How to specify the automl sdk version to be used?

[robert4os]

Is there a way to specify the automl version (i.e. via Conda dependencies)?

I am getting different results (even errors) because of different automl versions beeing used depending on the compute target (e.g. Training Cluster and local Compute Instances).

---

Document Details

⚠ Do not edit this section. It is required for docs.microsoft.com ➟ GitHub issue linking.

• ID: 0bc2b21e-6b1a-cb94-2857-147177a29d7c
• Version Independent ID: d14620a6-a2f6-49f1-632e-73903d41de8c
• Content: azureml.train.automl.automlconfig.AutoMLConfig class - Azure Machine Learning Python
• Content Source: AzureML-Docset/stable/docs-ref-autogen/azureml-train-automl-client/azureml.train.automl.automlconfig.AutoMLConfig.yml
• Service: machine-learning
• Sub-service: core
• GitHub Login: @j-martens
• Microsoft Alias: jmartens

[ram-msft]

@robert4os Thanks for the question. We are investigating the issue and will update you shortly.

[ram-msft]

@robert4os Can you please add more details about the errors that you are getting.
Please follow the below to specify the automl version or you can use the azureml-sdk[automl] == 1.0.85.
Following is the example notebook for AutoMLStep.
aml_conda_packages = ['pandas','scikit-learn','pyarrow','pytz','psutil']
aml_pip_packages = ['azureml-sdk', 'azureml-dataprep', 'azureml-train-automl==1.0.85']
aml_run_config.environment.python.conda_dependencies = CondaDependencies.create(
conda_packages = aml_conda_packages,
pip_packages = aml_pip_packages,
pin_sdk_version=False)

[robert4os]

Hi Ram, thank you for your response. I am using currently using automl_config = AutoMLConfig(training_data=train_data, **automl_settings) experiment = Experiment(ws, experiment_name) run = experiment.submit(automl_config, show_output=verbose) and there I did not find an option to specify the azurml[ml]-sdk to use on the target. Are you suggesting to wrap the call above in a Script and execute via PythonScriptStep ? Is there a doc or notebook to demonstrate this? Especially using it to execute Automl? The errors I found out, where unrelated to the version. Still it woul be very helpful to have control over the SDK version even for Automl calls like above. Thank you and best regards, Robert

…

On Fri, Mar 6, 2020 at 9:29 AM Ram ***@***.***> wrote: @robert4os <https://github.com/robert4os> Can you please add more details about the errors that you are getting. Please follow the below to specify the automl version. aml_conda_packages = ['pandas','scikit-learn','pyarrow','pytz','psutil'] aml_pip_packages = ['azureml-sdk', 'azureml-dataprep', 'azureml-train-automl==1.0.85'] aml_run_config.environment.python.conda_dependencies = CondaDependencies.create( conda_packages = aml_conda_packages, pip_packages = aml_pip_packages, pin_sdk_version=False) registration_input = model_data model_registration_step = PythonScriptStep(name='Register Model', script_name='register.py', compute_target=aml_compute, source_directory=scripts_folder, arguments=['--model_name', aml_trained_model_name, '--model_path', model_data, '--model_path_folder', aml_trained_model_folder, '--model_date_path', string_pipeline_param, '--runconfig',aml_run_config, '--trained_model_name', aml_trained_model_name ], inputs=[model_data], allow_reuse=False — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#842?email_source=notifications&email_token=AAMXJSUFRJG7QHKUA54HSBTRGCYADA5CNFSM4LCG6MW2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEOAQQ3I#issuecomment-595658861>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAMXJSXVV7MIEOSASMYTM4DRGCYADANCNFSM4LCG6MWQ> .

[swatig007]

@robert4os , the automl version installed locally is the one that gets used for remote training.
You can control the automl version installed locally using this example -
pip install azureml-train-automl=="1.0.85"

[ram-msft]

@robert4os We will now proceed to close this thread. If there are further questions regarding this matter, please tag me in your reply. We will gladly continue the discussion and we will reopen the issue.