Lightning-AI/pytorch-lightning

Distributed mode overwrites the user's choice for dataloaders shuffling

Open

#10.170 aperta il 27 ott 2021

Vedi su GitHub
 (8 commenti) (0 reazioni) (0 assegnatari)Python (3233 fork)batch import
data handlingfeaturehelp wanted

Metriche repository

Star
 (26.687 star)
Metriche merge PR
 (Merge medio 9g 15h) (3 PR mergiate in 30 g)

Descrizione

🐛 Bug

I ordered my training data in a specific manner and passed it to the DataLoader with shuffle=False (I use reload_dataloaders_every_n_epochs=1 to control it every epoch). Then, I found out that running in distributed mode re-creates the train dataloader and always set shuffle=True. The opposite happens for the val dataloader.

I think that the re-creation of the dataloaders must be using the same shuffle setting the original dataloaders had. Is there any reason not to do so?


Here is the related code with the hard-coded shuffle values:

train - shuffle forced to be true https://github.com/PyTorchLightning/pytorch-lightning/blob/c33df2639f19d49c5e7520294e3221efe402d684/pytorch_lightning/trainer/data_loading.py#L330-L333

val - shuffle forced to be true https://github.com/PyTorchLightning/pytorch-lightning/blob/c33df2639f19d49c5e7520294e3221efe402d684/pytorch_lightning/trainer/data_loading.py#L442-L443

cc @borda @justusschock @awaelchli

Guida contributor