Lightning-AI/pytorch-lightning

Returning None from training_step with multi GPU DDP training

Open

#5243 aperta il 23 dic 2020

Vedi su GitHub
 (26 commenti) (4 reazioni) (1 assegnatario)Python (3233 fork)batch import
distributedfeaturehelp wantedpriority: 1

Metriche repository

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

Descrizione

🐛 Bug

Returning None from training_step with multi GPU DDP training freezes the training without exception

To Reproduce

Starting multi-gpu training with a None-returning training_step function.

Example training_step function:

    def training_step(self, batch, batch_idx):
        data, target = batch
        model_outputs = self.forward(images)
        loss = calc_loss(model_outputs, target)

        if torch.isnan(loss) or random.random() < .05:
            return None

        return loss

Example trainer:

 trainer = Trainer(
    gpus=2,
    distributed_backend="ddp"
)

Expected behavior

To continue training with skipping the current batch as pointed out at here.

Environment

No specific environment is needed to reproduce this bug.

Additional context

This issue was mentioned here: #4956 but not with specifics.

Note: By the time this issue being investigated, a help for a workaround would be great!

Guida contributor