PaddlePaddle / PaddleCloud

PaddlePaddle Docker images and K8s operators for PaddleOCR/Detection developers to use on public/private cloud.
Apache License 2.0
284 stars 76 forks source link

Some points found in debugging. #427

Open gongweibao opened 7 years ago

gongweibao commented 7 years ago
typhoonzero commented 7 years ago

If pserver can't save checkpoint, the job can't be fault-tolerant, should the whole job exit?

We don't consider pserver fault tolerant currently. It's not a blocking issue, yet we need to fix it later.

If num_sample == 0 when we do evaluator, what cost should be displayed?

If num_sample == 0 trainer.py would never call event_handler

Should we support split a job with a file-list not only recordio format.

I agree, I remember there was an issue for that.

helinwang commented 7 years ago

If pserver can't save checkpoint, the job can't be fault-tolerant, should the whole job exit?

Did you found that pserver can't save checkpoint? I have a fix here: https://github.com/PaddlePaddle/Paddle/pull/5053

Should we support split a job with a file-list not only recordio format.

That is a good point, probably not for this experiment. (this can be fixed after refactor, everything becomes an OP).

It's easy to confuse to call it cloud_reader because we can cloud_reader train data and so we can read test data from the cloud? So rename it to get_task_train_data?

Do you mean we need to express the fact that cloud_reader is just for training, not for testing? Yes that's is a good point. get_task_train_data is a candidate, maybe we need to think for more candidates before we decide. Let's not worry about this one now, and focus on finishing the experiment and fixing all the bugs found by the experiment.

gongweibao commented 7 years ago

Ye, all of these are just for discussion and should not block the experiment.