calling iter twice messes up dataloaders with queues

[ben-da6]

Bug description

This bug has reappeared #18414

We now call iter() twice in different places:

• https://github.com/Lightning-AI/pytorch-lightning/blob/master/src/lightning/pytorch/loops/fit_loop.py#L263
• https://github.com/Lightning-AI/pytorch-lightning/blob/master/src/lightning/pytorch/loops/training_epoch_loop.py#L171C1-L172C1

What version are you seeing the problem on?

v2.1

How to reproduce the bug

import multiprocessing as mp
from queue import Queue
from typing import Iterator

import numpy as np
from lightning import Trainer
from lightning.pytorch.demos.boring_classes import BoringModel
from torch.utils.data import DataLoader, IterableDataset


class QueueDataset(IterableDataset):
    def __init__(self, queue: Queue) -> None:
        super().__init__()
        self.queue = queue

    def __iter__(self) -> Iterator:
        for k in range(5):
            print(f"getting {k}")
            tensor, index = self.queue.get(timeout=10)
            print(f"got {index}")
            yield tensor


if __name__ == "__main__":
    q = mp.Queue()
    arr = np.random.random([1, 32]).astype(np.float32)
    for ind in range(10):
        q.put((arr, ind))
    max_epoch = 1
    dataloader = DataLoader(QueueDataset(q), num_workers=1, batch_size=None, persistent_workers=True)
    trainer = Trainer(max_epochs=max_epoch, enable_progress_bar=False, devices=1)
    trainer.fit(BoringModel(), dataloader)
    trainer.save_checkpoint("model.ckpt")

    # q now has the next 5 elems in
    # resuming training we will hit the double iter() issue
    dataloader = DataLoader(QueueDataset(q), num_workers=1, batch_size=None, persistent_workers=True)
    trainer = Trainer(max_epochs=max_epoch + 1, enable_progress_bar=False, devices=1)
    trainer.fit(BoringModel(), dataloader, ckpt_path="model.ckpt")

Error messages and logs

relevant logs are:

# first epoch all good
getting 0
got 0
getting 1
got 1
getting 2
got 2
getting 3
got 3
getting 4
got 4

# second epoch we start getting from the queue twice!
# from fit loop iter()
getting 0
got 5
getting 1
got 6
getting 2
got 7
# from training_epoch loop iter()
getting 0
got 8
getting 1
got 9
getting 2

Environment

lighting==2.1.4

More info

No response

cc @justusschock @awaelchli @carmocca

[awaelchli]

This condition here is meant to prevent the iter() from getting called a second time, because in this case restarting should be True.

pytorch-lightning/src/lightning/pytorch/loops/training_epoch_loop.py

Lines 169 to 171 in 47c8f4c

	# `iter()` was called once in `FitLoop.setup_data()` already
	if self.trainer.current_epoch > 0 and not self.restarting:
	iter(data_fetcher) # creates the iterator inside the fetcher

But it isn't. The problem is that the fit loop sets restarting=False even though we are resuming, due to the logic here:

pytorch-lightning/src/lightning/pytorch/loops/fit_loop.py

Lines 123 to 128 in 47c8f4c

	def restarting(self, restarting: bool) -> None:
	# if the last epoch completely finished, we are not actually restarting
	values = self.epoch_progress.current.ready, self.epoch_progress.current.started
	epoch_unfinished = any(v != self.epoch_progress.current.processed for v in values)
	restarting = restarting and epoch_unfinished or self._iteration_based_training()
	_Loop.restarting.fset(self, restarting) # call the parent setter

This is tricky to solve @carmocca. The logic probably needs to be lifted up into the fit loop before epoch_loop.run(), with a different conditioning that does not rely on restarting.

[carmocca]

I didn't look too deeply. Couldn't we check restarting too for the FitLoop's iter call? We have a lot of tests around this so If a solution passes them we should be good.

[ben-da6]

The problem in the restarting property is self._iteration_based_training() is False

[ben-da6]

Also since this has appeared twice now, and its the sort of bug which is hard to track down could we add a test like my example?