-
Notifications
You must be signed in to change notification settings - Fork 154
[Feature request] Just in time evaluation of attributes #368
Copy link
Copy link
Open
Labels
back-endFor issues where the root is mostly occurring on the back-end and not a specific adapterFor issues where the root is mostly occurring on the back-end and not a specific adapterfront-endFor issues where the root is mostly occurring on the front-endFor issues where the root is mostly occurring on the front-endstat:awaiting model-navigatorThe issue is actively being worked on by our navigatorsThe issue is actively being worked on by our navigatorstype:featureFeature requestsFeature requests
Metadata
Metadata
Assignees
Labels
back-endFor issues where the root is mostly occurring on the back-end and not a specific adapterFor issues where the root is mostly occurring on the back-end and not a specific adapterfront-endFor issues where the root is mostly occurring on the front-endFor issues where the root is mostly occurring on the front-endstat:awaiting model-navigatorThe issue is actively being worked on by our navigatorsThe issue is actively being worked on by our navigatorstype:featureFeature requestsFeature requests
Type
Fields
Give feedbackNo fields configured for issues without a type.
Hello,
Writting my plugin, I wanted to add some attributes for the debug such as the tensor values for weight/biases.
However with big models this attribute seems to be a performance killer (potentially because of the json serialization of the graph).
Is it possible to set some attributes with a just-in-time (lazy) evaluation for some attributes? Ideally, the value of such attributes would only be computed when explicitly requested by the user (for example when they click on the node or expand the corresponding section in the UI).
From a plugin developper point of view, it would be equivalent to registering a callable that compute and return the value.
This feature would allow for advanced debugging capabilities without negatively impacting performance in typical workflows.
Regards,
Cyril