1 6 Advanced - Pytorch Lightning

.. list-table:: adv. user 1.6 :widths: 40 40 20 :header-rows: 1

- If
- Then
- Ref
- passed prepare_data_per_node to the Trainer
- set it as a property of DataHooks, accessible in the LightningModule and LightningDataModule instead
- PR8958_
- used process_position flag
- specify your ProgressBar callback and set it as process_position directly
- PR9222_
- used distributed training attributes add_to_queue and get_from_queue in LightningModule
- user the same methods in DDPStrategy(start_method='spawn')
- PR9118_
- called LightningModule.get_progress_bar_dict
- use the utility function pl.callbacks.progress.base.get_standard_metrics(module.trainer)
- PR9118_
- used LightningModule.on_post_move_to_device
- remove it as parameters tying happens automatically without the need of implementing your own logic
- PR9525_
- relied on Trainer.progress_bar_dict
- use ProgressBarBase.get_metrics
- PR9118_
- used LightningDistributed
- rely on the logic in DDPStrategy(start_method='...')
- PR9691_
- used the Accelerator collective API Accelerator.barrier, Accelerator.broadcast, and Accelerator.all_gather
- call Strategy collectives API directly, without going through Accelerator
- PR9677_
- used pytorch_lightning.core.decorators.parameter_validation
- rely on automatic parameters tying with pytorch_lightning.utilities.params_tying.set_shared_parameters
- PR9525_
- used LearningRateMonitor.lr_sch_names
- access them using LearningRateMonitor.lrs.keys() which will return the names of all the optimizers, even those without a scheduler.
- PR10066_
- implemented DataModule train_transforms, val_transforms, test_transforms, size, dims
- switch to LightningDataModule
- PR8851_