The report is in bmw_analysis.pdf. The main source file is the
notebook bmw_analysis.ipynb.
The dataset can be downloaded using
mkdir datasets
wget -O datasets/bmw.csv https://raw.githubusercontent.com/datacamp/careerhub-data/master/BMW%20Used%20Car%20Sales/bmw.csv
Dependencies can be installed using (optionally making a new conda environment first)
conda install numpy pandas matplotlib seaborn scikit-learn notebook
The project also depends on the local multi_model.py module.
For making the diagrams (optional), the draw_diagrams.py file should
be in the same directory as the notebook file, and graphviz is
needed. To install in e.g. Ubuntu use
pip install graphviz
sudo apt install graphviz # for Ubuntu
The flask and flask-restful packages are needed to run the webserver, in addition to the dependencies mentioned above
conda install flask flask-restful
Before running the server, the notebook needs to be run, to output the
bmw_linreg_model.pckl file which is needed by the webserver. The
server can then by run using
python flask_app.py
In production the server should be run through a WSGI layer, using
e.g. gunicorn (installable as pip install gunicorn) as
gunicorn flask_app:app
An example front end is implemented in the predict_price.html and
bmw_fetcher.js files. The front end should work should work by
simply opening the html file in a browser, while the javascript script
is in the same directory. The response after the first query might
take around 10 seconds to load, but subsequent requests should be
faster.
The bmw_fetcher.js has the URL for a version
of the flask app deployed to Heroku hardcoded into it, as well as the
feature ranges from the feature_ranges.json output file from the
notebook. If you want to test against a locally run server, or retrain
the model with new features, the relevant variables in
bmw_fetcher.js should be adjusted correspondingly.
The slides are based on the source file slides.org. This file is in
Emacs org-mode format, and should be exported to html using the
ox-reveal exporter. The figures
can be generated by running the bmw_analysis.ipynb notebook.