[Question] How do I retrieve the outputs of each trials of a sweep job?

[pievalentin]

• Package Name: azure.ai.ml.sweep
• Package Version: 1.0.0
• Operating System: Ubuntu
• Python Version: py38

I am running a Sweep job successfully on AzML and now I want to retrieve the outputs of each
trials for further exploration.
Thing is reading the both the doc and the code I can't see how I am supposed to retrieve them. The SweepJob class does not seem to expose the run like hyperdrive used to.

I would love if any one could give me directions ❤️

[azure-sdk]

Label prediction was below confidence level 0.6 for Model:ServiceLabels: 'Storage:0.1816569,Service Bus:0.09357062,Event Hubs:0.033976667'

[tjprescott]

@bandsina can you answer @pievalentin's question? cc/ @azureml-github

Thank you for your feedback. This has been routed to the support team for assistance.

[luigiw]

@pievalentin you can do ml_client.jobs.list(parent_job_name="sweep-job-name")

[pievalentin]

@luigiw Thank you for answer.
Continuing on your answer, I have the following code:

trial = next(ml_client.jobs.list(parent_job_name="upbeat_cumin_0lq28mbtln"))

mlflow.get_experiment(experiment_id=trial.experiment_name).artifact_location
# >> ''

The artifact_location is missing.
Even when I run

mlflow.artifacts.download_artifacts(run_id =trial.experiment_name,artifact_path='/outputs', dst_path='./training_result/tester.pkl')

Here is the error I have

RestException: RESOURCE_DOES_NOT_EXIST: Response: {'Error': {'Code': 'UserError', 'Severity': None, 'Message': 'Run ddfb3866-c07f-474b-XXXXXXXX was not found', 'MessageFormat': None, 'MessageParameters': None, 'ReferenceCode': None, 'DetailsUri': None, 'Target': None, 'Details': [], 'InnerError': None, 'DebugInfo': None, 'AdditionalInfo': None}, 'Correlation': {'operation': '4e538c0aeb63d3e68863b1f558a5a1f4', 'request': '98cf8a27389de16b'}, 'Environment': 'eastus2', 'Location': 'eastus2', 'Time': '2022-10-20T18:03:30.7030304+00:00', 'ComponentName': 'mlflow', 'error_code': 'RESOURCE_DOES_NOT_EXIST'}

Am I missing something?

PS: I could download the whole output of the sweep like this

ml_client.jobs.download(sweep_job.name,download_path='~/job_ouput')

but I would like to only pick the pickle that interest me and rename the file before pulling it to my local disk. Thats why I want to use a more granular API

[pievalentin]

Ok I guess I won't be able to have the same behavior as the sdk v1. Here is the closest thing I have for retrieving the ouput of a single trial of a sweep job.

trial = next(ml_client.jobs.list(parent_job_name=returned_sweep_job.name))
client = mlflow.MlflowClient()
finished_mlflow_run = client.get_run(trial.name)
client.list_artifacts(finished_mlflow_run.info.run_id)

This should output:

$ [<FileInfo: file_size=-1, is_dir=True, path='outputs'>, <FileInfo: file_size=-1, is_dir=True, path='system_logs'>, <FileInfo: file_size=-1, is_dir=True, path='user_logs'>]

Now you can download the output directory.

mlflow.artifacts.download_artifacts(run_id=finished_mlflow_run.info.run_id, artifact_path ="outputs",dst_path='sweep_trial_outputs')

I am not sure if the mlflow API enable file renaming.

Hope this help someone in the future