code block, one can find pred_dict. In this variable, keras_model specifies the relative path to the saved .h5 model; section_edge specifies the four coordinates of the 2D seismic section that is to be classified/visualized; plot_type, plot_ref and cord_syst are additional parameters that contrain the 2D slice which is to be examined, and they have to be consistent with section_edge; num_class and sample_step need to be consistent with the parameters defined during the model training; save_pred and save_location specify whether or not, and where to save the predictions.

Notes

Only train_dict is used during training, and only pred_dict is used in the prediction process. However, be sure to select the correct mode in output_dict1 before running train_demo.py or predict_demo.py.