You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In this how-to, you learn how to use the designer to train a model and set up a batch prediction pipeline and web service. Batch prediction allows for continuous and on-demand scoring of trained models on large data sets, optionally configured as a web service that can be triggered from any HTTP library.
18
+
In this article, you learn how to use the designer to create a batch prediction pipeline. Batch prediction lets you continuously score large datasets on-demand using a web service that can be triggered from any HTTP library.
20
19
21
-
For setting up batch scoring services using the SDK, see the accompanying [how-to](how-to-use-parallel-run-step.md).
22
-
23
-
In this how-to, you learn the following tasks:
20
+
In this how-to, you learn to do the following tasks:
24
21
25
22
> [!div class="checklist"]
26
-
> * Create a basic ML experiment in a pipeline
27
-
> * Create a parameterized batch inference pipeline
28
-
> * Manage and run pipelines manually or from a REST endpoint
23
+
> * Create and publish a batch inference pipeline
24
+
> * Consume a pipeline endpoint
25
+
> * Manage endpoint versions
26
+
27
+
To learn how to set up batch scoring services using the SDK, see the accompanying [how-to](how-to-run-batch-predictions.md).
29
28
30
29
## Prerequisites
31
30
32
-
1. If you don’t have an Azure subscription, create a free account before you begin. Try the [free or paid version of the Azure Machine Learning](https://aka.ms/AMLFree).
31
+
This how-to assumes you already have a training pipeline. For a guided introduction to the designer, complete [part one of the designer tutorial](tutorial-designer-automobile-price-train-score.md).
33
32
34
-
1. Create a [workspace](tutorial-1st-experiment-sdk-setup.md).
33
+
##Create a batch inference pipeline
35
34
36
-
1. Sign in to [Azure Machine Learning studio](https://ml.azure.com/).
35
+
Your training pipeline must be run at least once to be able to create an inferencing pipeline.
37
36
38
-
This how-to assumes basic knowledge of building a simple pipeline in the designer. For a guided introduction to the designer, complete the [tutorial](tutorial-designer-automobile-price-train-score.md).
37
+
1. Go to the **Designer** tab in your workspace.
39
38
40
-
## Create a pipeline
39
+
1. Select the training pipeline that trains the model want to use to make prediction.
41
40
42
-
To create a batch inference pipeline, you first need a machine learning experiment. To create one, navigate to the **Designer**tab in your workspace and create a new pipeline by selecting the **Easy-to-use prebuilt modules** option.
45
44
46
-
The following is a simple machine learning model for demonstration purposes. The data is a registered Dataset created from the Azure Open Datasets diabetes data. See the [how-to section](how-to-create-register-datasets.md#create-datasets-with-azure-open-datasets) for registering Datasets from Azure Open Datasets. The data is split into training and validation sets, and a boosted decision tree is trained and scored. The pipeline must be run at least once to be able to create an inferencing pipeline. Click the **Run** button to run the pipeline.
45
+
Now that the training pipeline has been run, you can create a batch inference pipeline.
To create predictions on new data, you can either manually connect a different dataset in this pipeline draft view or create a parameter for your dataset. Parameters let you change the behavior of the batch inferencing process at runtime.
58
+
59
+
In this section, you create a dataset parameter to specify a different dataset to make predictions on.
60
+
61
+
1. Select the dataset module.
62
+
63
+
1. A pane will appear to the right of the canvas. At the bottom of the pane, select **Set as pipeline parameter**.
64
+
65
+
Enter a name for the parameter, or accept the default value.
66
+
67
+
## Publish your batch inferencing pipeline
68
+
69
+
Now you're ready to deploy the inferencing pipeline. This will deploy the pipeline and make it available for others to use.
70
+
71
+
1. Select the **Publish** button.
72
+
73
+
1. In the dialog that appears, expand the drop-down for **PipelineEndpoint**, and select **New PipelineEndpoint**.
74
+
75
+
1. Provide an endpoint name and optional description.
76
+
77
+
Near the bottom of the dialog, you can see the parameter you configured with a default value of the dataset ID used during training.
