Skip to content

TensorBoard logs with Timestamps #708

Open
@merveenoyan

Description

@merveenoyan

I'd love to see if we could have TensorBoard logs as siblings named with timestamps of time they were pushed to Hub, for better experiment tracking. It will look like following:

logs 
> date-time of model run A
>> train for date-time A
>> test for date-time A
> date-time of model run B
>> train for date-time B 
>> test for date-time B 

At some point we could even keep different model configs at the same repository with different timestamps, better UX for people who like to do experiment tracking with Keras (a tree like above but for model files) and widget could work with the last model.
I also wish TB could keep the hyperparameters of models so that people wouldn't have to push models but their configs and tensorboards only so I might open a feature request over there as well. This way people don't have to try to open two repos at the same time and compare their TB logs side by side.
(edit: there's already such feature but I feel like we can implement it by default, see 97
This is particularly problematic though because we don't have full control over model training (like we do in transformers) so we don't have the authority to sign the timestamps. Instead the timestamps could come from the time people push their model+logs, we could organize inside the push and when they pull and push again we organize every time so that it will have the structure like above.

Metadata

Metadata

Assignees

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions