pl.Trainer.add_argparse_args() does not work with argparse subparsers

[mattbui]

🐛 Bug

https://github.com/PyTorchLightning/pytorch-lightning/blob/7114c2d570e689cdddc6d321ad29189d54184743/pytorch_lightning/utilities/argparse.py#L157-L160 will override the parent_parser so the sub parser will not work as expected. Please see the code in reproduce section.

I think the current version from master with use_argument_group=True will fix this. https://github.com/PyTorchLightning/pytorch-lightning/blob/efce2b77779467884df9a3d9c16c3176ea81a650/pytorch_lightning/utilities/argparse.py#L190-L197

Please reproduce using the BoringModel

To Reproduce

in a cli.py file

import argparse
import pytorch_lightning as pl

parser = argparse.ArgumentParser("")

sub_parsers = parser.add_subparsers()

train_parser = sub_parsers.add_parser("train")
train_parser.add_argument("--seed")
train_parser = pl.Trainer.add_argparse_args(train_parser)

args = parser.parse_args()

Run python cli.py train --help will not show the training flags for Trainer

Expected behavior

Run python cli.py train --help show the training flags for Trainer

Environment

• PyTorch Version (e.g., 1.0): 1.7.0
• OS (e.g., Linux): MacOS Big Sur 11.2.3
• How you installed PyTorch (conda, pip, source): conda
• Python version: 3.7.10

Additional context

[Borda]

@mattbui mind check if #4492 would be a good alternative for you?

[carmocca]

This is already fixed in master