Description

The devx for training over the API was very broken. A simple snippet like below doesn't work.
Nick uncovered this bug because Pella was bitten. Slack thread

This is one of those cases that just to debug, I had had to refactor first.

What actually solved the bug is:

• adding nocache=true to api calls
• waiting for version to have generating=False AND images>0 to consider it generated. Otherwise you can run into a very likely racing condition where the train endpoint finds a version that still looks empty.

    rf = Roboflow()
    project = rf.workspace("meh3").project("mosquitobao")
    version_number = project.generate_version(
        settings={
            "augmentation": {
                "bbblur": {"pixels": 1.5},
                "image": {"versions": 2},
            },
            "preprocessing": {
                "auto-orient": True,
            },
        }
    )
    # version_number = "61"
    print(version_number)
    version = project.version(version_number)
    model = version.train(
        speed="fast",  # Options: "fast" (default) or "accurate" (paid feature)
        checkpoint=None,  # Use a specific checkpoint to continue training
    )
    print(model)

Type of change

• Maintenance (non-breaking, non-user facing changing such as dependency update, adding test, modifying secrets, etc.)
• Bugfix

How has this change been tested, please provide a testcase or example of how you tested the change?

• Tested locally with the debugme.py file