PyTorch DataLoader uses same random seed for batches run in parallel
Solution 1:
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.