Calling BFI from Python

Introduction

The BFI R package provides Bayesian Federated Inference (BFI) for Gaussian, logistic, and survival regression models. Since there is no Python package for carrying out the BFI method so far, this vignette describes the usage of the R package BFI in the Python environment.

To move to the Python environment, we need to prepare our operating system. Although, we explain in the following how to prepare it in different systems (‘macOS’, ‘Ubuntu’ and ‘Windows’), we recommend using ‘Google Colab’ to write and execute (the following) Python code, as it can be done in any system through a browser without any preparations.

We illustrate a file-based workflow involving two local centers and a coordinating center. Each local center fits the agreed model and saves its estimation results in an RDS file. The coordinating center reads these files and performs the BFI aggregation without accessing individual-level observations.

Google Colab

First go to Google Colab from https://developers.google.com/colab or https://colab.research.google.com, and then click on New Notebook. Now, skip the next two sections and proceed to the Python Script section.

Python Installation

The steps to install the latest version of Python on ‘macOS’, ‘Ubuntu’ and ‘Windows’ are as follows:

On macOS

Open a terminal window (cmd + space and search ‘Terminal’) and install the package manager ‘Homebrew’ by executing the following command:

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/master/install.sh)"

Install the latest version of Python (Python 3) with the following command:

brew install python@3

Verify the installation by checking the Python version:

python3 --version

It’s assumed that R is also installed and configured on your system. If not, the following can do it:

# Install XCode command-line tools
xcode-select --install

# Install R
brew install --cask r

On Ubuntu

Update package repositories and get latest package information by running the following command in a terminal window:

sudo apt-get update -y

Verify if Python is indeed installed on your system by the following command:

python3 --version

If you already have Python (Python 3) installed on your system, you only need to upgrade it to the latest version as follows:

sudo apt-get upgrade python3

In case you did not have Python installed in the first place, the latest version of Python can be installed using the following command:

sudo apt-get install python3

Verify the installation by checking the Python version:

python3 --version

If you don’t have R installed and configured, just run the following:

# Install R
sudo apt-get install r-base

On Windows

By default, Python is usually not installed on Windows. However, you can check if it exists on the system by running one line of command on the command prompt. Go to Start and enter cmd in the search bar, and then click Command Prompt. Enter the command python --version or python3 --version in the command prompt.

To download Python, go to the official Python download website for Windows: https://www.python.org. Find a stable Python 3 release, and download the executable file for your system from the appropriate link.

After the installer is downloaded, open/run the Python installer to install Python. Select the Add Python 3.x to PATH checkbox, which enables users to launch Python from the command line, and then select Install Now.

Once the installation is complete, you can verify whether the Python installation is successful through the command line. Enter the command python --version or python3 --version in the command prompt.

It’s assumed that R is also installed on your system. If you don’t have R installed, just visit CRAN downloads (https://cran.r-project.org) and get the last version.

Installation of Required Modules

Here the required modules are installed through PIP. PIP (a Python package manager) helps us to install and use various packages/modules in Python programming.

Install the latest version of PIP and required modules by running the following commands on different systems.

On macOS

Install/Upgrade the latest version of PIP by running

# Installing PIP
python3 -m pip install --user --upgrade pip

Install required modules by running the following commands

# Installing the 'pandas' module
pip3 install pandas

# Installing the 'numpy' module
pip3 install numpy

# Installing the 'rpy2' module
pip3 install rpy2

On Ubuntu

To install PIP use the code

# Installing PIP
sudo apt-get install python3-pip

Verify you have successfully installed PIP by running

pip3 -V

Now, install required modules by typing the following commands

# Installing the 'pandas' module
sudo pip3 install pandas

# Installing the 'numpy' module
sudo pip3 install numpy

# Installing the 'rpy2' module
sudo pip3 install rpy2

If the lines above for installation of the modules did not install them, use the following:

# Installing the 'pandas' module
sudo apt-get install python3-pandas

# Installing the 'numpy' module
sudo apt-get install python3-numpy

# Installing the 'rpy2' module
sudo apt-get install python3-rpy2

Finally, use the show command to verify whether the module is now part of your Python packages:

pip3 show pandas
pip3 show numpy
pip3 show rpy2

On Windows

Usually, PIP is automatically installed if you are using Python downloaded from https://www.python.org. If you have followed the previous steps provided in this vignette, then you have PIP installed on your system. To check if PIP is already installed on Windows, you should open the command line again, type pip -V, and press Enter.

