OOP Task Trainer

Recently, ESPnet has a lot of duplicated lines in def train(args): ... of ASR, ASR-MIX, LM, MT, TTS, ST, etc. It is hard to maintain and update because lots of complicated things (custom fp, data loader, transfer learning) are copy-and-pasted in the huge GOD functions. I wanna unify those duplicated parts. How about unifying these with extensible OOP-based design?

# task-dependent part
class TaskInterface:
    @staticmethod
    def add_arguments(parser):

    def init_dataset(self, ...):

    def init_model(self, ...):

    def init_criterion(self, ...):

    def init_sampler(self, ...):


# task-independent training loop
class TaskRunner:
    def train(self, task: TaskInterface):

    def eval(self, task: TaskInterface):

    def run(self):

see also: https://github.com/pytorch/fairseq/blob/master/fairseq/tasks/fairseq_task.py

[ShigekiKarita]

Yeah, I’m also afraid that copying fairseq is not enough for us. My goal is not just splitting task-dependent and task-independent parts because the task-dependent part still can be duplicated.

(I believe that TaskRunner is just a common part of existing Trainer and Evaluator in all the tasks)

However, this idea can make relationship between tasks explicitly (instead of randomly importing symbols). For example, ASR mix class can be derived from ASR. And ST class can be derived from ASR and MT.

2019年9月9日(月) 21:28 Tomoki Hayashi [e-mail]:

I agree with your idea but I’m wondaring that whether we can remove duplicated lines if we use this OOP design.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/espnet/espnet/issues/1188?email_source=notifications&email_token=ABTOZ3Q4RTWHUJNDTLIECADQIY6NXA5CNFSM4IUYFX2KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD6HMC5Q#issuecomment-529449334, or mute the thread https://github.com/notifications/unsubscribe-auth/ABTOZ3RICTEWWGY7ZVAMOM3QIY6NXANCNFSM4IUYFX2A .

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.