Unexpected behavior with boolean program arguments

[alexarnimueller]

Unexpected behavior with boolean program arguments

[rizaozcelik]

I have created a pull request (#13) with a slightly different behavior than I described above. I implemented a mode flag that can be set to pretrain, finetune, and sample. I tried to keep the changes to the code minimum to minimize possible side effects.

[alexarnimueller]

Hi Riza

Great to hear you found my repo useful.
Thanks a lot for making me aware of this argparse issue. Indeed, the code as it is written now does not allow you to set the training flag to false. What actually needs to be done is to create a no-train flag.

Example:

from argparse import ArgumentParser

parser = ArgumentParser()
parser.add_argument('--notrain', default=False, action='store_true', help='Do not train')
args = parser.parse_args()

if args.notrain:
    print("Not training")
else:
    print("Training")

If you want to apply these changes and be a contributor, I would be very happy.

Thanks in advance

[alexarnimueller]

ah or an actual simpler version is to just change the behavior of the if train in main to something like if train and not finetune

Do you want to adapt it or should I?

[rizaozcelik]

Hi Alex,

Thanks for the prompt response. I would be happy to contribute to the repository and can apply the changes. I propose to leave the flow in the main untouched and change the argument behavior since arguments are the root cause for the strange behavior.

I propose to replace boolean train and finetune flags with a training_mode flag, which woud be set to train or finetune. This should make the behavior explicit for the users.

If you are also okay with that direction, I will create a pull request for you to review.