If PIP is not installed: First download get-pip.py to a folder on your computer. Then, open a command prompt and navigate to the folder containing the get-pip.py installer. Finally, run the following command:

python get-pip.py

PIP should now be installed successfully. If we receive a ‘file not found’ error, double check the directory path to the file. You can use the dir command to view the entire contents of a directory. More information can be found here: https://pip.pypa.io/en/stable/installation.

Now, install required modules by running the following commands:

# Installing the 'pandas' module
pip3 install pandas

# Installing the 'numpy' module
pip3 install numpy

# Installing the 'rpy2' module
pip3 install rpy2

Note for Windows users: On some systems, rpy2 may encounter difficulties loading R shared libraries, such as stats.dll. If this occurs, check that R_HOME points to the correct R installation. Additional DLL path configuration may be required.

Now, you can start coding in Python using Python’s command-line interpreter or IDLE application. Go to Start and enter python in the search bar. You can see Python 3.x and IDLE which can be used for coding. Open one of them and follow the following steps.

Python Script

It is assumed here that the required packages has been properly installed, which is the case in Google Colab. Now move on to work with the rpy2 package inside a Python script! If you are not in the Python script, first go to the Python Environment by typing the following command in the terminal window (you should skip this line if you are already in Google Colab):

python3

As of now, all the following code should be run in the Python environment. If you want to copy all code at once go to the last section.

Importing `Python` modules and functions

Use the following code to import required Python modules:

# import 'pandas' package
import pandas as pd

# import 'numpy' package
import numpy as np                

# import 'rpy2' package
import rpy2

and using the following code imports the required functions such as importr() and data():

import rpy2.robjects as ro
from rpy2.robjects.packages import importr, data
from rpy2.robjects.vectors import StrVector
from rpy2.robjects import numpy2ri, pandas2ri

Installing `R` packages from CRAN and GitHub

First load the packages preinstalled with R using importr() as follows:

# import R's "utils" package
utils = importr('utils')

# import R's "base" package
base = importr('base')

Then install the R packages stats and BFI from CRAN by typing

utils.chooseCRANmirror(ind=1)  # which selects the first mirror in the list
package_names = ('BFI',)
utils.install_packages(StrVector(package_names))

Now load the installed packages using

stats = importr('stats')
BFI = importr('BFI')

Installing the `BFI` package from GitHub (if necessary)

If you want to install the BFI package from GitHub (instead of CRAN), it can be installed and loaded using the following lines:

utils.chooseCRANmirror(ind=1)
utils.install_packages('devtools')
devtools = importr('devtools')

# Installing the 'BFI' package
devtools.install_github("hassanpazira/BFI", force = True)

# Loading the package
BFI = importr('BFI')

Applying BFI method to the simulated data

Now we generate two datasets independently from Gaussian distribution, and then apply main functions in the BFI package to these datasets:

Data Simulation for two Local Centers

First generate 30 samples randomly from Gaussian distribution with 3 covariates:

# model parameters: 'theta' and 'p' are assumed to be the same for both centers:
theta = np.array([1, 2, 3, 4, 0.75])   # intercept is theta[0], sigma2 is theta[4]
p = 3     # number of regression parameters without intercept

# Center 1:
n1 = 30   # sample size of center 1
X1 = np.random.normal(size=(n1, p)) 
X1.shape   # dimension of X1
X1 = pd.DataFrame(X1, columns=['X1', 'X2', 'X3'])
mu1 = theta[0] + np.dot(X1, np.delete(theta, [0, 4]))  # for gaussian: \eta = \mu 
y1 = np.random.normal(loc=mu1, scale=np.sqrt(theta[4]))
with (ro.default_converter + pandas2ri.converter).context():
    X1 = ro.conversion.py2rpy(X1)
y1 = ro.FloatVector(y1.tolist())

