How to migrate Modern(Active Learning)Document Understanding Skill to other Orchestrator

[rahulk11393 (rahul_@)]

Hi team,

I’ve been using Modern Document Understanding and deployed one skill there.
Is it possible to migrate this setup to another Orchestrator like we do for Classic ML Skill (AI Center) approach?

Thanks,
Rahul

[Baskar_Gurumoorthy]

Hi @rahulk11393

Please refer the below doc.

Document Understanding - Migrating existing projects

this below thread helps to migrate from dev to prod

How to migrate AI Center project from DEV to PROD?

What are the steps required to migrate AI Center project(ML Skill, data set…) from DEV tenant to PROD so that all the training provided to the model in DEV should also reflect in PROD as well? I tried to find the proper answer but didn’t find any.

Hope it helps!!!

[rahulk11393 (rahul_@)]

@Baskar_Gurumoorthy,

The article you shared is for the Classic design (AI Center) and my ask is for the (Active Learning) Modern Document Understanding project. Please check below snapshot:

[narender.bn (Nk)]

any solution on this?

[felipe.restrepo]

According to UiPath documentation:

Document Understanding - Automations in Document Understanding™

This link proposes a way to export the annotated docs coming from Action Center (human feedback) in modern du projects, but it doesn’t explain how to use the exported annotated docs to retrain (fine-tuning) the modern du model:

How To Implement A Document Understanding Training Loop On Cross Platform Activities or Studio Web

The new Document Understanding Cross Platform created for use in Studio Web and also available in Studio Desktop simplify a huge amount of work we need to do with in a Document Understanding project and also give us easier objects to work with as outputs and are well worth using over the legacy versions, however it is currently missing a key feature, the training loop to retrain your Machine Learning Model. In collaboration with UiPath we have figured out a workaround for this by leveraging the…

I hope that sooner than later UiPath modern du projects can use the human feedback to retrain the model automatically, but currently is not as easy as we expect. Other alternative could be:

1. When the human validation is needed, send the document to the action center
2. Create a copy of the original document in a storage location (shared drive, storage bucket, etc)
3. Get the human feedback and export the validated data for your use

Then periodically:

1. Get the original documents copied in the storage location, select the documents to be used for fine-tuning (remember to keep balanced datasets) and upload them in the modern du project
2. Annotate (again) the new uploaded docs (active learning will do the wrong prediction) and wait for the auto retraining to complete
3. Once the training is completed, deploy a new version of the du project
4. Update your code in Studio to point out to the new modern project version (this step will require for you to recreate the out_DocumentData variable because it changes with the new du project version)
5. Publish the a new version of the process package from studio

In summary, the modern du experience is not that good (yet) for projects that requires continuous fine-tuning