SyncRNG 1.3.3

Generate the same random numbers in R and Python

SyncRNG

Generate the same random numbers in R and Python.

Useful Links:

• SyncRNG on GitHub
• SyncRNG on PyPI
• SyncRNG on CRAN
• Blog post on SyncRNG

Contents: Introduction | Installation | Usage | Functionality | R: User defined RNG | Examples | Sampling without replacement | Sampling with replacement | Generating Normally Distributed Values | Creating the same train/test splits | Notes

Introduction

I created this package because I needed to have the same random numbers in both R and Python programs. Although both languages implement a Mersenne-Twister random number generator (RNG), the implementations are so different that it is not possible to get the same random numbers, even with the same seed.

SyncRNG is a "Tausworthe" RNG implemented in C and linked to both R and Python. Since both use the same underlying C code, the random numbers will be the same in both languages when the same seed is used. A Tausworthe generator is based on a linear feedback shift register and relatively easy to implement.

You can read more about my motivations for creating this here.

If you use SyncRNG in your work, please consider citing it. Here is a BibTeX entry you can use:

@misc{vandenburg2015syncrng,
  author={{Van den Burg}, G. J. J.},
  title={{SyncRNG}: Synchronised Random Numbers in {R} and {Python}},
  url={https://github.com/GjjvdBurg/SyncRNG},
  year={2015},
  note={Version 1.3}
}

Installation

Installing the R package can be done through CRAN:

> install.packages('SyncRNG')

The Python package can be installed using pip:

$ pip install syncrng

Usage

After installing the package, you can use the basic SyncRNG random number generator. In Python you can do:

>>> from SyncRNG import SyncRNG
>>> s = SyncRNG(seed=123456)
>>> for i in range(10):
>>>     print(s.randi())

And in R you can use:

> library(SyncRNG)
> s <- SyncRNG(seed=123456)
> for (i in 1:10) {
>    cat(s$randi(), '\n')
> }

You'll notice that the random numbers are indeed the same.

Functionality

In both R and Python the following methods are available for the SyncRNG class:

1. randi(): generate a random integer on the interval [0, 2^32).
2. rand(): generate a random floating point number on the interval [0.0, 1.0)
3. randbelow(n): generate a random integer below a given integer n.
4. shuffle(x): generate a permutation of a given list of numbers x.

Functionality is deliberately kept minimal to make maintaining this library easier. It is straightforward to build more advanced applications on the existing methods, as the examples below show.

R: User defined RNG

R allows the user to define a custom random number generator, which is then used for the common runif function in R. This has also been implemented in SyncRNG as of version 1.3.0. To enable this, run:

> library(SyncRNG)
> set.seed(123456, 'user', 'user')
> runif(10)

These numbers are between [0, 1) and multiplying by 2**32 - 1 gives the same results as above. Note that while this works for low-level random number generation using runif, it is not guaranteed that higher-level functions that build on this (such as rnorm and sample) translate easily to similar functions in Python. This has likely to do with R's internal implementation for these functions. Using random number primitives from SyncRNG directly is therefore generally more reliable. See the examples below for sampling and generating normally distributed values with SyncRNG.

Examples

This section contains several examples of functionality that can easily be built on top of the primitives that SyncRNG provides.

Sampling without replacement

Sampling without replacement can be done by leveraging the builtin shuffle method of SyncRNG:

R:

> library(SyncRNG)
> v <- 1:10
> s <- SyncRNG(seed=42)
> # Sample 5 values without replacement
> s$shuffle(v)[1:5]
[1] 6 9 2 4 5

Python:

>>> from SyncRNG import SyncRNG
>>> v = list(range(1, 11))
>>> s = SyncRNG(seed=42)
>>> # Sample 5 values without replacement
>>> s.shuffle(v)[:5]
[6, 9, 2, 4, 5]

Sampling with replacement

Sampling with replacement simply means generating random array indices. Note that these values are not (necessarily) the same as what is returned from R's sample function, even if we specify SyncRNG as the user-defined RNG (see above).

R:

> library(SyncRNG)
> v <- 1:10
> s <- SyncRNG(seed=42)
> u <- NULL
> # Sample 15 values with replacement
> for (k in 1:15) {
+ idx <- s$randi() %% length(v) + 1
+ u <- c(u, v[idx])
+ }
> u
[1] 10  1  1  9  3 10 10 10  9  4  1  9  6  3  6

Python:

>>> from SyncRNG import SyncRNG
>>> v = list(range(1, 11))
>>> s = SyncRNG(seed=42)
>>> u = []
>>> for k in range(15):
...     idx = s.randi() % len(v)
...     u.append(v[idx])
...
>>> u
[10, 1, 1, 9, 3, 10, 10, 10, 9, 4, 1, 9, 6, 3, 6]

Generating Normally Distributed Values

It is also straightforward to implement a Box-Muller transform to generate normally distributed samples.

