This is a Deep Learning Class Project carried at ESPCI under the supervision of Professor Alexandre Allauzen.
Team members: Matthieu Dagommer, Paul Boulgakoff, Godefroy Bichon, Germain L'Hostis.
We propose two solutions to the deep learning challenge "LANL earthquake prediction". See more information on the webpage : https://www.kaggle.com/c/LANL-Earthquake-Prediction. If you are comfortable with French language, feel free to read our report !
You can directly access the notebooks in Kaggle via these two links: CNN 1D: https://www.kaggle.com/code/matthieudagommer/earthquake-prediction-cnn1d. LSTM: https://www.kaggle.com/code/matthieudagommer/earthquake-prediction-lstm.
You can also run the code locally by cloning this repository, either the notebook or the modular Python version.
To download the repo via git, type:
git clone https://github.com/MatDagommer/earthquake-prediction.git
Check that you are able to launch python from your Git Bash shell.
python --version
If that's not the case, this blog explains how to set it up : https://prishitakapoor2.medium.com/configuring-git-bash-to-run-python-for-windows-a624aa4ae2c5. Check that you are using a virtual environment with the following versions:
Python : 3.10.4 Torch : 1.11
To retrieve the data, go to https://www.kaggle.com/competitions/LANL-Earthquake-Prediction/data and click "Download". You will have to unzip the content in the same folder as the scripts of this project.
In the Git Bash, you can run the program with default parameters with the following command. You need to choose a model: 0 stands for 1D CNN and 1 for LSTM.
python main.py -model 0
There are several parameters you can play with to customize the training. You can get the list of these arguments by typing:
python main.py --help
For instance, if you want to train a 1D CNN with 1000 epochs, a batch size of 300 and a learning rate of 0.001, you can type:
python main.py -model 0 -ep 1000 -bs 300 -lr 0.0001
The program will eventually ask you to name the model you have just trained so it can save its features in Models/<model_name>. A Matplotlib window should pop up with both the learning curve of your training and the distribution of error from the validation set at the best epoch.
