Create and explore automated machine learning experiments in the Azure portal (Preview)
In this article, you learn how to create, run, and explore automated machine learning experiments in the Azure portal without a single line of code. Automated machine learning automates the process of selecting the best algorithm to use for your specific data, so you can generate a machine learning model quickly. Learn more about automated machine learning.
If you prefer a more code-based experience, you can also configure your automated machine learning experiments in Python with the Azure Machine Learning SDK.
An Azure subscription. If you don’t have an Azure subscription, create a free account before you begin. Try the free or paid version of Azure Machine Learning service today.
An Azure Machine Learning service workspace. See Create an Azure Machine Learning service workspace.
Navigate to the left pane of your workspace. Select Automated Machine Learning under the Authoring (Preview) section.
If this is your first time doing any experiments with Automated Machine Learning, you'll see the following:
Otherwise, you will see your Automated machine learning dashboard with an overview of all of your automated machine learning experiments, including those created with the SDK. Here you can filter and explore your runs by date, experiment name, and run status.
Create an experiment
Select the Create Experiment button to populate the following form.
Enter your experiment name.
Select a compute for the data profiling and training job. A list of your existing computes is available in the dropdown. To create a new compute, follow the instructions in step 3.
Select the Create a new compute button to open the below pane, and configure your compute context for this experiment.
Field Description Compute name Enter a unique name that identifies your compute context. Virtual machine size Select the virtual machine size for your compute. Additional settings Min node: Enter the minimum number of nodes for your compute. The minimum number of nodes for AML compute is 0. To enable data profiling, you must have 1 or more nodes.
Max node: Enter the maximum number of nodes for your compute. The default is 6 nodes for an AML Compute.
To start the creation of your new compute, select Create. This can take a few moments.
Your compute name will indicate if the compute you select/create is profiling enabled. (See 7b for more details on data profiling).
Select a storage account for your data. Public preview only supports local file uploads and Azure Blob Storage accounts.
Select a storage container.
Select a data file from your storage container, or upload a file from your local computer to the container.
Use the preview and profile tabs to further configure your data for this experiment.
On the Preview tab, indicate if your data includes headers, and select the features (columns) for training using the Included switch buttons in each feature column.
On the Profile tab, you can view the data profile by feature, as well as the distribution, type, and summary statistics (mean, median, max/min, and so on) of each.
The following error message will appear if your compute context is not profiling enabled: Data profiling is only available for compute targets that are already running.
Select the training job type: classification, regression, or forecasting.
Select target column. The column which you would like to do the predictions on.
- Select time column: This column contains the time data to be used.
- Select forecast horizon: Indicate how many time units (minutes/hours/days/weeks/months/years) will the model be able to predict to the future. The further the model is required to predict into the future, the less accurate it will become. Learn more about forecasting and forecast horizon.
(Optional) Advanced settings: additional settings you can use to better control the training job.
Advanced settings Description Primary metric Main metric used for scoring your model. Learn more about model metrics. Exit criteria When any of these criteria are met, the training job ends before full completion.
Training job time (minutes): How long to allow the training job to run.
Max number of iterations: Maximum number of pipelines (iterations) to test in the training job. The job will not run more than the specified number of iterations.
Metric score threshold: Minimum metric score for all pipelines. This ensures that if you have a defined target metric you want to reach, you do not spend more time on the training job than necessary.
Preprocessing Select to enable or disable the preprocessing done by automated machine learning. Preprocessing includes automatic data cleansing, preparing, and transformation to generate synthetic features. Learn more about preprocessing. Validation Select one of the cross validation options to use in the training job. Learn more about cross validation. Concurrency Select the multi-core limits you would like to use when using multi-core compute. Blocked algorithm Select algorithms you want to exclude from the training job.
For more information on fields, click the information tool tip.
You can get a vast variety of summary statistics across your data set to verify whether your data set is ML-ready. For non-numeric columns, they include only basic statistics like min, max, and error count. For numeric columns, you can also review their statistical moments and estimated quantiles. Specifically, our data profile includes:
Feature: name of the column which is being summarized.
Profile: an in-line visualization based on the type inferred. For example, strings, booleans, and dates will have value counts, while decimals (numerics) have approximated histograms. This allows you to gain a quick understanding of the distribution of the data.
Type distribution: an in-line value count of types within a column. Nulls are their own type, so this visualization is useful for detecting odd or missing values.
Type: the inferred type of the column. Possible values include: strings, booleans, dates, and decimals.
Min: the minimum value of the column. Blank entries appear for features whose type does not have an inherent ordering (e.g. booleans).
Max: the maximum value of the column. Like "min," blank entries appear for features with irrelevant types.
Count: the total number of missing and non-missing entries in the column.
Not missing count: the number of entries in the column which are not missing. Note that empty strings and errors are treated as values, so they will not contribute to the "not missing count."
