bayesian-stats-examples
Python versions of things taught in the Bayesian Statistics courses on Coursera.
- Bayesian Statistics: From Concept to Data Analysis - taught by Herbert Lee from the University of California, Santa Cruz. Python notebooks for analogs of things taught in this class are in the folder
concept-to-data-analysis. - Bayesian Statistics: Techniques and Models - taught by Matthew Heiner, also from University of California, Santa Cruz. Python notebooks for models taught in this class are in the folder
techniques-and-models.
I decided to explore the subject further using the book Bayesian Analysis with Python, 2nd Edition (BAP) by Osvaldo Martin, one of the contributors to the PyMC3 project. The book claims to be introductory, but goes further than the courses above. The bayesian-analysis-exercises folder contain my answers to the exercises in the chapters. You will need to git clone the book's code repository at the top level of this project (git clone https://github.com/aloctavodia/BAP.git) to get access to the data files used by the book, and referenced from my notebooks.
Environment
Initially I just created a conda environment making sure I had the following packages installed.
- numpy
- scipy (for scipy.stats)
- statsmodels
- PyMC3
- matplotlib
Later when I started following the BAP book, I needed some additional packages such as arviz, graphviz, etc., so I just ran the BAP YAML file in the same environment.
conda env create --file bap.yml
Later I encountered issues with Theano not being able to compile the PyMC3 code that I couldn't figure out how to fix, so I created a completely new environment from a slightly modified clone of bap.yml, available as environment.yml. The modifications were to move the versions of numpy and pandas up for xarrow to be installed, which is needed by arviz.
conda env create --file environment.yml
In addition, when importing pymc3 and friends into the notebook, it complained about mkl and mkl-service so these needed to be installed as well.
conda install mkl
conda install mkl-service
Optionally, if your environment does not do this automatically, in addition to running Jupyter notebooks inside your newly created environment, you might need to create a Jupyter notebook kernel. To do this, run the following commands.
source activate bap
python -m ipykernel install --user --name bap --display-name "Python (BAP)"
To delete the environment and the kernel after you are done exploring, use the following commands:
source activate bap
jupyter kernelspec uninstall bap
source deactivate
conda remove --name bap --all
General impressions
- Course uses JAGS, PyMC3 has more natural syntax
- statsmodels.OLS is not as flexible compared to R's lm().
Caution
- Meant for understanding how to apply Python libraries to problems solved with R in the course.
- If used to answer quizzes, can lead to disappointment since answers don't always match exactly.