Revolutions

AI, Machine Learning and Data Science Roundup: May 2019

A monthly roundup of news about Artificial Intelligence, Machine Learning and Data Science. This is an eclectic collection of interesting blog posts, software announcements and data applications from Microsoft and elsewhere that I've noted over the past month or so.

Open Source AI, ML & Data Science News

PyTorch 1.1 is now available, with new support for Tensorboard and improvements to distributed training and JIT compilation.

JupyterHub 1.0 is released, a milestone for the multi-user Jupyter Notebook server.

matplotlib 3.1 is released, adding several improvements to the Python data visualization library.

Python is now included in Windows 10, with updates available via the Microsoft Store.

FastBert, a simple PyTorch interface for training text classifiers based on the popular language representation model BERT, is released.

torchvision 0.3, the PyTorch library of datasets and tools for computer vision, adds new models for semantic segmentation and object detection.

ML.NET 1.0, the open-source cross-platform machine learning framework for .NET developers, is now available.

ONNX Runtime 0.4 is released, adding support for Intel and NVIDIA accelerators to further reduce latency for deployed neural networks.

R 3.6.0 is released, with many new capabilities and improved memory usage and performance.

Industry News

The Wolfram Engine, featured in Mathematica and including a suite of algorithms for visualization, machine learning, NLP and more, is now available free for development (license required for production).  

Facebook releases Pythia, a new open-source deep learning framework based on PyTorch, for multitasking in the vision and language domain.

Google Cloud AutoML Natural Language provides an interactive UI to classify content and build a predictive ML model without coding.

Google Cloud TPU Pods, cloud-based clusters that can include more than 1,000 TPU chips as an "ML supercomputer", are now publicly available in beta. NVIDIA T4 GPUs are also now generally available in GCP.

Snips open-sources Tract, an embedded neural network inference engine designed for wake-word detection by virtual assistant devices.

Hewlett Packard Enterprises announced plans to acquire Cray, the supercomputer company.

Intel announces several open-source initiatives for AI and cloud technologies, including a Deep Learning Reference Stack optimized for Intel chipsets.

Microsoft News

Azure Machine Learning Service adds new MLOps capabilities, providing version control, audit trails, and packaging, deployment and monitoring support for machine learning models via an Azure DevOps extension. Also, model deployment to FPGA is now generally available in Azure (and additionally in preview for Databox Edge).

Further updates to Azure Machine Learning Service are now in preview, including a new drag-and-drop visual interface, a new form-based UI for automated machine learning, model interpretability, and hosted Python notebooks.

Azure Cognitive Services has added many new capabilities, highlights of which include:

• A new category of service: Decision. The services include Content Moderator, Anomaly Detector and a new service called Personalizer (in preview).
• Conversation Transcription, a new Speech service that can discriminate and transcribe audio from multiple speakers in real time.
• Ink Recognizer, a new Vision service that allows developers to add digital handwriting recognition to apps.
• Additional Cognitive Services offering container support: Anomaly Detector, Speech-to-Text and Text-to-Speech.
• Azure Search, which includes many cognitive skills including image recognition and language comprehension, is now generally available.
• Form Recognizer, a new knowledge discovery service that extracts key-value pairs and tables from forms in scanned and PDF documents, now in preview.

Visual Studio Code adds Remote Development, allowing use of remote Python workspaces over SSH, in Docker containers, and in Windows Subsystem for Linux. Other recent Python support improvements include Intellisense in the console and enhancements to the Python Language Server.

Azure Data Explorer now supports queries with custom Python code, as well as integration with Spark.

Azure SQL Database Edge, a small-footprint data engine with support for Python, R and Spark and optimized for edge devices and time-series data, is now in private preview.

Microsoft announces an end-to-end toolchain for autonomous systems (in preview), which developers can use to simulate and build robots and other AI-driven autonomous devices.

Microsoft R Open 3.5.3 is released; new R packages AzureKusto (interface to Data Explorer) and AzureGraph (interface to the Microsoft Graph), are available.

Learning resources

A beginner's tutorial on training a convolutional neural network, using only Python and numpy.

Rules of Machine Learning: Google's list of best practices for developers looking to create applications with machine learning capabilities. 

ODSC suggests 25 public data sets to get started with machine learning, spanning text, images, and tabular data.

Foundations of Data Science, a free book by Avrim Blum, John Hopcroft and Ravi Kannan with a focus on matrix decompositions and associated ML techniques.

Azure Open Datasets, a collection of curated public datasets easily accessible to Azure ML services.

Open Images v5, a large collection of annotated images including segmentation masks for 2.8 million objects in 350 categories, has been released by Google.

Microsoft Learn now offers several modules with free training for AI engineers and data scientists.

Six principles behind health data-driven organizations, a useful resource on building a data science culture by Francesca Lazzeri.

Applications

How Python is used at Netflix for personalization, machine learning, experimentation, statistical analysis and more.

Google develops a method to infer depth maps for video by using "mannequin challenge" videos as training data.

Samsung researchers develop a method of animating a single photograph of a person as a realistic talking head (video).

Google's "Translatotron", a speech-to-speech translation model that translates speech audio into a second language while retaining the original speaking voice.

AI-based applications help children with disabilities bridge language gaps.

Facebook researchers develop a method to synthesize full-body video of a person performing actions animated in real time under joystick control, based only on real source video.

LaLiga and BMW are using the Azure Bot Framework SDK to deliver specialized personal assistant applications.

Find previous editions of the monthly AI roundup here.

MRAN snapshots, and you

For almost five years, the entire CRAN repository of R packages has been archived on a daily basis at MRAN. If you use CRAN snapshots from MRAN, we'd love to hear how you use them in this survey. If you're not familiar with the concept, or just want to learn more, read on.

Every day since September 17, 2014, we (Microsoft and, before the acquisition, Revolution Analytics) have archived a snapshot of the entire CRAN repository as a service to the R community. These daily snapshots have several uses:

• As a longer-term archive of binary R packages. (CRAN keeps an archive of package source versions, but binary versions of packages are kept for a limited time. CRAN keeps package binaries only for the current R version and the prior major version, and only for the latest version of the package). 
• As a static CRAN repository you can use like the standard CRAN repository, but frozen in time. This means changes to CRAN packages won't affect the behavior of R scripts in the future (for better or worse). options(repos="https://cran.microsoft.com/snapshot/2017-03-15/") provides a CRAN repository that works with R 3.3.3, for example — and you can choose any date since September 17, 2014.
• The checkpoint package on CRAN provides a simple interface to these CRAN snapshots, allowing you use a specific CRAN snapshot by specifying a date, and making it easy to manage multiple R project each using different snapshots.
• Microsoft R Open, Microsoft R Client, Microsoft ML Server and SQL Server ML Services all use fixed CRAN repository snapshots from MRAN by default.
• The rocker project provides container instances for historical versions of R, tied to an appropriate CRAN snapshot from MRAN suitable for the corresponding R version.

Browse the MRAN time machine to find specific CRAN snapshots by date. (Tip: click the R logo to open the snapshot URL in its own new window.)

MRAN and the CRAN snapshot system was created at a time when reproducibility was an emerging concept in the R ecosystem. Now, there are several methods available to ensure that your R code works consistently, even as R and CRAN changes. Beyond virtualization and containers, you have packages like packrat and miniCRAN, RStudio's package manager, and the full suite of tools for reproducible research.

As CRAN has grown and changes to packages have become more frequent, maintaining MRAN is an increasingly resource-intensive process. We're contemplating changes, like changing the frequency of snapshots, or thinning the archive of snapshots that haven't been used. But before we do that we'd  like to hear from the community first. Have you used MRAN snapshots? If so, how are you using them? How many different snapshots have you used, and how often do you change that up? Please leave your feedback at the survey link below by June 14, and we'll use the feedback we gather in our decision-making process. Responses are anonymous, and we'll summarize the responses in a future blog post. Thanks in advance!

Take the MRAN survey here.

Introduction to AzureKusto

By Hong Ooi and Alex Kyllo

This post is to announce the availability of AzureKusto, the R interface to Azure Data Explorer (internally codenamed “Kusto”), a fast, fully managed data analytics service from Microsoft. It is available from CRAN, or you can install the development version from GitHub via devtools::install_github("cloudyr/AzureKusto").

