Getting Started with the AzureML Package
2016-01-29
Use this package to upload and download datasets to and from AzureML, to interrogate experiments, to publish R functions as AzureML web services, and to run R data through existing web services and retrieve the output.
Installation instructions
Install the development version of the package directly from GitHub with:
# Install devtools
if(!require("devtools")) install.packages("devtools")
devtools::install_github("RevolutionAnalytics/azureml")The package depends on:
jsonlitecurlminiCRANbase64encuuid
Some of the examples use data and functions in:
lme4ggplot2
Overview
AzureML provides an interface to publish web services on Microsoft Azure Machine Learning (Azure ML) from your local R environment. The main functions in the package cover the following topics:
- Workspace: connect to and manage AzureML workspaces
- Datasets: upload and download datasets to and from AzureML workspaces
- Publish: publish R functions as AzureML web services, and update or delete existing services
- Consume: apply any AzureML web service to your R data
Getting Started
To get started, please navigate to AzureML Studio and create a free account (not guest) or use your existing AzureML account. After logging in, under the “Settings” tab, copy and paste your Workspace ID from the “Name” sub-tab into your R console. From the “Authorization Tokens” sub-tab, copy your Primary Authorization Token into your R console. You will need this information to access all package functionality.
The package defines a Workspace class that represents an AzureML work space. Most of the functions in the package refer to a Workspace object directly or indirectly. Use the workspace() function to create Workspace objects, either by explicitly specifying an AzureML workspace ID and authorization token. Workspace objects are simply R environments that actively cache details about your AzureML sessions.
Obtaining AzureML Credentials
Before using the package, it is necessary to first obtain the security credentials to your Azure Machine Learning workspace. You can find this be logging in at https://studio.azureml.net. If you do not have an account, you can create a free account (not guest) to use these APIs.
Once logged in, you will be brought to the Studio landing page. Using the left-hand menu, navigate to the ‘Settings’ tab to find your Workspace ID. Note this, or copy it into your R session and store it is a variable, e.g. myWsID.
Next, within the ‘Settings’ tab, use the overhead menu to navigate to the ‘Authorization Tokens’ tab and similarly note your Primary Authorization Token.
library(AzureML)
ws <- workspace(
id = "your workspace ID",
auth = "your authorization token"
)or alternatively create a file in ~/.azureml/settings.json with the JSON structure (api_endpoint and management_endpoint are optional):
{"workspace": {
"id" : "test_id",
"authorization_token" : "test_token",
"api_endpoint" : "api_endpoint",
"management_endpoint" : "management_endpoint"
}}See ?workspace for more details.
Examining workspace datasets, experiments and services
The datasets(), experiments(), and services() functions return data frames that contain information about those objects available in the workspace.
The package caches R data frame objects describing available datasets, experiments and services in the workspace environment. That cache can be refreshed at any time with the refresh() function. The data frame objects make it relatively easy to sort and filter the datasets, experiments, and services in arbitrary ways. The functions also include filtering options for specific and common filters, like looking up something by name.
Use the download.datasets() and upload.dataset() functions to download or upload data between R and your Azure workspace. The download.intermediate.dataset() function can download ephemeral data from a port in an experiment that is not explicitly stored in your Azure workspace.
Use delete.datasets() to remove and delete datasets from the workspace.
The endpoints() function describes Azure web service endpoints, and works with supporting help functions like endpointHelp().
The publishWebService() function publishes a custom R function as an AzureML web service, available for use by any client. The updateWebService() and deleteWebServce() update or delete existing web services, respectively.
Use the consume() function to evaluate an Azure ML web service with new data uploaded to AzureML from your R environment.
Examples
Work with the AzureML package begins by defining a workspace object. The example below uses the configured workspace ID and authorization token in the ~/.azureml/settings.json file. Alternatively specify these settings explicitly in the workspace() function as outlined above. All of the examples require this step.
library(AzureML)
ws <- workspace()
ws## AzureML Workspace
## Workspace ID : dd01c7e4a424432c9a9f83142d5cfec4
## API endpoint : https://studio.azureml.netService availability
AzureML is a web service and sometimes operations can’t immediately proceed due to rate limiting or other factors. When this kind of thing occurs, the AzureML R package presents a warning and retries the service a few times before giving up with an error.
Datasets
AzureML datasets correspond more or less to R data frames. The AzureML package defines four basic dataset operations: list, upload, download, and delete.
List available datasets
The following example illustrates listing available datasets in your workspace.
head(datasets(ws)) # Or, equivalently: head(ws$datasets)## Name DataTypeId Size
## 1 Flight_On_Time_Performance_July_October_2013.csv GenericCSV 100506313
## 2 TestReadFromBlob Dataset 8102
## 3 New York weather GenericCSV 116989
## 4 airquality GenericTSV 2901
## 5 dataset-test-upload-2015-11-17--22-12-47 GenericTSV 2901
## 6 dataset-test-upload-2015-11-17--22-14-37 GenericTSV 2901
## ...
## 1 ...
## 2 ...
## 3 ...
## 4 ...
## 5 ...
## 6 ...
## ----------------------------------------------
## AzureML datasets data.frame variables include:
## [1] "VisualizeEndPoint" "SchemaEndPoint" "SchemaStatus"
## [4] "Id" "DataTypeId" "Name"
## [7] "Description" "FamilyId" "ResourceUploadId"
## [10] "SourceOrigin" "Size" "CreatedDate"
## [13] "Owner" "ExperimentId" "ClientVersion"
## [16] "PromotedFrom" "UploadedFromFilename" "ServiceVersion"
## [19] "IsLatest" "Category" "DownloadLocation"
## [22] "IsDeprecated" "Culture" "Batch"
## [25] "CreatedDateTicks"The list of datasets is presented as an a R data frame with class Datasets. Its print method shows a summary of the datasets, along with all of the available variables. Use any normal R data frame operation to manipulate the datasets. For example, to see the “Owner” value of each dataset:
head(ws$datasets$Owner, n=20)## [1] "apdevries" "apdevries"
## [3] "apdevries" "R"
## [5] "R" "R"
## [7] "R" "R"
## [9] "R" "R"
## [11] "R" "R"
## [13] "R" "Microsoft Corporation"
## [15] "Microsoft Corporation" "Microsoft Corporation"
## [17] "Microsoft Corporation" "Microsoft Corporation"
## [19] "Microsoft Corporation" "Microsoft Corporation"Downloading datasets
The next example illustrates downloading a specific dataset named “Airport Codes Dataset” from AzureML to your R session. This dataset is presented by AzureML as a “Generic CSV” dataset, and will be parsed by R’s read.table() function. (Other formats are parsed by an appropriate parser, for example read.arff().) The example illustrates passing additional arguments to the read.table() function used to parse the data from AzureML in this case.
airports <- download.datasets(ws, name = "Airport Codes Dataset", quote="\"")
head(airports)## airport_id city state name
## 1 10165 Adak Island AK Adak
## 2 10299 Anchorage AK Ted Stevens Anchorage International
## 3 10304 Aniak AK Aniak Airport
## 4 10754 Barrow AK Wiley Post/Will Rogers Memorial
## 5 10551 Bethel AK Bethel Airport
## 6 10926 Cordova AK Merle K Mudhole SmithYou can use download.datasets() to download more than one dataset as a time, returning the results in a list of data frames.
Uploading R data frames as AzureML datasets and deleting datasets
Use the upload.dataset() function to upload R data frames to AzureML.
upload.dataset(airquality, ws, name = "Air quality")## Name DataTypeId Size ...
## 1 Air quality GenericTSV 2901 ...
## ----------------------------------------------
## AzureML datasets data.frame variables include:
## [1] "VisualizeEndPoint" "SchemaEndPoint" "SchemaStatus"
## [4] "Id" "DataTypeId" "Name"
## [7] "Description" "FamilyId" "ResourceUploadId"
## [10] "SourceOrigin" "Size" "CreatedDate"
## [13] "Owner" "ExperimentId" "ClientVersion"
## [16] "PromotedFrom" "UploadedFromFilename" "ServiceVersion"
## [19] "IsLatest" "Category" "DownloadLocation"
## [22] "IsDeprecated" "Culture" "Batch"
## [25] "CreatedDateTicks"# Let's see what we've got:
head(download.datasets(ws, name = "Air quality"))## Ozone Solar.R Wind Temp Month Day
## 1 41 190 7.4 67 5 1
## 2 36 118 8.0 72 5 2
## 3 12 149 12.6 74 5 3
## 4 18 313 11.5 62 5 4
## 5 NA NA 14.3 56 5 5
## 6 28 NA 14.9 66 5 6Delete one or more AzureML datasets with delete.datasets():
delete.datasets(ws, name="Air quality")## Request failed with status 400. Waiting 2.3 seconds before retry## ...## Name Deleted status_code
## 1 Air quality TRUE 204Experiments
Use the experiments() function or simply use the ws$experiments data frame object directly to list details about experiments in your AzureML workspace. The experiments() function optionally filters experiments by ownership.
e <- experiments(ws, filter = "samples")
head(e)## Description CreationTime ...
## 1 Sample 6: Train, Test, Evaluate for Regression: 2015-08-27 21:34:57 ...
## 2 Text Classification: Step 2 of 5, text preproces 2015-08-27 21:39:38 ...
## 3 Quantile Regression: Car price prediction 2015-08-27 21:37:39 ...
## 4 Multiclass Classification: News categorization 2015-08-27 21:36:23 ...
## 5 Neural Network: Basic convolution 2015-08-27 21:36:48 ...
## 6 Text Classification: Step 3B of 5, unigrams TF-I 2015-08-27 21:39:49 ...
## -------------------------------------------------
## AzureML experiments data.frame variables include:
## [1] "ExperimentId"
## [2] "Description"
## [3] "Etag"
## [4] "Creator"
## [5] "IsArchived"
## [6] "JobId"
## [7] "VersionId"
## [8] "RunId"
## [9] "OriginalExperimentDocumentationLink"
## [10] "Summary"
## [11] "Category"
## [12] "Tags"
## [13] "StatusCode"
## [14] "StatusDetail"
## [15] "CreationTime"
## [16] "StartTime"
## [17] "EndTime"
## [18] "Metadata"The ws$experiments object is just an R data frame with class Experiments. Its print method shows a summary of the available experiments, but it can otherwise be manipulated like a normal R data frame.
The list of experiments in your workspace is cached in the workspace environment. Use the refresh() function to explicitly update the cache at any time, for example:
refresh(ws, "experiments")Web Services
The AzureML package helps you to publish R functions as AzureML web services that can be consumed anywhere. You can also use the AzureML package to run R data through an existing web service and collect the output.
Publishing a Web Service
The publishWebService() publishes an R function as an AzureML web service. Consider this simple example R function:
add <- function(x, y) {
x + y
}Use the function publishWebService() to publish the function as a service named “AzureML-vignette-silly”:
ws <- workspace()
api <- publishWebService(
ws,
fun = add,
name = "AzureML-vignette-silly",
inputSchema = list(
x = "numeric",
y = "numeric"
),
outputSchema = list(
ans = "numeric"
)
)The example publishes a function of two scalar numeric arguments, returning a single numeric scalar output value. Note that we explicitly define the web service input and output schema in the example. See the examples below for more flexible ways of defining web services with functions of data frames.
The result of publishWebService() is an Endpoint object, really just an R data frame with two elements: a list containing the details of the newly created web service, and a list of the endpoints of the web service. From here, you can pass the information on to another user, or use the information to use the web service from R:
class(api)## [1] "Endpoint" "data.frame"names(api)## [1] "Name" "Description"
## [3] "CreationTime" "WorkspaceId"
## [5] "WebServiceId" "HelpLocation"
## [7] "PrimaryKey" "SecondaryKey"
## [9] "ApiLocation" "PreventUpdate"
## [11] "MaxConcurrentCalls" "DiagnosticsTraceLevel"
## [13] "ThrottleLevel"The web service created is identical to a web service published through the Azure Machine Learning Studio. From the response, you can get the Web Service’s URL, API Key and Help Page URL, as shown above. The first two are needed to make calls to the web service. The latter has the sample code, sample request and other information for consuming the API from client apps such as mobile and web applications.
The new web service will show up on the ‘Web Services’ tab of the Studio interface, and the service will have a help page for each endpoint, e.g.
Note that AzureML allows multiple services to have the same name.
(helpPageUrl <- api$HelpLocation)## [1] "https://studio.azureml.net/apihelp/workspaces/dd01c7e4a424432c9a9f83142d5cfec4/webservices/10b27c1cc68d11e59a9e9fda44a82b45/endpoints/4ce411c50ec3486db0a5c191af0309c1"Once published, you can update a web service using the updateWebService() or publishWebService() functions. The updateWebService() function is just an alias for publishWebService(), except that the argument serviceId is compulsory.
api <- updateWebService(
ws,
fun = function(x, y) x - y,
inputSchema = list(
x = "numeric",
y = "numeric"
),
outputSchema = list(
ans = "numeric"
),
serviceId = api$WebServiceId # <<-- Required to update!
)The “AzureML-vignette-silly” service now substracts two numbers instead of adding them.
Discovering Web Services
Use the services() function to list in detail all of the available services in your AzureML workspace, or filter by web service name as shown below:
(webservices <- services(ws, name = "AzureML-vignette-silly"))## Id Name Description
## 22 0118e986c68c11e5bf7d7d3f8709ee1b AzureML-vignette-silly <NA>
## 24 7cca5998c68c11e581cfcf4167576cd7 AzureML-vignette-silly <NA>
## 26 10b27c1cc68d11e59a9e9fda44a82b45 AzureML-vignette-silly <NA>
## CreationTime WorkspaceId
## 22 2016-01-29T13:27:14.153Z dd01c7e4a424432c9a9f83142d5cfec4
## 24 2016-01-29T13:30:42.485Z dd01c7e4a424432c9a9f83142d5cfec4
## 26 2016-01-29T13:34:50.512Z dd01c7e4a424432c9a9f83142d5cfec4
## DefaultEndpointName EndpointCount
## 22 default 1
## 24 default 1
## 26 default 1Given a service, use the endpoints() function to list the AzureML service endpoints for the service:
ep <- endpoints(ws, webservices[1, ])
class(ep)## [1] "Endpoint" "data.frame"names(ep)## [1] "Name" "Description"
## [3] "CreationTime" "WorkspaceId"
## [5] "WebServiceId" "HelpLocation"
## [7] "PrimaryKey" "SecondaryKey"
## [9] "ApiLocation" "PreventUpdate"
## [11] "MaxConcurrentCalls" "DiagnosticsTraceLevel"
## [13] "ThrottleLevel"The returned Endpoints object contains all the information needed to consume a web service. The endpointHelp() function returns detailed information about an endpoint including its input and output schema and URI.
