All experiments were written as one (without provided tags)

tensorboard
1

Hi, I wanted to compare several runs that I did in the loop creating new SummaryWriter instances like this:

for experiment_name in experiments:
    logger = SummaryWriter(self._log_path, comment=experiment_name)
    ...
    for epoch in range(5):
        ...
        logger.add_scalar("Epoch Loss", loss, epoch)
        ...
    logger.close()

In the log path I got several files like this:

events.out.tfevents.1609675249.nlp-vm.13735.0
events.out.tfevents.1609679736.nlp-vm.13735.1
events.out.tfevents.1609687200.nlp-vm.13735.2
events.out.tfevents.1609691662.nlp-vm.13735.3
events.out.tfevents.1609699158.nlp-vm.13735.4
events.out.tfevents.1609703743.nlp-vm.13735.5
events.out.tfevents.1609711308.nlp-vm.13735.6
events.out.tfevents.1609716054.nlp-vm.13735.7

But the Tensorboard displays all runs as one:

image

image

What I should do to fix it and can I do it without re-running all experiments?

2

Found: the events files that I have in the one folder should be in the separate and the folder name will be displayed as an experiment name.
Also found the important note in the SummaryWriter docstring:

comment (string): Comment log_dir suffix appended to the default
``log_dir``. If ``log_dir`` is assigned, this argument has no effect.