Deprecate process_position from the Trainer constructor

[ananthsub]

🚀 Feature

Remove this two arguments from the Trainer constructor

Motivation

We are auditing the Lightning components and APIs to assess opportunities for improvements:

• https://docs.google.com/document/d/1xHU7-iQSpp9KJTjI3As2EM0mfNHHr37WZYpDpwLkivA/edit#
• Review Lightning architecture & API #7740

The Trainer today has over 50 arguments to its constructor. This number is growing with each feature release, and makes the trainer cluttered. It also hurts the extensibility of the trainer: a number of arguments passed to the Trainer are to customize other utilities. Plumbing arguments through the trainer creates an undesirable coupling: when the underlying components change, the framework is forced to make breaking API changes in at least two places:

1. The underlying component
2. The trainer constructor

Example: #8062

• we now have to change the trainer API in order to address weights_summary implementation changes (or deprecate weights_summary off the constructor in favor of a callback: Deprecate summarize() off LightningModule #8478 )

Example: #8780
-log_gpu_memory accepting min_max is hyper-specific to nvidia-smi, and isn't applicable for torch.cuda.memory stats

Upcoming examples:

• Supporting new types of progress bars feat: Add Rich Progress Bar #8929
• But our current trainer arguments for progress bars could be specific to the TQDM implementation

Pitch

1. Deprecate process_position off the Trainer constructor in v1.5

2. In version 1.7, remove process_positionfrom the Trainer entirely

To customize this, users can still construct the ProgressBar callback object and pass it to the Trainer.

Alternatives

Additional context

---

If you enjoy Lightning, check out our other projects! ⚡

• Metrics: Machine learning metrics for distributed, scalable PyTorch applications.

• Flash: The fastest way to get a Lightning baseline! A collection of tasks for fast prototyping, baselining, finetuning and solving problems with deep learning

• Bolts: Pretrained SOTA Deep Learning models, callbacks and more for research and production with PyTorch Lightning and PyTorch

• Lightning Transformers: Flexible interface for high performance research using SOTA Transformers leveraging Pytorch Lightning, Transformers, and Hydra.

[ananthsub]

@kaushikb11 would this make integrating new progress bars like Rich more straightforward?

[awaelchli]

Agree with process_position, nobody uses it and I believe the documentation is slightly misleading.
I am not sure about the refresh rate though. Removing that may introduce a major inconvenience for users.

[ananthsub]

Agree with process_position, nobody uses it and I believe the documentation is slightly misleading.
I am not sure about the refresh rate though. Removing that may introduce a major inconvenience for users.

Trainer(callbacks=[ProgressBar(refresh_rate=X)])

vs

Trainer(progress_bar_refresh_rate=X)

This doesn't seem inconvenient for me. Especially if users already have other callbacks they use, this is very natural. Wdyt?

Motivation:

• Keep Trainer constructor specific to what the Trainer actually needs
• Allow the ProgressBar to add more args without the Trainer needing to be aware or set defaults (and even allow us to test out multiple progress bar implementations without needing to shoehorn this arg into all of them)
• Consistency with other callbacks: customizations for frameworks provided by the callback like ModelCheckpoint and EarlyStopping are expected to be instantiated this way. Why do we treat ProgressBar differently?
• Long-term health & expectations for more components: The trainer needing to hold arguments for each of these isn't scaling