How to create a custom label dataset by feeding manifest programmatically

Hello,

Of course there is a way! ;)

I tried to do that 1 or 2 months ago and I've just written a blog post on this with OpenImages dataset as example.

You can check it here https://www.legiasquad.com/articles/retraining-aws-rekognition-with-custom-labels-on-an-annotated-dataset/

There is also a code sample that I used to generate this :)

I hope this helps!

Hi, I have a manifest built from my dataset. The ask is to be able to run a script so that bboxes are created from the manifest prior to training, and without needing to interact with any GUI, e.g. clicking on the radio button "Import images labeled by SageMaker Ground Truth". Thanks all the same.

Hello

If you have a manifest (with or without bounding boxes) in the prescribed format, you can kick off a training without the console by using the CLI via the create-project-version command:
https://docs.aws.amazon.com/cli/latest/reference/rekognition/create-project-version.html

You can also use the SDK to programmatically kick off a training using the CreateProjectVersion API: https://docs.aws.amazon.com/rekognition/latest/dg/API_CreateProjectVersion.html

Could you please elaborate on "script so that bboxes are created from the manifest prior to training"?

For now, I need to interact with any GUI, e.g. clicking on the radio button "Import images labeled by SageMaker Ground Truth" and specify the manifest link in bucket. Once I submit that through the GUI, I can see that bounding boxes are drawn on the image in the console based on the bbox coordinates in the manifest. Then I click on Train button, the training will start.

I want to avoid these manual steps using the GUI. If I can pass the manifest file and kick off the training using that , I will return here and mark this question as answered.

It seems it does, from reading this example: https://docs.aws.amazon.com/rekognition/latest/customlabels-dg/tm-sdk.html

thanks.

Following the SDK, using CreateProjectVersion, I can kick off training and implicitly creating the datasets by supplying two manifests, one for training and another for testing.

Also want to credit sdoloris. His websites provides a tutorial for creating the manifest that is helpful. Marking my question answered. Thank you.

Edited by: mymingle on Mar 17, 2020 4:03 PM

Great - please feel free to reach out if you have any other questions !