tensorboard.md

Using Tensorboard

Setting up

First, you need a fresh (>= 1.1.0) pytorch version installed, as tensorboard was supported only recently. Tensorboard itself has to be installed separately, see instructions.

Using with EGG

By default, Tensorboard reporting is disabled; to enable it, you need to set the flag --tensorboard when launching your game. Another useful flag sets the path for the output logs, --tensorboard_dir (default is ./runs/). Make sure that different experiments use different directories, otherwise the results would be mixed.

An example command then would be:

python -m egg.zoo.mnist_autoenc.train --tensorboard --tensorboard_dir=./runs/mnist_example/

After training is finished, the logs can be analyzed by running tensorboard and opening its page on a browser (http://127.0.0.1:6006/):

tensorboard --logdir=./runs/ 

Using in your game

To get access to the SummaryWriter instance (see pytorch description),users have to call the egg.core.utils.get_summary_writer() function. Since the writer writes the statistics asynchronously, it is necessary to call egg.core.close() at the end of a program - otherwise the last datapoints might be lost.