recipes

Recipes

How to create a project

You can refer to the following documentation. You can also refer to our recipe on how to create a project (run it here)

How to update interface settings

1. Edit new_interface_settings.json.

2. Execute:

python update_interface_settings.py

How to massively add users

1. Edit new_users.yml.

2. Execute:

python add_users.py

Note: You need to have ADMIN right at Organization level to create users.

How to import assets

Check out our recipe on how to import assets (run it here)

How to delete all assets

1. Execute:

python delete_all_assets.py

How to push pre-annotation to existing assets

Check out our recipe on how to import predictions (run it here)

How to export and parse labels

Check out our recipe on how to export labels (run it here)

How to do Named-Entity Recognition with Google

Example based on Enron email dataset. Its downloads the data, get the first 50 emails, pre-label them with Google NLP, push both assets and predictions to Kili and prioritize the assets.

1. Create a NER project and retrieve its ID.

2. Make sure that your GCP authentication is set up properly (or follow this tutorial).

3. Execute:

python google_ner_pre_labeling.py

How to use Python functions to query data

You can query every asset, label, or project-related information through the API. A comprehensive example is our recipe on how to use the query methods (run it here) You can also refer to the documentation.

How to use GraphQL Playground to query data

If you prefer, you can directly query GraphQL API without using kili-playground.

1. Generate an API key in Kili interface in My account, under the tab API KEY. Store it in some place secured.

2. Go to http://cloud.kili-technology.com/api/label/v2/graphql

3. In the bottom left corner of the screen, click on HTTP headers and write the retrieved token in the authorization headers:

{
  "Authorization": "X-API-Key: YOUR_API_KEY"
}

4. Launch any query/mutation:

query {
  users(where: { email: "YOUR_EMAIL" }, first: 10, skip: 0) {
    id
    activated
    email
  }
}

How to do online learning with YOLOv3 (for object detection)

1. Create a project for single-class object detection

2. Update settings to respect YOLOv3's handling of classes (key is an integer starting from zero and value is in lower case)

{
  "jobs": {
    "OBJECT_DETECTION": {
      "mlTask": "OBJECT_DETECTION",
      "content": {
        "input": "radio",
        "categories": {
          "0": {
            "name": "face"
          },
          ...
        }
      },
      "required": true,
      "tools": [
        "polygon"
      ],
      "instruction": "Detect following objects"
    }
  }
}

You can use Kili Playground's recipe update_interface_settings to programmatically update the interface.

3. Build the docker in the folder:

cd image-object-detection-with-yolo
docker build -t kili-playground-yolo .

4. Launch it by setting EMAIL/PASSWORD/PROJECT_ID/API_ENDPOINT environment variables:

docker run -it -e "EMAIL=myemail@kili-technology.com" \
  -e "PASSWORD=my_password" \
  -e "PROJECT_ID=1234567890" \
  -e "JOB_ID=job_12345" \
  -e "API_ENDPOINT=https://cloud.kili-technology.com/api/label/v2/graphql" \
  --network="host" kili-playground-yolo

You can also launch the script manually:

git clone git@github.com:ultralytics/yolov3.git /path/to/yolov3
python main.py --weights /path/to/weights/yolov3.pt \
  --email=myemail@kili-technology.com \
  --password=mypassword \
  --project_id=1234567890 \
  --job_id=job_12345 \
  --api_endpoint=https://cloud.kili-technology.com/api/label/v2/graphql \
  --yolo_path=/path/to/yolov3

5. The script will continuously create predictions on non-labelled assets.

How to import OCR metadata

1. Edit new_assets.yml where metadata has the format of ./examples/invoice.json and the content points to the URL of invoice.png.

2. Execute:

python import_assets.py

How to set assets only to be labeled by some users

1. Edit new_assets.yml and change the field toBeLabeledBy. When this field no exists, assets can be labeled by everyone.

2. Execute:

python set_asset_to_be_labeled_by.py

How to authenticate to the API ?

1. Generate an API key in Kili interface in My account, under the tab API KEY. Store it in some place secured.

2. Start automating your tasks :

from kili.client import Kili

kili = Kili(api_key=api_key)

How to delete one asset identified by its external id ?

Use:

assets = kili.assets(project_id=project_id, external_id_contains=[external_id])
kili.delete_many_from_dataset(asset_ids=[a['id] for a in assets])

How to update instructions in project?

Currently instructions can be set as a link to a PDF or an external web page.

You can update instructions with:

python update_project_instructions.py

How to get labeler statistics?

You can get an excel file summing up the asset / labeler / day doing:

python get_labeler_stats.py