AzureKusto provides an interface (including DBI compliant methods for connecting to Kusto clusters and submitting Kusto Query Language (KQL) statements, as well as a dbplyr style backend that translates dplyr queries into KQL statements. On the administrator side, it extends the AzureRMR framework to allow for creating clusters and managing database principals.

Connecting to a cluster

To connect to a Data Explorer cluster, call the kusto_database_endpoint() function. Once you are connected, call run_query() to execute queries and command statements.

library(AzureKusto)

## Connect to a Data Explorer cluster with (default) device code authentication
Samples <- kusto_database_endpoint(
    server="https://help.kusto.windows.net",
    database="Samples")

res <- run_query(Samples,
    "StormEvents | summarize EventCount = count() by State | order by State asc")
head(res)

##            State EventCount
## 1        ALABAMA       1315
## 2         ALASKA        257
## 3 AMERICAN SAMOA         16
## 4        ARIZONA        340
## 5       ARKANSAS       1028
## 6 ATLANTIC NORTH        188

# run_query can also handle command statements, which begin with a '.' character
res <- run_query(Samples, ".show tables | count")
res[[1]]

##   Count
## 1     5

dplyr Interface

The package also implements a dplyr-style interface for building a query upon a tbl_kusto object and then running it on the remote Kusto database and returning the result as a regular tibble object with collect(). All the standard verbs are supported.

library(dplyr)
StormEvents <- tbl_kusto(Samples, "StormEvents")
q <- StormEvents %>%
    group_by(State) %>%
    summarize(EventCount=n()) %>%
    arrange(State)

show_query(q)
## <KQL> database('Samples').['StormEvents']
## | summarize ['EventCount'] = count() by ['State']
## | order by ['State'] asc

collect(q)
## # A tibble: 67 x 2
##    State          EventCount
##    <chr>               <dbl>
##  1 ALABAMA              1315
##  2 ALASKA                257
##  3 AMERICAN SAMOA         16
## ...

DBI interface

AzureKusto implements a subset of the DBI specification for interfacing with databases in R.

The following methods are supported:

• Connections: dbConnect, dbDisconnect, dbCanConnect
• Table management: dbExistsTable, dbCreateTable, dbRemoveTable, dbReadTable, dbWriteTable
• Querying: dbGetQuery, dbSendQuery, dbFetch, dbSendStatement, dbExecute, dbListFields, dbColumnInfo

It should be noted, though, that Data Explorer is quite different to the SQL databases that DBI targets. This affects the behaviour of certain DBI methods and renders other moot.

library(DBI)

Samples <- dbConnect(AzureKusto(),
                     server="https://help.kusto.windows.net",
                     database="Samples")

dbListTables(Samples)
## [1] "StormEvents"       "demo_make_series1" "demo_series2"     
## [4] "demo_series3"      "demo_many_series1"

dbExistsTable(Samples, "StormEvents")
##[1] TRUE

dbGetQuery(Samples, "StormEvents | summarize ct = count()")
##      ct
## 1 59066

If you have any questions, comments or other feedback, please feel free to open an issue on the GitHub repo.

And one more thing...

As of Build 2019, Data Explorer can also run R (and Python) scripts in-database. For more information on this feature, currently in public preview, see the Azure blog and the documentation article.

 

 

What's new in R 3.6.0

A major update to the open-source R language, R 3.6.0, was released on April 26 and is now available for download for Windows, Mac and Linux. As a major update, it has many new features, user-visible changes and bug fixes. You can read the details in the release announcement, and in this blog post I'll highlight the most significant ones.

Changes to random number generation. R 3.6.0 changes the method used to generate random integers in the sample function. In prior versions, the probability of generating each integer could vary from equal by up to 0.04% (or possibly more if generating more than a million different integers). This change will mainly be relevant to people who do large-scale simulations, and it also means that scripts using the sample function will generate different results in R 3.6.0 than they did in prior versions of R. If you need to keep the results the same (for reproducibility or for automated testing), you can revert to the old behavior by adding RNGkind(sample.kind="Rounding")) to the top of your script.

