Load a model into an existing dataset
complete
P
PerceptiLabs
This feature builds on top of this one: https://perceptilabs.canny.io/feature-requests/p/allow-for-re-use-of-data-and-order-models-based-on-what-dataset-they-were-create
This would make loading shared models easier by letting a model be loaded straight into an existing dataset, avoiding the need to change the dataset path.
It would do so by adding this button:
R
Robert Lundberg
complete
A first experimental version of this has been added and will be released this week.
Improvements to it will come early next year after some underlying structure has been improved.
R
Robert Lundberg
in progress
R
Robert Lundberg
planned
R
Robert Lundberg
Merged in a post:
Enable easy importing of models
P
PerceptiLabs
Currently, a model can be shared with another person, but it's clunky to change what dataset it uses.
This feature will add the ability to use a dataset that's already listed in the tool as the new dataset for the model.
It will also have warnings if the new dataset does not match up with the old one.
R
Robert Lundberg
under review
Julian Moore
The suggestion to allow components to be grouped, shared etc. via a PL repository would effectively achieve the same thing with fewer limitations and greater flexibility (maybe ;) )
If trained components/sub-graphs can be shared, then the amount of data involved would increase a lot, but even giant models have "only" 10^9, 10^10 parameters and since movies of ~10+GB are streamed now 10GB of data for a trained model doesn't seem so much.
R
Robert Lundberg
Julian Moore: True, grouping and sharing components would accomplish close to the same thing as sharing a model, with exception to pre-processing settings :)
Sharing a trained sub-graph is another interesting topic that may prove slightly trickier than sharing a full model, considering that some cutting and pasting in the checkpoint may be needed.
The divider between sharing a component/group and sharing a model from a UX perspective here would be:
* The model can instantly be applied to data and ran, either as trained or untrained.
* The component/group can bee added inside an existing model to be connected up to the rest of the graph.