Hierarchical transfer learning for deep learning velocity model building

Prerequisites and installation

Install the latest version of git. Make sure to update git if it is already installed.

Clone this repository through:

git clone https://github.com/CloudyOverhead/velocity-model-building-using-transfer-learning.git

Navigate to the package's directory and run:

pip install .

This should install all required packages, including SeisCL. However, please follow the additional steps for completing the installation of SeisCL in the package's README.md.

Generating the seismic data and training the neural networks

Generating velocity models and the seismic data

Use the following commands to generate each dataset:

python vmbtl --dataset Article1D --generate

python vmbtl --dataset Article2D --generate

python vmbtl --dataset USGS --generate

python vmbtl --dataset Article1DSteep --generate

python vmbtl --dataset Article2DSteep --generate

python vmbtl --dataset Marmousi --generate

You may use the --gpus option to control the quantity of GPUs dedicated to generation and the --plot option to ensure the results are correct.

Training the neural networks

Use the following commands to produce the train the neural networks. Make sure to create the ./logs directory if necessary (mkdir logs). Use export i=0 through export i=15, except export i=2 and export i=9 for each given command to generate the ensemble of 14 networks.

python vmbtl/automated_training.py --dataset Article1D --params Hyperparameters1D --noise --seed "($(($i*3)), $(($i*3+1)), $(($i*3+2)))" --destdir logs/weights_1d/$i --gpus 2

python vmbtl/automated_training.py --dataset Article2D --params Hyperparameters2D --noise --seed "($(($i*3+100)), $(($i*3+101)), $(($i*3+102)))" --restore_from "('$(pwd)/logs/weights_1d/$i/checkpoint_000060', None, None)" --destdir logs/weights_2d/$i --gpus 2

python vmbtl/automated_training.py --dataset Article2D --params Hyperparameters2DNoTL --noise --seed "($(($i*3+200)), $(($i*3+201)), $(($i*3+202)))" --learning_rate 8E-4 --destdir logs/weights_2d_no_tl_8E-4/$i --gpus 2

python vmbtl/automated_training.py --dataset Article2D --params Hyperparameters2DNoTL --noise --seed "($(($i*3+300)), $(($i*3+301)), $(($i*3+302)))" --learning_rate 8E-5 --destdir logs/weights_2d_no_tl_8E-5/$i --gpus 2

python vmbtl/automated_training.py --dataset Article1DSteep --params Hyperparameters1DSteep --seed "($(($i*3+1000)), $(($i*3+1001)), $(($i*3+1002)))" --destdir logs/weights_1d_steep/$i --destdir logs/weights_1d_steep/$i --gpus 2

python vmbtl/automated_training.py --dataset Article2DSteep --params Hyperparameters2DSteep --seed "($(($i*3+1100)), $(($i*3+1101)), $(($i*3+1102)))" --restore_from "('$(pwd)/logs/weights_1d_steep/$i/checkpoint_000060', None, None)" --destdir logs/weights_2d_steep/$i --gpus 2

Reproducing the figures

To generate the figures, use the following command, which will produce the predictions on the test examples as well.

python vmbtl/postprocess

The figures will be available under the ./figures directory.