Precision defaults on trainer constructor

[ananthsub]

I'm going through the Trainer constructor as part of the API review #7740

• Why does the amp_backend on the Trainer constructor have a default value of ‘native’ ? By default, it’s a no-op because the default value for precision is 32. Should this be optional instead?
• Why does amp_level have a default value of 'O2'? This can be confusing with the default for amp_backend because because pytorch native amp doesn’t expose an optimization level setting. This is used only for apex amp. Should this be optional instead?

https://github.com/PyTorchLightning/pytorch-lightning/blob/5329b0d11358b72a42d596d3b70b7010e35c45af/pytorch_lightning/trainer/trainer.py#L140-L157