Event banner
Event details
- Good luck and we wait for the event
- The event has already passed! You can read the summary linked above to see all the questions and answers.
nswitanekMicrosoft
What are best practices for deciding whether or how to configure a new automl run, given that you've already run one? To effectively review all the child runs from several automl runs, does the AML studio UI provide all functionality or are there useful helper scripts for collecting metrics via SDK that enable more effective review of runs?- cesardlYou can review AutoML Jobs/Runs in the Azure ML Studio UI the same way than when using AutoML v1. Basically, the backend is the same, hence, the UI for reviewing the Jobs/Runs is also shared across v1 and v2.
- Thanks everyone for joining and participating in our AMA about Automated ML! We'll be posting a summary of questions and answers in this space soon.
- Just 11 minutes to go! Get your questions in!
- How many employees are required for at least 1 ML in my company?
- nswitanekHello! This question is out of scope for this particular AMA event, which is specifically around Automated ML. We'd recommend posting this question in the Machine Language space on Tech Community: https://techcommunity.microsoft.com/t5/ai-machine-learning/bd-p/MachineLearning
- nswitanekWhat guidance can you provide on transitioning to automl with SDK v2? What are the installation steps to make sure all the package dependencies for azureml automl are all updated appropriately? What gotchas should a user of the new SDK be aware of when coming to it from SDK v1?
- cesardl
See reply to this for details.
- cesardl
Hi Nick, AutoML Python SDK uses completely new Python packages, and the way you do it is by installing in a Conda environment the Azure ML Python SDK v2 packages, explained here:
https://github.com/Azure/azureml-examples/tree/main/sdk
Basically, it's simply this pip command on top of your activated conda environment:
pip install --pre azure-ai-ml
Gotchas / new things:
- By default, models in v2 are saved as MLFlow models (MLFlow model) but you can also deploy it as a regular .pkl file model as in v1. In reality, the models created by AutoML in v1 and v2 are "the same" because AutoML backend is the same. What's new in v2 is the Python SDK and CLI.
- CLI is completely new for AutoML, since AutoML was not part of Azure ML CLI v1, previously.
- Instead of using Tabular Datasets (v1) you use MLTable (v2) which you currently define in a .YAML file pointing to your data (i.e. a .CSV file). In the next Public Preview you will also have Python functions to create an MLTable from your data files using "data plane" functions for MLTable.
- I have two questions. 1) Is it the same ML than AI or Cognitive Services? 2) If I want to detech objects in a Film because the source is a Film and not Pictures, Can I use AI Builder too?
- SwatiGharseHI Jorge, the products target different personas when building computer vision models. Cognitive services targets a developer persona and allows users to easily train models that can be accessed via APIs. AzureML's AutoML solution targets data scientists / ML Pros and allows them additional control around model training and model deployment. AI Builder is a no code experience that allows you to use intelligence within Power Apps.
- Thank you so much Swati Gharse now I need to left for an presential event. thank you so much for your answer
FeidowuIs there a Teams Meeting-Link for the session? All I have is the link below and it shows that the session is Live but I can't see to Join: https://techcommunity.microsoft.com/t5/artificial-intelligence-machine/automated-ml-ama-ask-microsoft-anything/ec-p/3310467#M1- Yes, this whole event is text-based and takes place in the space you're at. If you have any questions, please post them!
- AMAs are text based.
- I would like to see an explanation of the relationship between quota and Azure ML. What vms are best for a small, medium and large assisted labeling jobs? What vms are best for small, medium and large Object Detection Model training?
- SwatiGharseAutomated ML models for computer vision tasks require GPU SKUs and support NC and ND families. We recommend the NCsv3-series (with v100 GPUs) for faster training. A compute target with a multi-GPU VM SKU leverages multiple GPUs to also speed up training. Additionally, when you set up a compute target with multiple nodes you can conduct faster model training through parallelism when tuning hyperparameters for your model.
- Thanks Swati, I work for a startup. We do not need the fastest training. We are willing to wait. What we need is the most cost efficient SKU. Do you have a recommendation for most cost efficient sku for Assisted Labeling and Object Detection Model training?
- Is there MlOps for Automated ML? Where would be the best place to start?
- cesardlMLOps for AutomatedML is practically the same MLOps approaches for the general Azure ML. The only difference is that with AutoML SDK 1 you use the AutoMLStep in Azure ML pipelines (SDK 1) or using AutoML Jobs functions in Pipelines (SDK 2), but the rest of the approaches are the same. You can start here: https://docs.microsoft.com/en-us/azure/cloud-adoption-framework/manage/mlops-machine-learning
- cesardlAlso, see this example notebook using AutoMLStep in Pipelines (Python SDK 1): https://github.com/Azure/azureml-examples/tree/main/python-sdk/tutorials/automl-with-azureml/continuous-retraining
