Bind max steps and lr iterations

[Rilwan-Adewoyin]

Context & Current Setup:

• User defines training.lr.iterations which effects how many steps the model learns for
• However, this often does not align with the training.max_epochs value which is often larger
• GPU hours potentially wasted as max_epochs =/= training.lr.iterations
• Alternatively unexpected results from training length not aligning to iteration schedule behaviour

Changes Aim to tie training.lr.iterations to the training length

1. Change default method to control run length to be by training.max_steps instead of training.max_epochs
2. Changed default setting in config such that training.lr.iterations defaults to training.max_steps
3. User now defines training.max_steps in order to define length of training
4. User still has the optionality

[FussyDuck]

All committers have signed the CLA.

[mchantry]

Looks good to me. I will run it by the ATS. Don't merge yet.

[anaprietonem]

I think having this functionality it's great, so thanks @Rilwan-Adewoyin for implementing it!
just a quick question, what happens when/if they user pass both max_steps and max_epochs? will the code then run until max_epochs is reached and the scheduler used the max_steps?
My two cents here is that probably it could be nice to add some logger info to indicate this!

[Rilwan-Adewoyin]

Hey thanks,
So the default pytorch lightning behaviour is that the code will run until the smallest of max_steps or max_epochs is reached, but the scheduler will be aligned to max_steps.

Yep that sounds like a good idea, so add a logger.info message when the user sets both