Then generate randomly 50 samples from Gaussian distribution with 3 covariates:

# Center 2:
n2 = 50   # sample size of center 2
X2 = np.random.normal(size=(n2, p))
X2 = pd.DataFrame(X2, columns=['X1', 'X2', 'X3'])
mu2 = theta[0] + np.dot(X2, np.delete(theta, [0, 4]))
y2 = np.random.normal(loc=mu2, scale=np.sqrt(theta[4]))
with (ro.default_converter + pandas2ri.converter).context():
    X2 = ro.conversion.py2rpy(X2)
y2 = ro.FloatVector(y2.tolist())

MAP Estimates at the Local Centers

The following compute the Maximum A Posterior (MAP) estimators of the parameters for center 1:

# Creating an inverse covariance matrix for a Gaussian prior for center 1
Lambda1 = BFI.inv_prior_cov(X=X1, family="gaussian", **{"lambda": 0.01})
fit1 = BFI.MAP_estimation(y=y1, X=X1, family="gaussian", Lambda=Lambda1)
print(fit1)
theta_hat1 = fit1.rx2("theta_hat") # MAP estimates of the intercept and coefficients
A_hat1 = fit1.rx2("A_hat")         # minus the curvature matrix
summary_1 = ro.r["summary"](fit1, cur_mat = True)
ro.r["print"](summary_1)

Obtaining the MAP estimators of the parameters for center 2 using the following:

# Creating an inverse covariance matrix for Gaussian prior for center 2
Lambda2 = BFI.inv_prior_cov(X=X2, family="gaussian", **{"lambda": 0.01})
# MAP estimates
fit2 = BFI.MAP_estimation(y=y2, X=X2, family="gaussian", Lambda=Lambda2)
theta_hat2 = fit2.rx2("theta_hat") # MAP estimates of the parameters
A_hat2 = fit2.rx2("A_hat")         # minus the curvature matrix around 'theta_hat2'
summary_2 = ro.r["summary"](fit2, cur_mat = True)
ro.r["print"](summary_2)

Sending the Required Files to Central Server

The local estimation results are saved as RDS files, which preserve the R objects and can be read by a coordinating center using either R or Python through rpy2.

Each center saves its own local estimation results. The resulting files are then transferred to the coordinating center using an approved transfer mechanism.

At center 1:

# Save the local estimation results at center 1
base.saveRDS(fit1, file="fit1.rds")

At center 2:

# Save the local estimation results at center 2
base.saveRDS(fit2, file="fit2.rds")

Central server with R

If the central server uses R, the received RDS files can be loaded directly. If it uses Python, see the next subsection: Central server with Python.

The received files are assumed to be available in the current working directory.

# Load the received local estimation results
fit1 <- readRDS("fit1.rds")
fit2 <- readRDS("fit2.rds")

The loaded objects contain the information required for central BFI aggregation, including the local parameter estimates, curvature matrices, and prior precision matrices.

theta_hat1 <- fit1$theta_hat
A_hat1 <- fit1$A_hat

theta_hat2 <- fit2$theta_hat
A_hat2 <- fit2$A_hat

To perform the central aggregation in R using bfi(), see the BFI reference documentation.

Central server with Python

If the central server is using Python, follow the instructions below. We will show how to do this for fit1 at center 1. The process is the same for center 2.

The local BFI results are saved as RDS files, which can also be read in Python using rpy2. The received files are assumed to be available in the current working directory.

# Load the received RDS files
fit1 = base.readRDS("fit1.rds")
fit2 = base.readRDS("fit2.rds")

The loaded objects contain the local estimation results required for central BFI aggregation.

The variable fit1 contains all information needed for further analysis in the central server. To see how to proceed and obtain the BFI estimates on the central server, follow the next section.

