find_unused_parameters in the lightning trainer (1.3.2)

[shamanez]

🐛 Bug

If there are unsued parameters in the model, should we still explicitly mentioned the trainer as follows?

plugins=[DDPPlugin(find_unused_parameters=True)]

trainer = pl.Trainer.from_argparse_args( args, weights_summary=None, callbacks=[logging_callback], logger=logger, plugins=[DDPPlugin(find_unused_parameters=True)], **train_params,)

[awaelchli]

No, it is set to True by default and you want to turn it off unless you need it :)

Reference:
https://pytorch-lightning.readthedocs.io/en/1.3.3/benchmarking/performance.html#when-using-ddp-set-find-unused-parameters-false

[shamanez]

Thanks. Git it