randomtransforms.py

import random


class CMWC(random.Random):
    # code from here https://code.activestate.com/recipes/576707/
    'Long period random number generator: Complementary Multiply with Carry'
    # http://en.wikipedia.org/wiki/Multiply-with-carry

    a = 3636507990
    logb = 32
    b = 2 ** logb
    r = 1359

    def _gen_word(self):
        i = self.i
        xc, self.c = divmod(self.a * self.Q[i] + self.c, self.b)
        x = self.Q[i] = self.b - 1 - xc
        self.i = 0 if i + 1 == self.r else i + 1
        return x

    def getrandbits(self, k):
        while self.bits < k:
            self.f = (self.f << self.logb) | self._gen_word()
            self.bits += self.logb
        x = self.f & ((1 << k) - 1)
        self.f >>= k
        self.bits -= k
        return x

    def random(self, RECIP_BPF=random.RECIP_BPF, BPF=random.BPF):
        return self.getrandbits(BPF) * RECIP_BPF

    def seed(self, seed=None):
        seeder = random.Random(seed)
        Q = [seeder.randrange(0x100000000) for i in range(self.r)]
        c = seeder.randrange(0x100000000)
        self.setstate(0, 0, 0, c, Q)

    def getstate(self):
        return self.f, self.bits, self.i, self.c, tuple(self.Q)

    def setstate(self, f, bits, i, c, Q):
        self.f, self.bits, self.i, self.c, self.Q = f, bits, i, c, list(Q)


class myrandomtransform():
    def __init__(self, seed=None):
        self.rng = CMWC()
        if seed is not None:
            self.rng.seed(seed)

    def run(self):
        return self.rng.random()

a = myrandomtransform(2)
b = myrandomtransform(2)

avalues = []
bvalues = []
for i in range(10):
    if random.random() > 0.5:
        avalues.append(a.run())
    else:
        bvalues.append(b.run())
print(avalues)
print(bvalues)

Why so complicated?

class RandomTransform:

    def __init__(self, seed: int = None):
        self.rng = random.Random(seed)

    def run(self):
        return self.rng.random()

should do the same thing, no?

Hello,

Yes and no. :-) The point here is to control the random number generator (RNG) to prevent:

• non deterministic behavior due to low level routines, other implementation hidden issues
• change of behavior between different RNG versions

I was intending to use it to comment on this issue but somehow I did not (can't remember why): pytorch/vision#1234

So it is not trying to be super random, just enough for testing purposes (as long as the testing does not require high quality randomness) while guaranteeing reproducibility of the randomness which is paramount for testing.