PyTorch DataLoader uses same random seed for batches run in parallel

user_9067615 0
Viewed 3844

There is a bug in PyTorch/Numpy where when loading batches in parallel with a DataLoader (i.e. setting num_workers > 1), the same NumPy random seed is used for each worker, resulting in any random functions applied being identical across parallelized batches.

Minimal example:

import numpy as np
from torch.utils.data import Dataset, DataLoader

class RandomDataset(Dataset):
    def __getitem__(self, index):
        return np.random.randint(0, 1000, 2)

    def __len__(self):
        return 9
    
dataset = RandomDataset()
dataloader = DataLoader(dataset, batch_size=1, num_workers=3)

for batch in dataloader:
    print(batch)

As you can see, for each parallelized set of batches (3), the results are the same:

# First 3 batches
tensor([[891, 674]])
tensor([[891, 674]])
tensor([[891, 674]])
# Second 3 batches
tensor([[545, 977]])
tensor([[545, 977]])
tensor([[545, 977]])
# Third 3 batches
tensor([[880, 688]])
tensor([[880, 688]])
tensor([[880, 688]])

What is the recommended/most elegant way to fix this? i.e. have each batch produce a different randomization, irrespective of the number of workers.

dataloader pytorch parallel-processing numpy python
2 Answers

It seems this works, at least in Colab:

dataloader = DataLoader(dataset, batch_size=1, num_workers=3, 
    worker_init_fn = lambda id: np.random.seed(id) )

EDIT:

it produces identical output (i.e. the same problem) when iterated over epochs. – iacob

Best fix I have found so far:

...
dataloader = DataLoader(ds, num_workers= num_w, 
           worker_init_fn = lambda id: np.random.seed(id + epoch * num_w ))

for epoch in range ( 2 ):
    for batch in dataloader:
        print(batch)
    print()

Still can't suggest closed form, thing depends on a var (epoch) then called. Ideally It must be something like worker_init_fn = lambda id: np.random.seed(id + EAGER_EVAL(np.random.randint(10000) ) where EAGER_EVAL evaluate seed on loader construction, before lambda is passed as parameter. Is it possible in python, I wonder.

user_40945740

To add onto Alexey's answer, if you persist the workers then you only need to set the seed per worker once on initialization using the worker_init_fn. There is a dataloader param explicitly for this.

Another alternative, prob best, is to use torch itself for RNG as it handles parallelism by default, instead of python's builtin random or numpys' random. Then use torch.seed(my_fav_number) before u call any data loaders for deterministic results.

user_135434110
Related