You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I'd like more flexibility in the directory hierarchy used for writing tensorboard data.
Specifically, I want to make the experiment-name directory optional, removing one level in the hierarchy.
Also I'd like to be able to specify the name of the version directory. Currently it is "version_N" but if the constructor version parameter is a string, I think it could just directly use that.
Motivation
You could specify the version on the command line or generate it from the current datetime, which I find more meaningful than "version_73". Likewise if the tensorboard data is not going to be preserved, it's much more readable in the tensorboard gui to skip the experiment name.
Here's the directory layout you could have and how it appears in TensorBoard.
I think the change also cleans up the code a little, removing duplication.
Without experiments directory: TensorBoardLogger(tensorboard_dir, name="")
Without experiments directory and with explicitly named version directory: TensorBoardLogger(tensorboard_dir, name="", version="2020-02-07-220525")
checkpoints/2020-02-07-220525
checkpoints/2020-02-07-220525/events.out.tfevents.1581113127.glia.8438.0
checkpoints/2020-02-07-220525/meta_tags.csv
Current code: TensorBoardLogger(tensorboard_dir)
checkpoints/default
checkpoints/default/version_0
checkpoints/default/version_0/events.out.tfevents.1581112657.glia.8299.0
checkpoints/default/version_0/meta_tags.csv
checkpoints/default/version_1
checkpoints/default/version_1/events.out.tfevents.1581112777.glia.8341.0
checkpoints/default/version_1/meta_tags.csv
🚀 Feature
I'd like more flexibility in the directory hierarchy used for writing tensorboard data.
Specifically, I want to make the experiment-name directory optional, removing one level in the hierarchy.
Also I'd like to be able to specify the name of the version directory. Currently it is "version_N" but if the constructor version parameter is a string, I think it could just directly use that.
Motivation
You could specify the version on the command line or generate it from the current datetime, which I find more meaningful than "version_73". Likewise if the tensorboard data is not going to be preserved, it's much more readable in the tensorboard gui to skip the experiment name.
Pitch
I have a branch with working code + tests
https://github.com/bobkemp/pytorch-lightning/tree/tensorboard-path-generalisation
Alternatives
None
Additional context
The text was updated successfully, but these errors were encountered: