Tips for Stan installation for Windows Users
This recitation was written by Julian Wagner and updated by Matteo Guareschi
Here are two big issues I encountered when trying to install cmdstanpy on my Windows machines.
A space in the username kills the install
Windows is terrible about keeping the user from introducing a space into the username, e.g. Julian Wagner. A space in the username leads to a space in the user folder, e.g. /c/Users/Julian Wagner/; I myself ended up with a space in the username for one of my Windows machines (curse you Windows for using a space!!), and it killed the install. I tried to make a workaround for this, but was not able to solve it. If you have a space in your user folder (you can check in
File explorer > Local disk (C:) > Users and look at the folder there), I suggest you make a new user login for your machine that has no space. If you make this new user an admin, you will be able to access the files from your existing login on the new login. To make a new user login, you can do Settings > Accounts > Other Users > Add someone else to this PC > I don't have this person's sign-in information > Add a user without a Microsoft account (in Windows 11:
Settings > Accounts > Family & other users > Add other user > I don't have this person's sign-in information > Add a user without a Microsoft account) . Now you can enter a new username and the key is to pick a new username without a space!!, e.g. jwagner. With that done, you can click on the new user and do Change account type > Administrator if you want to access files from your existing user with the new user. From the start menu, you can click on the little icon that looks
like a person and click on the new username to switch to the new user login. It will take a bit to get set up for logging in. If you did a local install of Git Bash/Anaconda, then you will need to redo the setup from bebi103a and bebi103b to do installs for the new user login. As irksome as it is, I am not aware of how to succeed with the install with a User folder with a space in it.
Two ways to install Stan
Using Conda
Possibly the easiest way to install Stan is to use the package from conda-forge. Open a Powershell Prompt from Anaconda Navigator and run the command:
conda install -c conda-forge cmdstan
cmdstanpy.set_cmdstan_path(os.path(r"C:\Users\USER\anaconda3\Library\bin\cmdstan")
USER with your username. Now rerun the test code and, hopefully, it will compile and show the plot. If not, it might be time to uninstall cmdstan and cmdstanpy and try with the second method:conda remove cmdstan pip uninstall cmdstanpy pip install cmdstanpyset_cmdstan_path command. Press the Start button on your keyboard (or on the bar at the bottom of the screen) and type “Advanced System Settings”, in the Advanced tab you will see Environment Variables… Open it and click on New… Set variable name to CMDSTAN and the value to: C:\Users\USER\anaconda3\Library\bin\cmdstan, again replacing USER with your
username.Using Git Bash
So far in the course, we have often used a terminal launched from within the Anaconda Navigator to do installs of packages. When attempting to install cmdstanpy on Windows (on two separate machines) from such a terminal I ran into this error message after running python -c "import cmdstanpy; cmdstanpy.install_cmdstan()"
Command "make build" failed
'cut' is not recognized as an internal or external command,
operable program or batch file.
'cut' is not recognized as an internal or external command,
operable program or batch file.
INFO: Could not find files for the given pattern(s).
The syntax of the command is incorrect.
mingw32-make: *** [bin/diagnose.exe] Error 1
which seems pretty cryptic and frustrating! As far as I can tell, this error means that the installation requires a Unix command called cut that is not available in the terminal launched through the Anaconda Navigator. Luckily, you should be able to get around this problem by instead running the install through Git Bash. I am assuming that you have installed Git for Windows as recommended by Justin in
bebi103a, which gets you access to the Git Bash app. When I initially installed Git Bash, commands like conda and jupyter lab and python did not work by default due to path issues. To fix this, from in the Git Bash terminal, I ran
/c/Users/jwagner/anaconda3/Scripts/conda init bash
If you have a global install of Anaconda instead of a local, the command will instead be something like
/c/ProgramData/Anaconda3/Scripts/conda init bash
In general, you will need to find the absolute path to the folder where Anaconda is installed, and the conda executable will be in the /Scripts/ folder.
You can check that the conda command works by running conda list in the Git Bash terminal which will print out all the packages you have installed in your conda environment. Make sure you have the C++ toolchain installed (you can skip running this if you already ran it previously):
conda install libpython m2w64-toolchain -c msys2
You can now launch JupyterLab from within this Git Bash terminal by typing jupyter lab. From in this Jupyter session, click New Launcher and under Console click Python 3
In this python console, run import cmdstanpy and then cmdstanpy.install_cmdstan(). This may take a few minutes.
NOTE: this command may look like it hangs (never finished running). Just let it run for a while, and then you can launch a new Jupyter notebook and try to run Justin’s test script to see if the install worked.
The install took a couple of minutes to run for me, but this approach worked on both the Windows machines I tried it on.
Now try the code below and, if it works, try launching JupyterLab as you normally do (through Anaconda Navigator) and running it again. You may now encounter an error. If that is the case, the best suggestion is to start always running JupyterLab from Git Bash.
Justin’s test code
To test your installation of CmdStanPy, you can run this test code on the famous “schools” hierarchical model.
[1]:
import numpy as np
import cmdstanpy
import arviz as az
import bokeh.plotting
import bokeh.io
import bokeh.resources
bokeh.io.output_notebook(bokeh.resources.INLINE)
schools_data = {
"J": 8,
"y": [28, 8, -3, 7, -1, 1, 18, 12],
"sigma": [15, 10, 16, 11, 9, 11, 10, 18],
}
schools_code = """
data {
int<lower=0> J; // number of schools
vector[J] y; // estimated treatment effects
vector<lower=0>[J] sigma; // s.e. of effect estimates
}
parameters {
real mu;
real<lower=0> tau;
vector[J] eta;
}
transformed parameters {
vector[J] theta = mu + tau * eta;
}
model {
eta ~ normal(0, 1);
y ~ normal(theta, sigma);
}
"""
with open("schools_code.stan", "w") as f:
f.write(schools_code)
sm = cmdstanpy.CmdStanModel(stan_file="schools_code.stan")
samples = sm.sample(data=schools_data, output_dir="./")
samples = az.from_cmdstanpy(samples)
# Make a plot of samples
p = bokeh.plotting.figure(
frame_height=250, frame_width=250, x_axis_label="μ", y_axis_label="τ"
)
p.circle(
np.ravel(samples.posterior["mu"]),
np.ravel(samples.posterior["tau"]),
alpha=0.1
)
bokeh.io.show(p)
INFO:cmdstanpy:compiling stan file /Users/bois/Dropbox/git/bebi103_course/2022/b/content/recitations/04/schools_code.stan to exe file /Users/bois/Dropbox/git/bebi103_course/2022/b/content/recitations/04/schools_code
INFO:cmdstanpy:compiled model executable: /Users/bois/Dropbox/git/bebi103_course/2022/b/content/recitations/04/schools_code
INFO:cmdstanpy:CmdStan start procesing
INFO:cmdstanpy:CmdStan done processing.
[2]:
%load_ext watermark
%watermark -v -p numpy,bokeh,cmdstanpy,arviz,jupyterlab
Python implementation: CPython
Python version : 3.9.7
IPython version : 7.29.0
numpy : 1.20.3
bokeh : 2.3.3
cmdstanpy : 1.0.0
arviz : 0.11.4
jupyterlab: 3.2.1