Quantiles (at 0.1, 1, 5, 25, 50, 75, 95, 99, and 99.9% intervals): the approximated values at each quantile to provide a sense of the distribution of the data. Blank entries appear for features with irrelevant types.
Mean: the arithmetic mean of the column. Blank entries appear for features with irrelevant types.
Standard deviation: the standard deviation of the column. Blank entries appear for features with irrelevant types.
Variance: the variance of the column. Blank entries appear for features with irrelevant types.
Skewness: the skewness of the column. Blank entries appear for features with irrelevant types.
Kurtosis: the kurtosis of the column. Blank entries appear for features with irrelevant types.
When configuring your experiments, you can enable the advanced setting
Preprocess. Doing so means that the following data preprocessing and featurization steps are performed automatically.
|Drop high cardinality or no variance features||Drop these from training and validation sets, including features with all values missing, same value across all rows or with extremely high cardinality (for example, hashes, IDs, or GUIDs).|
|Impute missing values||For numerical features, impute with average of values in the column.
For categorical features, impute with most frequent value.
|Generate additional features||For DateTime features: Year, Month, Day, Day of week, Day of year, Quarter, Week of the year, Hour, Minute, Second.
For Text features: Term frequency based on unigrams, bi-grams, and tri-character-grams.
|Transform and encode||Numeric features with few unique values are transformed into categorical features.
One-hot encoding is performed for low cardinality categorical; for high cardinality, one-hot-hash encoding.
|Word embeddings||Text featurizer that converts vectors of text tokens into sentence vectors using a pre-trained model. Each word’s embedding vector in a document is aggregated together to produce a document feature vector.|
|Target encodings||For categorical features, maps each category with averaged target value for regression problems, and to the class probability for each class for classification problems. Frequency-based weighting and k-fold cross validation is applied to reduce over fitting of the mapping and noise caused by sparse data categories.|
|Text target encoding||For text input, a stacked linear model with bag-of-words is used to generate the probability of each class.|
|Weight of Evidence (WoE)||Calculates WoE as a measure of correlation of categorical columns to the target column. It is calculated as the log of the ratio of in-class vs out-of-class probabilities. This step outputs one numerical feature column per class and removes the need to explicitly impute missing values and outlier treatment.|
|Cluster Distance||Trains a k-means clustering model on all numerical columns. Outputs k new features, one new numerical feature per cluster, containing the distance of each sample to the centroid of each cluster.|
Run experiment and view results
To run the experiment, click Start. The experiment preparing process takes a couple of minutes.
View experiment details
Once the experiment preparation phase is done, you'll see the Run Detail screen. This gives you a full list of the models created. By default, the model that scores the highest based on your parameters is at the top of the list. As the training job tries out more models, they are added to the iteration list and chart. Use the iteration chart to get a quick comparison of the metrics for the models produced so far.
Training jobs can take a while for each pipeline to finish running.
View training run details
Drill down on any of the output models to see training run details, like performance metrics and distribution charts. Learn more about charts.
Once you have the best model at hand, it is time to deploy it as a web service to predict on new data.
Automated ML helps you with deploying the model without writing code:
You have a few options for deployment.
If you want to deploy the best model based on the metric criteria you set for the experiment, select Deploy Best Model from the Run Detail page.
If you want to deploy a specific model iteration, drill down on the model to open its specific run detail page and select Deploy Model.
First step is to register the model into the service. Select "Register model" and wait for the registration process to complete.
Once the model is registered, you'll be able to download the scoring script (scoring.py) and the environment script (condaEnv.yml) to be used during deployment.
When the scoring script and the environment script are downloaded, go to the Assets blade of the left navigation pane and select Models.
Select the model you registered, and select "Create image".
You can identify the model by its description, which will include the run ID, iteration number, in the following format: <Run_ID>_<Iteration_number>_Model
Enter a name for the image.
Select the Browse button next to the "Scoring File" box to upload the scoring file (scoring.py) you previously downloaded.
Select the Browse button next to the "Conda File" box to upload the environment file (condaEnv.yml) you previously downloaded.
You can use your own scoring script and conda file, as well as upload additional files. Learn more about scoring script.
File names must be under 32 characters and must begin and end with alphanumerics. May include dashes, underscores, dots, and alphanumerics between. Spaces are not allowed.
Select the "Create" button to start the image creation. This will take a few minutes to complete, once done, you will see a message on the top bar.
Go to the "Images" tab, check the checkbox next to the image you want to deploy, and select "Create deployment". Learn more about deployments.
There are 2 options for deployment.
- Azure Container Instance (ACI) - This is used more for testing purpose rather than operational deployment at scale. Make sure to fill in the values for at least one core for CPU Reserve Capacity, and at least one gigabyte (GB) for Memory Reserve Capacity
- Azure Kubernetes Service (AKS)) - This option is for deployment at scale. You will need to have an AKS based compute ready.
When done, select Create. Deploying the model can take several minutes for each pipeline to finish running.
That's it! You have an operational web service to generate predictions.