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
ELEKTRONN2's current workflow for data visualization during training (preview predictions, training error plots, model graphs) is that everything is written automatically into png files in the save directory at fixed time intervals and with fixed image sizes/display options etc.
A much better way to deal with this data would be to (optionally) make it available from a web server that runs during training and serves visualizations and previews on user demand. This would prevent littering save directories with hundreds of plots and would make visualization much more accessible and extensible.
A possible design inspiration would be TensorBoard.
The recently announced dash framework might be usable for implementing this feature. I am not sure if it will work in our use case without a paid license, though (which would be a no-go).
A probably more mature and less "fancy" alternative to dash is Bokeh. It also looks promising and it has the advantage of being fully free and having an excellent documentation.
The text was updated successfully, but these errors were encountered:
ELEKTRONN2's current workflow for data visualization during training (preview predictions, training error plots, model graphs) is that everything is written automatically into png files in the save directory at fixed time intervals and with fixed image sizes/display options etc.
A much better way to deal with this data would be to (optionally) make it available from a web server that runs during training and serves visualizations and previews on user demand. This would prevent littering save directories with hundreds of plots and would make visualization much more accessible and extensible.
A possible design inspiration would be TensorBoard.
The recently announced dash framework might be usable for implementing this feature. I am not sure if it will work in our use case without a paid license, though (which would be a no-go).
A probably more mature and less "fancy" alternative to dash is Bokeh. It also looks promising and it has the advantage of being fully free and having an excellent documentation.
The text was updated successfully, but these errors were encountered: