Creating and deploying a Custom Vision predictive service

Hong Ooi

The basic idea behind Custom Vision is to take a pre-built image recognition model supplied by Azure, and customise it for your needs by supplying a set of images with which to update it. All model training and prediction is done in the cloud, so you don’t need a powerful machine. Similarly, since you are starting with a model that has already been trained, you don’t need a very large dataset or long training times to obtain good predictions (ideally). This vignette walks you through the process of creating and deploying a Custom Vision predictive service.

Creating the resources

You can create the Custom Vision resources using the AzureRMR framework for interacting with Resource Manager. Note that Custom Vision requires at least two resources to be created: one for training, and one for prediction. The available service tiers for Custom Vision are F0 (free, limited to 2 projects for training and 10k transactions/month for prediction) and S0.

rg <- AzureRMR::get_azure_login("yourtenant")$

res <- rg$create_cognitive_service("mycustvis",
    service_type="CustomVision.Training", service_tier="S0")
pred_res <- rg$create_cognitive_service("mycustvispred",
    service_type="CustomVision.Prediction", service_tier="S0")


Custom Vision defines two different types of endpoint: a training endpoint, and a prediction endpoint. Somewhat confusingly, they can both share the same hostname, but use different paths and authentication keys. To start, call the customvision_training_endpoint function with the service URL and key.

url <- res$properties$endpoint
key <- res$list_keys()[1]

endp <- customvision_training_endpoint(url=url, key=key)

Creating the project

Custom Vision is organised hierarchically. At the top level, we have a project, which represents the data and model for a specific task. Within a project, we have one or more iterations of the model, built on different sets of training images. Each iteration in a project is independent: you can create (train) an iteration, deploy it, and delete it without affecting other iterations.

You can see the projects that currently exist on the endpoint by calling list_projects. This returns a named list of project objects:

Azure Custom Vision project 'general_compact' (304fc776-d860-490a-b4ec-5964bb134743)
  Domain: classification.general.compact (0732100f-1a38-4e49-a514-c9b44c697ab5)
  Export target: standard
  Classification type: Multiclass

Azure Custom Vision project 'general_multilabel' (c485f10b-cb54-47a3-b585-624488335f58)
  Domain: classification.general (ee85a74c-405e-4adc-bb47-ffa8ca0c9f31)
  Export target: none
  Classification type: Multilabel

Azure Custom Vision project 'logo_obj' (af82557f-6ead-401c-afd6-bb9d5a3b042b)
  Domain: object_detection.logo (1d8ffafe-ec40-4fb2-8f90-72b3b6cecea4)
  Export target: none
  Classification type: NA

There are three different types of projects, as implied by the list above:

The functions to create these projects are create_classification_project (which is used to create both multiclass and multilabel projects) and create_object_detection_project. Let’s create a classification project:

testproj <- create_classification_project(endp, "testproj", export_target="standard")
Azure Custom Vision project 'testproj' (db368447-e5da-4cd7-8799-0ccd8157323e)
  Domain: classification.general.compact (0732100f-1a38-4e49-a514-c9b44c697ab5)
  Export target: standard
  Classification type: Multiclass

Here, we specify the export target to be standard to support exporting the final model to one of various standalone formats, eg TensorFlow, CoreML or ONNX. The default is none, in which case the model stays on the Custom Vision server. The advantage of none is that the model can be more complex, resulting in potentially better accuracy. The type of project is multiclass classification, and the domain (the initial model used as the basis for training) is general. Other possible domains for classification include landmarks and retail.

Adding and tagging images

Since a Custom Vision model is trained in Azure and not locally, we need to upload some images. The data we’ll use comes from the Microsoft Computer Vision Best Practices project. This is a simple set of images containing 4 kinds of objects one might find in a fridge: cans, cartons, milk bottles, and water bottles.


The generic function to add images to a project is add_images, which takes a vector of filenames, Internet URLs or raw vectors as the images to upload. The method for classification projects also has an argument tags which can be used to assign labels to the images as they are uploaded.

add_images returns a vector of image IDs, which are how Custom Vision keeps track of the images it uses. It should be noted that Custom Vision does not keep a record of the source filename or URL; it works only with image IDs. A future release of AzureVision may automatically track the source metadata, allowing you to associate an ID with an actual image. For now, this must be done manually.

Let’s upload the fridge objects to the project. We’ll keep aside 5 images from each class of object to use as validation data.

cans <- dir("fridgeObjects/can", full.names=TRUE)
cartons <- dir("fridgeObjects/carton", full.names=TRUE)
milk <- dir("fridgeObjects/milk_bottle", full.names=TRUE)
water <- dir("fridgeObjects/water_bottle", full.names=TRUE)

# upload all but 5 images from cans and cartons, and tag them
can_ids <- add_images(testproj, cans[-(1:5)], tags="can")
carton_ids <- add_images(testproj, cartons[-(1:5)], tags="carton")

If you don’t tag the images at upload time, you can do so later with add_image_tags:

# upload all but 5 images from milk and water bottles
milk_ids <- add_images(testproj, milk[-(1:5)])
water_ids <- add_images(testproj, water[-(1:5)])

add_image_tags(testproj, milk_ids, tags="milk_bottle")
add_image_tags(testproj, water_ids, tags="water_bottle")

