Disable UserWarning for DataLoaders with num_workers=0

[marcm-ml]

Proposed refactoring or deprecation

Disable dataloader UserWarning for num_workers=0 since 0 disables multi-process loading (according to pytorch docs it loads in the main process whereas 1 would create a separate process)

Motivation

I understand the motivation behind utilising more processes for data loading but the warning should not appear if the user specifically requested to use the main process (by setting num_workers=0)

Pitch

Change

https://github.com/PyTorchLightning/pytorch-lightning/blob/c33df2639f19d49c5e7520294e3221efe402d684/pytorch_lightning/trainer/data_loading.py#L109-L115

to

elif 0 < dataloader.num_workers <= 2 < num_cpus and not using_spawn:
    rank_zero_warn( 
        f"The dataloader, {name}, does not have many workers which may be a bottleneck." 
        " Consider increasing the value of the `num_workers` argument`" 
        f" (try {num_cpus} which is the number of cpus on this machine)" 
        " in the `DataLoader` init to improve performance." 
    ) 

[Programmer-RD-AI]

Good Idea

[awaelchli]

the warning should not appear if the user specifically requested to use the main process (by setting num_workers=0)

I don't see a way of knowing if the user set num_workers=0 intentionally. The default setting in the PyTorch DataLoader is 0. Is the warning appearing multiple times for you?

[marcm-ml]

You are correct. Maybe I misused the word "specifically" there. The warning appears in scenarios where it should not imo, e.g. when debugging a model. There I set various trainer debug flags but also num_workers=0 for all my dataloaders. Hence, I see the message as many times as there are dataloaders with num_workers=0. Also, when predicting I sometimes use dataloaders with no workers as well since it takes more time to create the workers than loading the data itself.

[carmocca]

This is a controversial warning for the reasons you mentioned. On the other hand, it has helped many users to find this simple bottleneck.

Since there's no way of us knowing, our past recommendation has been to disable the warning:

import warnings
warnings.filterwarnings("ignore", ".*does not have many workers.*")

[awaelchli]

RFC @PyTorchLightning/core-contributors

The intention of the warning was to apply for num_workers=0 as well, as this may be accidentally forgotten by the user and could bottleneck the data loading during training. The warning appears multiple times because the message differs from dataloader to dataloader.

Option 🚀 : do not show warning for 0
Option 🎉 : adjust logic or message such that it only prints once
Option 👀 : keep as is

My preference is towards 🎉 or 👀

[marcm-ml]

I am leaning towards 🚀 but 🎉 is a good trade-off

[carmocca]

I personally don't like option 🎉 because if we implement it and you actually forget the value in N dataloaders, you need to:

1. run your code
2. see the warning
3. fix the num_workers value

N times.

And if you don't want the warning to appear (actually want num_workers=0|1) you still need to filter the warning anyways

[marcm-ml]

I see your concerns and after some thought I believe it is the right choice for the sake of all users to keep this warning for num_workers=0. Perhaps the reason and/or the recommendation can be documented somewhere outside this issue, e.g. under Best Practices -> Dataloaders -> num_workers [1] in addition to what is already written there? Something like:

Dataloaders

[...]

num_workers

The question of how many workers to specify in num_workers is tricky. Here’s a summary of some references, [1], and our suggestions:

1. num_workers=0 means ONLY the main process will load batches (that can be a bottleneck). In case this is intentional, users can disable the warning about low worker count by using:

    import warnings
    warnings.filterwarnings("ignore", ".*does not have many workers.*")

2. num_workers=1 means ONLY ...

[1] Best Practices

[carmocca]

Updating the docs with that sounds great! Feel free to open a PR for it :)