[BUG] Creating a task requires an estimation procedure ID, but procedure is not correctly created

[L-Agbaria]

It is important to first note that the tasks/task.py appears to be slightly different from the version obtainable through pip install "openml==0.15.1", which would've elevated the priority while reducing the effort had the GitHub version not experienced the same bug.

The create_task() function requires an estimation_procedure_id parameter despite it having a default value in all cases, which all funnel towards the same OpenMLTask class where the handling of estimation procedure appears to break*.
It was tested without publication on the main server, as well as with publication on a local docker instance (test server does not have any procedures), where in both cases the create_task() function creates a task with Estimation Procedure equal to None and no data split (errors follow).

import openml
from openml.tasks import TaskType

openml.config.server = "..."
openml.config.apikey = "..."

did = 128  # iris

try:
    my_task = openml.tasks.create_task(
        task_type=TaskType.SUPERVISED_CLASSIFICATION,
        dataset_id=did,
        target_name="class",
        evaluation_measure="predictive_accuracy",
        estimation_procedure_id=1,
    )
    my_task.publish()
except openml.exceptions.OpenMLServerException as e:
    # Error code for 'task already exists'
    if e.code == 614:
        # Lookup task
        tasks = openml.tasks.list_tasks(data_id=did, output_format='dataframe')
        tasks = tasks.query(
            'task_type == "Supervised Classification" '
            'and estimation_procedure == "10-fold Crossvalidation" '
            'and evaluation_measures == "predictive_accuracy"'
        )
        display(tasks)
        task_id = tasks.loc[:, "tid"].values[0]
        print("Task already exists. Task ID is", task_id)
    
print(my_task)
print(my_task.estimation_procedure)
print(my_task.get_split_dimensions())

* Unconfirmed, however it is a good starting point.

[RugvedBane]

Hi, I've looked into this issue.

The root cause appears to be in create_task() in openml/tasks/functions.py. When building the task object, it only passes estimation_procedure_id to the constructor, but OpenMLTask.init also expects estimation_procedure_type, estimation_parameters, and data_splits_url, these are never populated, leaving the task object incomplete.

When a task is fetched from the server (_create_task_from_xml), all four fields are correctly extracted from the XML response. create_task() needs similar logic to fetch and populate these fields using the given estimation_procedure_id.

I'd like to work on a fix for this. Could I be assigned?

[RugvedBane]

After looking into the code, I've found the root cause.

The issue is in _parse_publish_response() in openml/tasks/task.py. After publishing, it only saves task_id from the server response it never populates estimation_procedure_type, estimation_parameters, and data_splits_url. So the task object remains incomplete after publish().

The fix would be to fetch the complete task from the server using the returned task_id right after publishing, then populate the missing fields from it. This is minimal only 4 lines added, no new API endpoints needed.

Happy to submit a PR if I can be assigned.