How to resolve could not infer the batch_size from the batch

[Zhengxinyang]

I encountered the following error report：

pytorch_lightning.utilities.exceptions.MisconfigurationException: We could not infer the batch_size from the batch. Either simplify its structure or provide the batch_size as self.log(..., batch_size=batch_size).

How can I fix it without changing the data loader?

[akihironitta]

How can I fix it without changing the data loader?

@Zhengxinyang You can fix the error by providing the batch_size as self.log(..., batch_size=batch_size) as the error suggests. (Or, are you asking something different?)

[Zhengxinyang]

Thanks. In fact, I found this error triggered by Trainer(... , track_grad_norm=2), how can I fix it?

[akihironitta]

@Zhengxinyang track_grad_norm shouldn't impact it at all AFAIK. I've tried to reproduce it with PL master, however, I cannot reproduce the behaviour. https://github.com/akihironitta/gist/blob/repro/13683-track-grad-norm-batch-size/pl_boring_model/main.py

Could you provide the env detail and script to reproduce?

[rubencart]

Running into the same error, also only when track_grad_norm=2. I think it is because tracking the gradient norm makes a call to log_dict without batch_size=..., see part of the stacktrace below.

  File ".../miniconda3/envs/satenv4/lib/python3.10/site-packages/pytorch_lightning/plugins/precision/precision_plugin.py", line 85, in _after_closure
    self._track_grad_norm(trainer)
  File ".../miniconda3/envs/satenv4/lib/python3.10/site-packages/pytorch_lightning/plugins/precision/precision_plugin.py", line 135, in _track_grad_norm
    trainer.lightning_module.log_grad_norm(grad_norm_dict)
  File ".../miniconda3/envs/satenv4/lib/python3.10/site-packages/pytorch_lightning/core/module.py", line 565, in log_grad_norm
    self.log_dict(grad_norm_dict, on_step=True, on_epoch=True, prog_bar=False, logger=True)

[janfer95]

Are there any news regarding this? I run into the same error when using track_grad_norm=1 in a pytorch_geometric / pytorch_lightning code, even though I do pass self.log(..., batch_size=batch_size in training_step.

Is there maybe a way to pass it directly to track_grad_norm?

[So-cean]

same problem, when using torch_geometric. it seems that lightning treated pyg data's num_nodes as batch_size