78
+
79
+
1. Select **Publish**.
80
+
81
+
51
82
52
-
Now that the pipeline has been run, there is a new option available next to **Run** and **Publish** called **Create inference pipeline**. Click the dropdown and select **Batch inference pipeline**.
The result is a default batch inference pipeline. This includes a node for your original pipeline experiment setup, a node for raw data for scoring, and a node to score the raw data against your original pipeline.
86
+
Now, you have a published pipeline with a dataset parameter. The pipeline will use the trained model created in the training pipeline to score the dataset you provide as a parameter.
You can add other nodes to change the behavior of the batch inferencing process. In this example, you add a node for randomly sampling from the input data before scoring. Create a **Partition and Sample** node and place it between the raw data and scoring nodes. Next, click on the **Partition and Sample** node to gain access to the settings and parameters.
90
+
In this section, you will set up a manual pipeline run and alter the pipeline parameter to score new data.
1. After the deployment is complete, go to the **Endpoints** section.
63
93
64
-
The *Rate of sampling* parameter controls what percent of the original data set to take a random sample from. This is a parameter that will be useful to adjust frequently, so you enable it as a pipeline parameter. Pipeline parameters can be changed at runtime, and can be specified in a payload object when rerunning the pipeline from a REST endpoint.
94
+
1. Select **Pipeline endpoints**.
65
95
66
-
To enable this field as a pipeline parameter, click the ellipses above the field and then click **Add to pipeline parameter**.
This screen shows all published pipelines published under this endpoint.
73
103
74
-
## Deploy batch inferencing pipeline
104
+
1. Select the pipeline you published.
75
105
76
-
Now you are ready to deploy the pipeline. Click the **Deploy** button, which opens the interface to set up an endpoint. Click the dropdown and select **New PipelineEndpoint**.
106
+
The pipeline details page shows you a detailed run history and connection string information for your pipeline.
107
+
108
+
1. Select **Run** to create a manual run of the pipeline.
1. Change the parameter to use a different dataset.
113
+
114
+
1. Select **Run** to run the pipeline.
79
115
80
-
Give the endpoint a name and optional description. Near the bottom you see the `sample-rate` parameter you configured with a default value of 0.8. When you're ready, click **Deploy**.
You can find information on how to consume pipeline endpoints and published pipeline in the **Endpoints** section.
83
119
84
-
## Manage endpoints
120
+
You can find the REST endpoint of a pipeline endpoint in the run overview panel. By calling the endpoint, you are consuming its default published pipeline.
85
121
86
-
After deployment is complete, go to the **Endpoints**tab and click the name of the endpoint you just created.
122
+
You can also consume a published pipeline in the **Published pipelines**page. Select a published pipeline and find the REST endpoint of it.
This screen shows all published pipelines under the specific endpoint. Click on your inferencing pipeline.
126
+
To make a REST call, you will need an OAuth 2.0 bearer-type authentication header. See the following [tutorial section](tutorial-pipeline-batch-scoring-classification.md#publish-and-run-from-a-rest-endpoint) for more detail on setting up authentication to your workspace and making a parameterized REST call.
The pipeline details page shows you detailed run history and connection string information for your pipeline. Click the **Run** button to create a manual run of the pipeline.
130
+
The designer assigns a version to each subsequent pipeline that you publish to an endpoint. You can specify the pipeline version that you want to execute as a parameter in your REST call. If you don't specify a version number, the designer will use the default pipeline.
When you publish a pipeline, you can choose to make it the new default pipeline for that endpoint.
97
133
98
-
In the run setup, you can provide a description for the run, and change the value for any pipeline parameters. This time, rerun the inferencing pipeline with a sample rate of 0.9. Click **Run** to run the pipeline.
You can also set a new default pipeline in the **Published pipelines** tab of your endpoint.
101
137
102
-
The **Consume** tab contains the REST endpoint for rerunning your pipeline. To make a rest call, you will need an OAuth 2.0 bearer-type authentication header. See the following [tutorial section](tutorial-pipeline-batch-scoring-classification.md#publish-and-run-from-a-rest-endpoint) for more detail on setting up authentication to your workspace and making a parameterized REST call.
0 commit comments