BFI at Central Server (using Python)

On the central server, first load the files saved to a proper address:

# Read the two received files
fit1 = base.readRDS("fit1.rds")
fit2 = base.readRDS("fit2.rds")

Now, The main function bfi() can be used to obtain the BFI estimates:

theta_hat1 = fit1.rx2("theta_hat")
A_hat1 = fit1.rx2("A_hat")
theta_hat2 = fit2.rx2("theta_hat")
A_hat2 = fit2.rx2("A_hat")
theta_hats = base.list(theta_hat1, theta_hat2)
A_hats = base.list(A_hat1, A_hat2)
Lambda1 = fit1.rx2("Lambda")
Lambda2 = fit2.rx2("Lambda")
check_prior = ro.r("function(x, y) stopifnot(isTRUE(all.equal(x, y)))")
check_prior(Lambda1, Lambda2)
fit_bfi = BFI.bfi(theta_hats=theta_hats, A_hats=A_hats, 
                  Lambda=Lambda1, family="gaussian")
print(fit_bfi)
summary_bfi = ro.r["summary"](fit_bfi, cur_mat=True)
ro.r["print"](summary_bfi)

Datasets included in the `BFI` package

To find a list of all datasets included in the package run:

print(utils.data(package = "BFI"))

In order to use the datasets available from the BFI package, use the following code:

Nurses = data(BFI).fetch('Nurses')['Nurses']  # is equivalent to BFI::Nurses in R
print("Dimension of the 'Nurses' data: \n", base.dim(Nurses))
print("Colnames of the 'Nurses' data: \n", base.colnames(Nurses))

trauma = data(BFI).fetch('trauma')['trauma']  # is equivalent to BFI::trauma in R
print("Dimension of the 'trauma' data: \n", base.dim(trauma))
print("Colnames of the 'trauma' data: \n", base.colnames(trauma))

All code together

If you need all the code at once to copy and paste them for example into Google Colab, here’s the complete code:

# import 'pandas' package
import pandas as pd

# import 'numpy' package
import numpy as np                

# import 'rpy2' package
import rpy2                  

import rpy2.robjects as ro
from rpy2.robjects.packages import importr, data
from rpy2.robjects.vectors import StrVector
from rpy2.robjects import numpy2ri, pandas2ri

# import R's "utils" package
utils = importr('utils')

# import R's "base" package
base = importr('base')

utils.chooseCRANmirror(ind=1)  # which selects the first mirror in the list
package_names = ('BFI',)
utils.install_packages(StrVector(package_names))

# loading the installed packages
stats = importr('stats')
BFI = importr('BFI')

## Examples

# model parameters: 'theta' and 'p' are assumed to be the same for both centers:
theta = np.array([1, 2, 3, 4, 0.75])   # intercept is theta[0], sigma2 is theta[4]
p = 3     # number of regression parameters without intercept

# Data Simulation for center 1
n1 = 30   # sample size of center 1
X1 = np.random.normal(size=(n1, p)) 
X1.shape   # dimension of X1
X1 = pd.DataFrame(X1, columns=['X1', 'X2', 'X3'])
mu1 = theta[0] + np.dot(X1, np.delete(theta, [0, 4]))  # for gaussian: \eta = \mu 
y1 = np.random.normal(loc=mu1, scale=np.sqrt(theta[4]))
with (ro.default_converter + pandas2ri.converter).context():
    X1 = ro.conversion.py2rpy(X1)
y1 = ro.FloatVector(y1.tolist())

# Data Simulation for center 2
n2 = 50   # sample size of center 2
X2 = np.random.normal(size=(n2, p))
X2 = pd.DataFrame(X2, columns=['X1', 'X2', 'X3'])
mu2 = theta[0] + np.dot(X2, np.delete(theta, [0, 4]))
y2 = np.random.normal(loc=mu2, scale=np.sqrt(theta[4]))
with (ro.default_converter + pandas2ri.converter).context():
    X2 = ro.conversion.py2rpy(X2)
y2 = ro.FloatVector(y2.tolist())

## MAP estimates at center 1
# Creating an inverse covariance matrix for a Gaussian prior for center 1
Lambda1 = BFI.inv_prior_cov(X=X1, family="gaussian", **{"lambda": 0.01})
fit1 = BFI.MAP_estimation(y=y1, X=X1, family="gaussian", Lambda=Lambda1)
print(fit1)
theta_hat1 = fit1.rx2("theta_hat") # MAP estimates of the intercept and coefficients
A_hat1 = fit1.rx2("A_hat")         # minus the curvature matrix
summary_1 = ro.r["summary"](fit1, cur_mat = True)
ro.r["print"](summary_1)
base.saveRDS(fit1, file="fit1.rds")