Other image functions to be aware of include list_images, remove_images, and add_image_regions (which is for object detection projects). A useful one is browse_images, which takes a vector of IDs and displays the corresponding images in your browser.

browse_images(testproj, water_ids[1:5])

Training the model

Having uploaded the data, we can train the Custom Vision model with train_model. This trains the model on the server and returns a model iteration, which is the result of running the training algorithm on the current set of images. Each time you call train_model, for example to update the model after adding or removing images, you will obtain a different model iteration. In general, you can rely on AzureVision to keep track of the iterations for you, and automatically return the relevant results for the latest iteration.

mod <- train_model(testproj)
Azure Custom Vision model
  Project/iteration: testproj/Iteration 1 (f243bb4c-e4f8-473e-9df0-190a407472be)

Optional arguments to train_model include:

Other model iteration management functions are get_model (to retrieve a previously trained iteration), list_models (retrieve all previously trained iterations), and delete_model.

We can examine the model performance on the training data (which may be different to the current data!) with the summary method. For this toy problem, the model manages to obtain a perfect fit.

                                    id         name precision precisionStdDeviation recall
1 22ddd4bc-2031-43a1-b0ef-eb6b219eb6f7          can         1                     0      1
2 301db6f9-b701-4dc6-8650-a9cf3fe4bb2e       carton         1                     0      1
3 594ad770-83e5-4c77-825d-9249dae4a2c6  milk_bottle         1                     0      1
4 eda5869a-cc75-41df-9c4c-717c10f79739 water_bottle         1                     0      1

  recallStdDeviation averagePrecision    
1                  0                1
2                  0                1
3                  0                1
4                  0                1

[1] 1

[1] 0

[1] 1

[1] 0

[1] 1

Obtaining predictions from the trained model is done with the predict method. By default, this returns the predicted tag (class label) for the image, but you can also get the predicted class probabilities by specifying type="prob".

validation_imgs <- c(cans[1:5], cartons[1:5], milk[1:5], water[1:5])
validation_tags <- rep(c("can", "carton", "milk_bottle", "water_bottle"), each=5)

predicted_tags <- predict(mod, validation_imgs)

table(predicted_tags, validation_tags)
predicted_tags can carton milk_bottle water_bottle
  can            4      0           0            0
  carton         0      5           0            0
  milk_bottle    1      0           5            0
  water_bottle   0      0           0            5
head(predict(mod, validation_imgs, type="prob"))
              can       carton  milk_bottle water_bottle
[1,] 9.999968e-01 8.977501e-08 5.855104e-11 3.154334e-06
[2,] 9.732912e-01 3.454168e-10 4.610847e-06 2.670425e-02
[3,] 3.019476e-01 5.779990e-04 6.974699e-01 4.506565e-06
[4,] 5.072662e-01 2.849253e-03 4.856858e-01 4.198686e-03
[5,] 9.962270e-01 5.411842e-07 3.540882e-03 2.316211e-04
[6,] 3.145034e-11 1.000000e+00 2.574793e-10 4.242047e-14

This shows that the model got 19 out of 20 predictions correct on the validation data, misclassifying one of the cans as a milk bottle.


Publishing to a prediction resource

The code above demonstrates using the training endpoint to obtain predictions, which is really meant only for model testing and validation. For production purposes, we would normally publish a trained model to a Custom Vision prediction resource. Among other things, a user with access to the training endpoint has complete freedom to modify the model and the data, whereas access to the prediction endpoint only allows getting predictions.

Publishing a model requires knowing the Azure resource ID of the prediction resource. Here, we’ll use the resource object that was created earlier using AzureRMR; you can also obtain this information from the Azure Portal.

# publish to the prediction resource we created above
publish_model(mod, "iteration1", pred_res)

Once a model has been published, we can obtain predictions from the prediction endpoint in a manner very similar to previously. We create a predictive service object with classification_service, and then call the predict method. Note that a required input is the project ID; you can supply this directly or via the project object.

pred_url <- pred_res$properties$endpoint
pred_key <- pred_res$list_keys()[1]

pred_endp <- customvision_prediction_endpoint(url=pred_url, key=pred_key)

project_id <- testproj$project$id
pred_svc <- classification_service(pred_endp, project_id, "iteration1")

# predictions from prediction endpoint -- same as before
predsvc_tags <- predict(pred_svc, validation_imgs)
table(predsvc_tags, validation_tags)
predsvc_tags   can carton milk_bottle water_bottle
  can            4      0           0            0
  carton         0      5           0            0
  milk_bottle    1      0           5            0
  water_bottle   0      0           0            5

Exporting as standalone

As an alternative to deploying the model to an online predictive service resource, you can also export the model to a standalone format. This is only possible if the project was created to support exporting. The formats supported include:

To export the model, call export_model and specify the target format. By default, the model will be downloaded to your local machine, but export_model also (invisibly) returns a URL from where it can be downloaded independently.

export_model(mod, "tensorflow")
Downloading to
trying URL 'https://irisprodae...'
Content type 'application/octet-stream' length 4673656 bytes (4.5 MB)
downloaded 4.5 MB

See also