R:

library(SyncRNG)

# Generate n numbers from N(mu, sigma^2)
syncrng.box.muller <- function(mu, sigma, n, seed=0, rng=NULL)
{
    if (is.null(rng)) {
        rng <- SyncRNG(seed=seed)
    }

    two.pi <- 2 * pi
    ngen <- ceiling(n / 2)
    out <- replicate(2 * ngen, 0.0)

    for (i in 1:ngen) {
        u1 <- 0.0
        u2 <- 0.0

        while (u1 == 0) { u1 <- rng$rand(); }
        while (u2 == 0) { u2 <- rng$rand(); }

        mag <- sigma * sqrt(-2.0 * log(u1))
        z0 <- mag * cos(two.pi * u2) + mu
        z1 <- mag * sin(two.pi * u2) + mu

        out[2*i - 1] = z0;
        out[2*i] = z1;
    }
    return(out[1:n]);
}

> syncrng.box.muller(1.0, 3.0, 11, seed=123)
 [1]  9.6062905  1.4132851  1.0223211  1.7554504 13.5366881  1.0793818
 [7]  2.5734537  1.1689116  0.5588834 -6.1701509  3.2221119

Python:

import math
from SyncRNG import SyncRNG

def syncrng_box_muller(mu, sigma, n, seed=0, rng=None):
    """Generate n numbers from N(mu, sigma^2)"""
    rng = SyncRNG(seed=seed) if rng is None else rng

    two_pi = 2 * math.pi
    ngen = math.ceil(n / 2)
    out = [0.0] * 2 * ngen

    for i in range(ngen):
        u1 = 0.0
        u2 = 0.0

        while u1 == 0:
            u1 = rng.rand()
        while u2 == 0:
            u2 = rng.rand()

        mag = sigma * math.sqrt(-2.0 * math.log(u1))
        z0 = mag * math.cos(two_pi * u2) + mu
        z1 = mag * math.sin(two_pi * u2) + mu

        out[2*i] = z0
        out[2*i + 1] = z1

    return out[:n]

>>> syncrng_box_muller(1.0, 3.0, 11, seed=123)
[9.60629048280169, 1.4132850614143178, 1.0223211130311138, 1.7554504380249232, 
13.536688052073458, 1.0793818230927306, 2.5734537321359925, 
1.1689116061110083, 0.5588834007200677, -6.1701508943037195, 
3.2221118937024342]

Creating the same train/test splits

A common use case for this package is to create the same train and test splits in R and Python. Below are some code examples that illustrate how to do this. Both assume you have a matrix X with 100 rows.

R:

# This function creates a list with train and test indices for each fold
k.fold <- function(n, K, shuffle=TRUE, seed=0)
{
	idxs <- c(1:n)
	if (shuffle) {
		rng <- SyncRNG(seed=seed)
		idxs <- rng$shuffle(idxs)
	}

	# Determine fold sizes
        fsizes <- c(1:K)*0 + floor(n / K)
        mod <- n %% K
        if (mod > 0)
		fsizes[1:mod] <- fsizes[1:mod] + 1

        out <- list(n=n, num.folds=K)
	current <- 1
        for (f in 1:K) {
		fs <- fsizes[f]
		startidx <- current
		stopidx <- current + fs - 1
		test.idx <- idxs[startidx:stopidx]
		train.idx <- idxs[!(idxs %in% test.idx)]
		out$testidxs[[f]] <- test.idx
		out$trainidxs[[f]] <- train.idx
		current <- stopidx
	}
	return(out)
}

# Which you can use as follows
folds <- k.fold(nrow(X), K=10, shuffle=T, seed=123)
for (f in 1:folds$num.folds) {
        X.train <- X[folds$trainidx[[f]], ]
        X.test <- X[folds$testidx[[f]], ]

        # continue using X.train and X.test here
}

Python:

def k_fold(n, K, shuffle=True, seed=0):
    """Generator for train and test indices"""
    idxs = list(range(n))
    if shuffle:
        rng = SyncRNG(seed=seed)
        idxs = rng.shuffle(idxs)

    fsizes = [n // K]*K
    mod = n % K
    if mod > 0:
        fsizes[:mod] = [x+1 for x in fsizes[:mod]]

    current = 0
    for fs in fsizes:
        startidx = current
        stopidx = current + fs
        test_idx = idxs[startidx:stopidx]
        train_idx = [x for x in idxs if not x in test_idx]
        yield train_idx, test_idx
        current = stopidx

# Which you can use as follows
kf = k_fold(X.shape[0], K=3, shuffle=True, seed=123)
for trainidx, testidx in kf:
    X_train = X[trainidx, :]
    X_test = X[testidx, :]

    # continue using X_train and X_test here

Notes

The random numbers are uniformly distributed on [0, 2^32 - 1]. No attention has been paid to thread-safety and you shouldn't use this random number generator for cryptographic applications.

If you have questions, comments, or suggestions about SyncRNG or you encounter a problem, please open an issue on GitHub. Please don't hesitate to contact me, you're helping to make this project better for everyone! If you prefer not to use Github you can email me at gertjanvandenburg at gmail dot com.