Rohit Gupta

> also like @carmocca's suggestion for adding a flag mixed_precision=True | False even though that's adding one more thing to remember for users. I think it may be fine if...

I like it... just not sure `precision-type-amp_backend` is a good way to get the configuration. Should we do dict? ```py {'precision: 16, 'mixed': False, amp_type: 'apex'} ``` key names can...

> they are consistent with our other registry based trainer arguments Well, sort of. They are indicated by the argument name in their respective strategy class (`"deepspeed_stage_3")`, but here one...

```py class Mod1(LightningModule): def __init__(self, same_arg): super().__init__() self.same_arg = same_arg self.save_hyperparameters() class Mod2(LightningModule): def __init__(self, same_arg): super().__init__() self.same_arg = same_arg self.save_hyperparameters() class Parent(): def __init__(self, same_arg="parent", diff_arg="test"): super().__init__() self.m1 =...

I'll try to complete focal loss and smoothing loss this week #121

I guess it should be `batches * num_epochs` but why would it be ` * ab_size`?

okay. yes... I didn't see max_epochs there. should be something like ``` total = (total / accumulation_factor)*max_epochs. ```

> After some analysis, it seems the last epoch at which the model was saved is rerun with some restart internal flags do you mean the checkpoint that was restored...

then it should start from the new epoch and rerun the complete new epoch with newly generated metrics that will be used to update LR schedulers. Can you share a...

> @rohitgr7 We could potentially have a hydra specific launcher after #11643 is finalized. In that sense, I like the approach of this PR of creating a utility function encapsulating...