Consuming Web Services
Use the consume() function to send data to your newly published web service API for scoring.
df <- data.frame(
x = 1:5,
y = 6:10
)
s <- services(ws, name = "AzureML-vignette-silly")
s <- tail(s, 1) # use the last published function, in case of duplicate function names
ep <- endpoints(ws, s)
consume(ep, df)## ans
## 1 -5
## 2 -5
## 3 -5
## 4 -5
## 5 -5Alternatively, the endpoint primary key and API location can be found on the help page for that specific endpoint, which can be found on Azure Machine Learning Studio. Using the Help Page URL, you can access sample code to build clients that can consume this web service in real time to make predictions.
Deleting a Web Service
Use deleteWebservice() to remove a webservice endpoint that you no longer need or want (like these silly examples):
deleteWebService(ws, name = "AzureML-vignette-silly")Other examples of publishing web services
The simplest and perhaps most useful way to define a web service uses functions that take a single data frame argument and return a vector or data frame of results. The next example trains a generalized boosted regression model using the gbm package, publishes the model as a web service with name “AzureML-vignette-gbm”, and runs example data through the model for prediction using the consume() function.
library(AzureML)
library(MASS)
library(gbm)## Loading required package: survival## Loading required package: lattice## Loading required package: splines## Loading required package: parallel## Loaded gbm 2.1.1ws <- workspace()
test <- Boston[1:5, 1:13]
set.seed(123)
gbm1 <- gbm(medv ~ .,
distribution = "gaussian",
n.trees = 5000,
interaction.depth = 8,
n.minobsinnode = 1,
shrinkage = 0.01,
cv.folds = 5,
data = Boston,
n.cores = 1) # You can set this to n.cores = NULL to use all cores
best.iter <- gbm.perf(gbm1, method="cv", plot=FALSE)
mypredict <- function(newdata)
{
require(gbm)
predict(gbm1, newdata, best.iter)
}
# Example use of the prediction function
print(mypredict(test))## [1] 24.54431 21.15155 33.88859 34.06615 34.93906# Publish the service
ep <- publishWebService(ws = ws, fun = mypredict, name = "AzureML-vignette-gbm",
inputSchema = test)
# Consume test data, comparing with result above
print(consume(ep, test))## Request failed with status 401. Waiting 6.7 seconds before retry## .......## ans
## 1 24.54431
## 2 21.15155
## 3 33.88859
## 4 34.06615
## 5 34.93906Notice that we don’t need to explicitly specific the inputSchema or outputSchema arguments when working with functions that use data frame I/O. When finished with this example, we can delete the example service with:
deleteWebService(ws, "AzureML-vignette-gbm")Tips on writing functions used in web services
Try to use the data frame I/O interface as illustrated in the last example above. It’s simpler and more robust than using functions of scalars or lists and exhibits faster execution for large data sets.
Use require() in your function to explicitly load required packages.
The publishWebServce() function uses codetools to bundle objects required by your function following R lexical scoping rules. The previous example, for instance, uses the best.iter and gbm1 variables inside of the mypredict() function. publishWebService() identified that and included their definitions in the R environment in which the function is evaluated in AzureML. Fine-grained control over the export of variables is provided by the publishWebService() function in case you need it (see the help page for details).
Use the packages option of publishWebService() to explicitly bundle required packages and their dependencies (but not suggested dependencies) using miniCRAN. This lets you upload packages to AzureML that may not otherwise be available in that environment already, using the correct R version and platform used by AzureML.
Be aware that the version of R running in AzureML may not be the same as the version of R that you are running locally. That means that some packages might not be available, or sometimes package behavior in the AzureML version of R might be different that what you observe locally. This is generally more of an issue for cutting-edge packages.
JSON is used to transfer data between your local R environment and the R services running in AzureML–numeric values experience a change of base, which can lead to a small loss of precision in some circumstances. If you really, really need to move binary objects between your local R session and the AzureML R service you might try base64 encoding the data, for example.