Log_save_interval and row_log_interval

Trainer
#1

What do log_save_interval and row_log_interval from Trainer flags do?
I noticed logging in tensorboard is done at row_log_interval (default 50) while in wandb it doesn’t.

GH issue: https://github.com/PyTorchLightning/pytorch-lightning/issues/3339

#2

Why not for wandb?
I’m looking at the trainer code and it is applied to whatever logger is used:

should_log_metrics = (batch_idx + 1) % self.row_log_interval == 0 or self.should_stop
        if should_log_metrics or self.fast_dev_run:
            # logs user requested information to logger
            metrics = batch_output.batch_log_metrics
            grad_norm_dic = batch_output.grad_norm_dic
            if len(metrics) > 0 or len(grad_norm_dic) > 0:
                self.log_metrics(metrics, grad_norm_dic)

row_log_interval is for how many steps should be skipped before adding a new row to the summary.
log_save_interval is the interval at which this summary is actually written to disk. this is the slowest part, and determines how often you can “refresh” your tensorboard.

log_save_interval may not apply to all loggers if they do their own thing, like sending the data to the cloud.

#3

By summary you mean all the logs from previous steps are stored and at log_save_interval, it saves them all to the disk the clear up summary state of TensorboardLogger? So with this logger if I set row_log_interval=20 and log_save_interval=1000 for total batches of 100, it won’t create any logs, right?

#4

Yes, row_log_interval applies to all logger.
I changed x-axis to step in the wandb plot so it showed from 0, 1, 2, …

So I miss understood it