In this project I use TensorFlow and a recurrent neural network to produce a model that can classify two types of earthquakes. The data is taken from the UCR Time Series Classification Archive (Yanping Chen, Eamonn Keogh, Bing Hu, Nurjahan Begum, Anthony Bagnall, Abdullah Mueen and Gustavo Batista (2015)). This work was inspired by Rob Romijners's blog post on LSTM time-series classification. I credit Understanding LSTM Networks that provided an excellent introduction to RNNs and how an LSTM works. I also credit the detailed blog posts Recurrent Neural Networks in Tensorflow I, Recurrent Neural Networks in Tensorflow II, and the links there in for providing a good background of how to create an RNN using TensorFlow.
This project is a work in progress. If you find it interesting and would like to contribute I welcome input.
Anaconda can be downloaded from here. The notebooks in this repo use Python 3.
Create a conda environment and all required packages, less TensorFlow:
conda create --name tf_python3 python=3 pandas numpy matplotlib jupyter natsort
Enter conda environment:
source activate tf_python3
Install TensorFlow from here:
conda install -c jjhelmus tensorflow=0.10.0rc0
Run the following pip command to install all required packages, except TensorFlow:
pip3 install pandas numpy matplotlib jupyter natsort
Then, install TensorFlow from here:
pip3 install -i https://pypi.anaconda.org/jjhelmus/simple tensorflow
All TensorFlow models are saved in directories with the following format: tmp/logs/run_?? where the question marks refer to the model run number. To start tensorboard, execute the following command:
tensorboard --logdir=tmp/logs/run_??
This project was originally presented at the San Francisco Python Meetup on November 9th, 2016. The slides associated
with that talk are found in the Jupyter notebook Classifying Earthquakes slides.ipynb. A slightly different version
of that talk was presented at Udacity on January 7th, 2017 and those slides are in the Jupyter notebook
Classifying Earthquakes slides - Udacity.ipynb.
If you download one of these notebooks and would like to compile the slides as html, run the following command:
jupyter nbconvert Classifying\ Earthquakes\ slides.ipynb --to slides --post serve