Trainer.parse_argparser does not yield sensible default for default_root_dir

[jeremyjordan]

🐛 Bug

Using Trainer.parse_argparser returns True for default_root_dir, however, a string is expected.

To Reproduce

Steps to reproduce the behavior:

>>> from pytorch_lightning import Trainer
>>> from argparse import ArgumentParser, Namespace
>>> parser = ArgumentParser(add_help=False)
>>> parser = Trainer.add_argparse_args(parent_parser=parser)
>>> args = Trainer.parse_argparser(parser)
>>> args
Namespace(accumulate_grad_batches=1, amp_level='O1', auto_lr_find=False, auto_scale_batch_size=False, auto_select_gpus=False, benchmark=False, check_val_every_n_epoch=1, checkpoint_callback=True, default_root_dir=True, deterministic=False, distributed_backend=True, early_stop_callback=False, fast_dev_run=False, gpus=<function Trainer._arg_default at 0x1219efdd0>, gradient_clip_val=0, log_gpu_memory=True, log_save_interval=100, logger=True, max_epochs=1000, max_steps=True, min_epochs=1, min_steps=True, num_nodes=1, num_processes=1, num_sanity_val_steps=2, overfit_pct=0.0, precision=32, print_nan_grads=False, process_position=0, profiler=True, progress_bar_callback=True, progress_bar_refresh_rate=1, reload_dataloaders_every_epoch=False, replace_sampler_ddp=True, resume_from_checkpoint=True, row_log_interval=10, terminate_on_nan=False, test_percent_check=1.0, tpu_cores=True, track_grad_norm=-1, train_percent_check=1.0, truncated_bptt_steps=True, val_check_interval=1.0, val_percent_check=1.0, weights_save_path=True, weights_summary='full')

[jeremyjordan]

@Borda any ideas?

[jeremyjordan]

Same unexpected behavior for weights_save_path which is causing some tests to fail in #1504, I can temporarily resolve this by treating True the same as None (eg. select default value for the user) but this might not be ideal.

[lezwon]

@jeremyjordan is there an update on this? I was facing this issue while writing CLI tests for #2094. I can continue with those tests once this issue is fixed.

[EspenHa]

This issue actually affects all paramters that default to None in __init__, for instance: min_steps, max_steps, log_gpu_memory, distributed_backend, weights_save_path, truncated_bptt_steps, and resume_from_checkpoint are all set to True.

I have a branch with a test for this here: https://github.com/EspenHa/pytorch-lightning/tree/add_argparse_test
I also implemented a fix here: https://github.com/EspenHa/pytorch-lightning/tree/fix_argparse_bug

@Borda If you would like, I can submit these as a PR?

[Borda]

@EspenHa great, pls send a PR 🚀

[EspenHa]

@Borda PR submitted here: #2526 🚀