* Replace np.random by random

[vfdev-5]

To be able to fix random state with a single command:

random.seed(12345)

instead of

np.random.seed(12345)
random.seed(12345)

[alykhantejani]

this would now cause breaking changes in users code if they were previously setting np.random.seed to fix the seed in the transforms...

Perhaps a better solution would be to pass in a random state to the functions, and if it is None it uses the current random state (got by np.random.get_state).

Thoughts @fmassa ?

[ducha-aiki]

What about linking seed to actually torch.manual_seed ?

[vfdev-5]

@ducha-aiki in this case all torchvision random stuff is needed to rewrite with torch.random what is much more breaking change than just drop np.random

[fmassa]

I think we need to have some uniformity in here, so we might need to go through a breaking change (which is very small though).
I think we might want to pass a seed argument to the random transforms. But just passing the seed is not enough because we don't want to always have the same transforms, nor to always have to manually change the seed.

Maybe we will need to pass a initial random state, which will be updated by the random transforms.
Something like

def random_flip(img, random_state=None):
    if random_state is not None:
        torch.set_rng_state(random_state)
    # perform the random operations using torch
    ...
    # now update the random state
    if random_state is not None:
        random_state.copy_(torch.get_rng_state)
    # and maybe set the torch rng state back to its original value

But that's a bit annoying and might bring some additional overhead and maybe issues on multithreading behaving the same for different threads.
What do you think?

[fmassa]

I think this should be unified. Thanks!

[laoreja]

I use np.random in my dataset class, and noticed the issue on multithreading behaving the same for different threads.
My fix is adding worker_init_fn=lambda x: np.random.seed((torch.initial_seed() + x) % (2 ** 32)) to the DataLoader initialization.

If I also use your transform functions that uses the standard random library, should I add a line to the worker_init_fn also, and actually I do not understand what's the problem behind the multithreading, can you give some explanation?

[vfdev-5]

@laoreja random and np.random behave differently if multithreading type is "fork".
Take a look at this snippet:

import multiprocessing as mp

import random
import numpy as np

def task():
    print("-- Task --")
    print(random.random(), np.random.rand())
    print(random.random(), np.random.rand())
    print("-- End Task --")
    
random.seed(12345)
np.random.seed(12345)

workers = [mp.Process(target=task, args=()) for i in range(4)]

print("Run")
for w in workers:
    w.start()

and the output

Run
-- Task --
0.6062518296570525 0.9296160928171479
0.7142844140096419 0.3163755545817859
-- End Task --
-- Task --
0.6959326394013083 0.9296160928171479
0.31228532647566887 0.3163755545817859
-- End Task --
-- Task --
0.4936887984487045 0.9296160928171479
0.7509582156538067 0.3163755545817859
-- Task --
-- End Task --
0.6168844031764428 0.9296160928171479
0.5782864572237567 0.3163755545817859
-- End Task --

So, you do not need to add new seed for the workers if use random.

HTH

[laoreja]

Thanks!