Changes to R's serialization format. If you want to save R data to disk with R 3.6.0 and read it back with R version 3.4.4 or earlier, you'll need to use readRDS("mydata.Rd",version=2) to save your data in the old serialization format, which has been updated to Version 3 for this release. (The same applies to the functions save, serialize, and byte-compiled R code.)  The R 3.5 series had forwards-compatibility in mind, and can already read data serialized in the Version 3 format.

Improvements to base graphics. You now have more options the appearance of axis labels (and perpendicular labels no longer overlap), better control over text positioning, a formula specification for barplots, and color palettes with better visual perception.

Improvements to package installation and loading, which should eliminate problems with partially-installed packages and reduce the space required for installed packages.

More functions now support vectors with more than 2 billion elements, including which and pmin/pmax.

Various speed improvements to functions including outer, substring, stopifnot, and the $ operator for data frames.

Improvements to statistical functions, including standard errors for T tests and better influence measures for multivariate models.

R now uses less memory, thanks to improvements to functions like drop, unclass and seq to use the ALTREP system to avoid duplicating data.

More control over R's memory usage, including the ability limit the amount of memory R will use for data. (Attempting to exceed this limit will generate a "cannot allocate memory" error.) This is particularly useful when R is being used in production, to limit the impact of a wayward R process on other applications in a shared system.

Better consistency between platforms. This has been an ongoing process, but R now has fewer instances of functions (or function arguments) that are only available on limited platforms (e.g. on Windows but not on Linux). The documentation is now more consistent between platforms, too. This should mean fewer instances of finding R code that doesn't run on your machine because it was written on a different platform.

There are many other specialized changes as well, which you can find in the release notes. Other than the issues raised above, most code from prior versions of R should run fine in R 3.6.0, but you will need to re-install any R packages you use, as they won't carry over from your R 3.5.x installation. Now that a couple of weeks have passed since the release, most packages should be readily available on CRAN for R 3.6.0.

As R enters its 20th year of continuous stable releases, please do take a moment to reflect on the ongoing commitment of the R Core Team for their diligence in improving the R engine multiple times a year. Thank you to everyone who has contributed. If you'd like to support the R project, here's some information on contributing to the R Foundation.

Final note: the code name for R 3.6.0 is "Planting of a Tree". The R code-names generally refer to Peanuts comics: if anyone can identify what the R 3.6.0 codename is referring to, please let us know in the comments! Update: Thanks for the comments! The consensus seems to be that the name refers to Arbor Day, which coincided with R 3.6.0's release day of April 26.

Ah, the "It's Arbor Day, Charlie Brown!" origin strip…https://t.co/aOdAqZBKaE pic.twitter.com/VrSCFaU74h

— Mara Averick (@dataandme) May 10, 2019

R-announce mailing list: R 3.6.0 is released

Availability of Microsoft R Open 3.5.2 and 3.5.3

It's taken a little bit longer than usual, but Microsoft R Open 3.5.2 (MRO) is now available for download for Windows and Linux. This update is based on R 3.5.2, and accordingly fixes a few minor bugs compared to MRO 3.5.1. The main change you will note is that new CRAN packages released since R 3.5.1 can now be used with this version of MRO.

Microsoft R Open 3.5.3, based on R 3.5.3, will be available next week, on May 10. Microsoft R Open 3.6.0, based on the recently-released R 3.6.0, is currently under development, and we'll make an announcement here when it's available too.

One thing to note: as of version 3.5.2 Microsoft R Open is no longer distributed for MacOS systems. If you want to continue to take advantage of the Accelerate framework on MacOS for, it's not too difficult to tweak the CRAN binary for MacOS to enable multi-threaded computing, and the additional open source bundled packages (like checkpoint and iterators) are available to install from CRAN or GitHub. 

As always, we hope you find Microsoft R Open useful, and if you have any comments or questions please visit the Microsoft R Open forum. You can follow the development of Microsoft R Open at the MRO Github repository. To download Microsoft R Open, simply follow the link below.

MRAN: Download Microsoft R Open