DocumentationRecipesReferenceGraphQLChangelog
Log In

May 16th, 2023

Changelog (v2.133.0)

New

  • [Labeling] Interactive segmentation has been improved. It now uses the top-tier SAM model which outputs masks of much higher quality. We’ve also implemented a caching system for image embeddings. What this means is: getting the first interactively-generated mask may take a few seconds, but any subsequent masks will take much shorter to compute.

  • [Python SDK] We’ve added a new method get_plugin_build_errors . You can use it to retrieve a paginated list of errors that occurred for a specific plugin. For details, refer to our Python SDK documentation.

  • [Python SDK] Extracting label data from the jsonResponse field is not an easy task. That’s because it is a complex, nested dictionary. This is why we’ve developed a label parser that extracts label data in a more convenient way. With label parsing you can now replace complex queries like this one:
    my_label["jsonResponse"]["jobs"]["MY_JOB_NAME"]["categories"][0]["name"]
    with short, easily-understandable ones like this one:
    my_label.jobs["MY_JOB_NAME"].category.name

    Additionally, the new ParsedLabel class enables your IDE to explore the available attributes during development.

For details, refer to our SDK documentation.

You can also follow the programmatic tutorial that we’ve prepared for you.

Changed

  • [Labeling] When the project ontology does not contain any jobs with required status, the Show incomplete jobs filter is no longer visible in the labeling interface. Fewer unnecessary UI items means enhanced focus on important tasks.
  • [Labeling] In number input fields, it is now possible to enter negative numbers and numbers with an exponent (e.g. 5e7) for scientific notations.
  • [Python SDK] kili.assets() now has two parameters external_id_in and external_id_strictly_in, that allow to filter the assets using their external ids. For details, refer to our SDK documentation.

Fixed

  • [Labeling] In previous Kili versions, when users attempted to change a class directly from a pop-up menu in the labeling interface, sometimes not all of the available values were being shown in the dropdown list. We’ve fixed this issue.
  • [Export] Exporting labeled assets used to malfunction if class filters were applied to the asset queue. We’ve fixed this issue.
  • [Python SDK] In projects that use assets located in external cloud storage, requesting an export into the COCO or Pascal VOC formats used to fail with a generic error. We’ve fixed this: the updated error points you to the root cause of the issue: no access to the asset file.