## MAP estimates at center 2
# Creating an inverse covariance matrix for Gaussian prior for center 2
Lambda2 = BFI.inv_prior_cov(X=X2, family="gaussian", **{"lambda": 0.01})
# MAP estimates
fit2 = BFI.MAP_estimation(y=y2, X=X2, family="gaussian", Lambda=Lambda2)
theta_hat2 = fit2.rx2("theta_hat") # MAP estimates of the parameters
A_hat2 = fit2.rx2("A_hat")         # minus the curvature matrix around 'theta_hat2'
summary_2 = ro.r["summary"](fit2, cur_mat = True)
ro.r["print"](summary_2)
base.saveRDS(fit2, file="fit2.rds")

## BFI at central server
# Load the file using the relative path to the file
fit1 = base.readRDS("fit1.rds")
fit2 = base.readRDS("fit2.rds")

theta_hat1 = fit1.rx2("theta_hat")
theta_hat2 = fit2.rx2("theta_hat")
A_hat1 = fit1.rx2("A_hat")
A_hat2 = fit2.rx2("A_hat")

theta_hats = base.list(theta_hat1, theta_hat2)
A_hats = base.list(A_hat1, A_hat2)

Lambda1 = fit1.rx2("Lambda")
Lambda2 = fit2.rx2("Lambda")

check_prior = ro.r("function(x, y) stopifnot(isTRUE(all.equal(x, y)))")
check_prior(Lambda1, Lambda2)

fit_bfi = BFI.bfi(theta_hats=theta_hats, A_hats=A_hats, Lambda=Lambda1, family="gaussian")
print(fit_bfi)
summary_bfi = ro.r["summary"](fit_bfi, cur_mat=True)
ro.r["print"](summary_bfi)

# To find a list of all datasets included in the package:
print(utils.data(package = "BFI"))

Nurses = data(BFI).fetch('Nurses')['Nurses']
print("Dimension of the 'Nurses' data: \n", base.dim(Nurses))
print("Colnames of the 'Nurses' data: \n", base.colnames(Nurses))

trauma = data(BFI).fetch('trauma')['trauma']
print("Dimension of the 'trauma' data: \n", base.dim(trauma))
print("Colnames of the 'trauma' data: \n", base.colnames(trauma))

Contact

If you find any errors, have any suggestions, or would like to request that something be added, please file an issue at issue report or send an email to: h.pazira@arq.org.

Hassan Pazira

2024-10-12

Introduction

Google Colab

Python Installation

On macOS

On Ubuntu

On Windows

Installation of Required Modules

On macOS

On Ubuntu

On Windows

Python Script

Importing `Python` modules and functions

Installing `R` packages from CRAN and GitHub

Installing the `BFI` package from GitHub (if necessary)

Applying BFI method to the simulated data

Data Simulation for two Local Centers

MAP Estimates at the Local Centers

Sending the Required Files to Central Server

Central server with R

Central server with Python

BFI at Central Server (using Python)

Datasets included in the `BFI` package

All code together

Contact

Calling BFI from Python

Hassan Pazira

2024-10-12

Introduction

Google Colab

Python Installation

On macOS

On Ubuntu

On Windows

Installation of Required Modules

On macOS

On Ubuntu

On Windows

Python Script

Importing Python modules and functions

Installing R packages from CRAN and GitHub

Installing the BFI package from GitHub (if necessary)

Applying BFI method to the simulated data

Data Simulation for two Local Centers

MAP Estimates at the Local Centers

Sending the Required Files to Central Server

Central server with R

Central server with Python

BFI at Central Server (using Python)

Datasets included in the BFI package

All code together

Contact

Importing `Python` modules and functions

Installing `R` packages from CRAN and GitHub

Installing the `BFI` package from GitHub (if necessary)

Datasets included in the `BFI` package