Python and Random State Sharing with NumPy and PyTorch, depending on different multiprocessing start method (i.e. fork vs spawn)

multiprocessing_seed_numpy.py

import numpy as np
import multiprocessing as mp
from argparse import ArgumentParser


def get_current_seed(wid):
    s = np.random.get_state()[1][0]
    print(f"Pool Worker {wid}: Random State {s}")


if __name__ == "__main__":
    parser = ArgumentParser()
    parser.add_argument(
        "--with-fork",
        action="store_true",
        help="Force fork method",
    )
    args = parser.parse_args()

    if args.with_fork:
        mp.set_start_method("fork")
    print("Start Method: ", mp.get_start_method())

    SEED = 123456
    # set seed
    np.random.seed(SEED)
    print("Main Process : Random State", np.random.get_state()[1][0])

    with mp.Pool(5) as p:
        p.map(get_current_seed, range(5))

multiprocessing_seed_torch.py

# Identical to NumPy version, but using PyTorch random

import torch as th
import multiprocessing as mp
from argparse import ArgumentParser


def get_current_seed(wid):
    s = th.random.initial_seed()
    print(f"Pool Worker {wid}: Random State {s}")


if __name__ == "__main__":
    parser = ArgumentParser()
    parser.add_argument(
        "--with-fork",
        action="store_true",
        help="Force fork method",
    )
    args = parser.parse_args()

    if args.with_fork:
        mp.set_start_method("fork")
    print("Start Method: ", mp.get_start_method())

    SEED = 123456
    # set seed
    th.random.manual_seed(SEED)
    print("Main Process : Random State", th.random.initial_seed())

    with mp.Pool(5) as p:
        p.map(get_current_seed, range(5))

run.md

Replicability

This output has been generated using Python 3.8 on macOS, which is the first release of Py3 on macOS to use spawn as the default start_method (more here)

numpy

❯ python multiprocessing_seed_numpy.py
Start Method:  spawn
Main Process : Random State 123456
Pool Worker 0: Random State 2147483648
Pool Worker 1: Random State 2147483648
Pool Worker 3: Random State 2147483648
Pool Worker 2: Random State 2147483648
Pool Worker 4: Random State 2147483648

❯ python multiprocessing_seed_numpy.py --with-fork
Start Method:  fork
Main Process : Random State 123456
Pool Worker 0: Random State 123456
Pool Worker 1: Random State 123456
Pool Worker 2: Random State 123456
Pool Worker 3: Random State 123456
Pool Worker 4: Random State 123456

torch

❯ python multiprocessing_seed_torch.py
Start Method:  spawn
Main Process : Random State 123456
Pool Worker 0: Random State 12008651253415256287
Pool Worker 1: Random State 12008651253415256287
Pool Worker 2: Random State 12008651253415256287
Pool Worker 3: Random State 12008651253415256287
Pool Worker 4: Random State 12008651253415256287

❯ python multiprocessing_seed_torch.py --with-fork
Start Method:  fork
Main Process : Random State 123456
Pool Worker 0: Random State 123456
Pool Worker 1: Random State 123456
Pool Worker 2: Random State 123456
Pool Worker 3: Random State 123456
Pool Worker 4: Random State 123456