When i use machine learning, it will call twice, and the result repeat twice

[fanqie03]

version: machine learning config:

action :

first interactive call first interactive result

second interactive call second interactive result

third call third result

fourth call fourth result

use default mmdetection examples

[fanqie03]

How to use machine learning? The result is a little strange. Did I use it wrong?

[KonstantinKorotaev]

Hi @mengfu188 You used it right. It seems that duplicate results are connected with double calls of interactive annotation - our front end merges all results. I have asked our front end team to check this issue. Meanwhile, if you don't want to wait the fix you can add filter for duplicate results in ML backend.

[fanqie03]

Hello @KonstantinKorotaev Another question is if i use bicycle tag to auto rectangle. The front end will turn all labels into bicycles. But the back end gives different labels. I use this example https://github.com/heartexlabs/label-studio-ml-backend/tree/master/label_studio_ml/examples/mmdetection

How to use machine learning? The result is a little strange. Did I use it wrong?

[andreaswimmer]

I have the exact same problem. The browser calls interactive-annotating twice and all the labels are added twice.

I am running the latest version:

  "release": "1.6.0",
  "label-studio-os-package": {
    "version": "1.6.0",
    "short_version": "1.6",
    "latest_version_from_pypi": "1.6.0",
    "latest_version_upload_time": "2022-09-27T17:09:34",
    "current_version_is_outdated": false
  },

  "label-studio-os-backend": {
    "message": "Bump rc5",
    "commit": "b0df5dd19a5400e8d26fb8269e2add87799af055",
    "date": "2022-09-27 16:32:57 +0100",
    "branch": "HEAD",
    "version": "1.6.0+0.gb0df5dd.dirty"
  },

  "label-studio-frontend": {
    "message": "fix: DEV-3510: fix video regions w/o label",
    "commit": "5d4231c2a8b9703f8c43abdf62948ac892d6f742",
    "branch": "master",
    "date": "2022/09/27 13:58:00"
  },

  "dm2": {
    "message": "feat: DEV-3373: Enable video preview (#106)  * Enable video preview  ...",
    "commit": "8374cbfeceb5aa9870278fc5371f0585d452c111",
    "branch": "master",
    "date": "2022-09-27T08:11:03Z"
  },

  "label-studio-converter": {
    "version": "0.0.44"
  },

  "label-studio-ml": {
    "version": "1.0.8.rc3"
  }
}

[andreaswimmer]

If anybody else has the same problem, as a workaround, I cache the last request and return empty predictions if the current one is equal to the last one. This technically creates race condition, if two users initiate a prediction, but my deployment is small enough to accept that risk.

    last_tasks = None

    def predict(self, tasks: List[dict], **kwargs):
        # https://github.com/heartexlabs/label-studio/issues/2546
        if tasks == self.last_tasks:
            logging.warning("Double request, returning 0 predictions.")
            self.last_tasks = None
            return []
        self.last_tasks = tasks
        # actual prediction starts here