Yann-Aël Le Borgne, Jacopo De Stefani and Gianluca Bontempi - Machine Learning Group
The course slides (analytics part) are available in the Analytics_Course_Slides folder.
This repository contains the material for the exercise classes of the ULB/VUB Big Data Analytics master course (first semester 2018) - Advanced analytics part.
These hands-on sessions provide:
-
Session 1 : An introduction to Spark and its Machine Learning (ML) library. The case study for the first session is a churn prediction problem: How to predict which customers will quit a subscription to a given service? The session covers the basics for loading and formatting a dataset for training an ML algorithm using Spark ML library, and illustrates the use of different Spark ML algorithms and accuracy metrics to address the prediction problem.
-
Sessions 2-4: An in-depth coverage of the use of the Map/Reduce programming model for distributing machine learning algorithms, and their implementation in Spark. Sessions 2, 3, and 4 cover, respectively, the Map/Reduce implementations from scratch of
- Session 2: Linear regression (ordinary least squares and stochastic gradient descent). The algorithms are applied on an artificial dataset, and illustrate the numpy and Map/Reduce implementations for OLS and SGD.
- Session 3: Clustering with K-means. The algorithm is first applied on an artificial dataset, and then on a clustering problem for image compression.
- Session 4: Recommender system with alternating least squares, using as a case study a movie recommendation problem.
After detailing the Map/Reduce techniques for solving these problems, each session ends with an example on how to use the corresponding algorithm with Spark ML, and get insights into how Spark distributes the task using the Spark user interface.
-
Session 5: An overview of a deep learning framework (Keras/Tensorflow), and its use for image classification using convolutional neural networks.
The material is available as a set of Jupyter notebooks.
From the command line, use
git clone https://github.com/Yannael/BigDataAnalytics_INFOH515
If using the course cluster, you will have to use SFTP to send this folder to the cluster.
After setting up your environment - see below, you should be able to run the notebooks in Check_Setup
These notebooks rely on different technologies and frameworks for Big Data and machine learning (Spark, Kafka, Keras and Tensorflow). We summarize below different ways to have your environment set up.
Install Anaconda Python (see https://www.anaconda.com/download/, choose Linux distribution, and Python 3.6 version for 64-bit x86 systems).
Make sure the binaries are in your PATH. Anaconda installer proposes to add them at the end of the installation process. If you decline, you may later add
export ANACONDA_HOME=where_you_installed_anaconda
export PATH=$ANACONDA_HOME/bin:$PATH
Download from https://spark.apache.org/downloads.html (Use version 2.2.1, prebuilt for Apache Hadoop 2.7). Untar and add executables to your PATH, as well as Python libraries to PYTHONPATH
export SPARK_HOME=where_you_untarred_spark
export PATH=$SPARK_HOME/bin:$SPARK_HOME/sbin:$PATH
export PYTHONPATH="$SPARK_HOME/python/lib/pyspark.zip:$SPARK_HOME/python/lib/py4j-0.10.4-src.zip"
Download from https://kafka.apache.org/downloads, and untar archive. Start with
export KAFKA_HOME=where_you_untarred_kafka
nohup $KAFKA_HOME/bin/zookeeper-server-start.sh $KAFKA_HOME/config/zookeeper.properties > $HOME/zookeeper.log 2>&1 &
nohup $KAFKA_HOME/bin/kafka-server-start.sh $KAFKA_HOME/config/server.properties > $HOME/kafka.log 2>&1 &
Install with pip
pip install tensorflow
pip install keras
The notebook is part of Anaconda. Start Jupyter notebook with
jupyter notebook
and open in the browser at 127.0.0.1:8888
In order to ease the setting-up of the environment, we also prepared a Docker container that provides a ready-to-use environment. See docker folder for installing Docker, downloading the course container, and get started with it.
Note that the Dockerfile script essentially follows the steps for the 'local' installation.
Anaconda can be found in /serveur/logiciels/anaconda3. Simply use
export ANACONDA_HOME=/serveur/logiciels/anaconda3
export PATH=$ANACONDA_HOME/bin:$PATH
to use Anaconda Python.
Download from https://spark.apache.org/downloads.html (Use version 2.2.1, prebuilt for Apache Hadoop 2.7).
Untar and add executables to your PATH, as well as Python libraries to PYTHONPATH
export SPARK_HOME=where_you_untarred_spark
export PATH=$SPARK_HOME/bin:$SPARK_HOME/sbin:$PATH
export PYTHONPATH="SPARK_HOME/python/lib/pyspark.zip:SPARK_HOME/python/lib/py4j-0.10.4-src.zip"
Due to permission restrictions, you should manually add the execution flags to the executables that will be needed to start spark sessions, more specifically
chmod a+x $SPARK_HOME/bin/pyspark
chmod a+x $SPARK_HOME/bin/spark-submit
chmod a+x $SPARK_HOME/bin/spark-class
You may then start pyspark from the command line, and open the Spark UI at 127.0.0.1:4040.
The notebook is part of Anaconda. Start Jupyter notebook with
jupyter notebook
and open in the browser at 127.0.0.1:8888
In general (wired connection on campus, from home, or reasonably open wifi):
ssh -p 30 -L 8000:jupyter:8000 -L 8888:cdh02:8888 -L 8088:cdh02:8088 [email protected]
On Eduroam (port 30 is blocked)
ssh -p 3128 -L 8000:jupyter:8000 -L 8888:cdh02:8888 -L 8088:cdh02:8088 [email protected]
Note:
- The 'LaPlaine' and 'Solbosch' Wifi will not work due to strict port restrictions.
- -L is for port redirections. This is needed to have access to any 'Web' services from the cluster.
- Port redirections can be a bit tricky. To sum up:
- Jupyter is available on port 8000, at jupyter:8000
- Hue is available on port 8888, at hue:8888
- The Hadoop web UI is available on port 8088, at cdh02:8888
- More generally: There are a number of other Web services available from the cluster (to be detailed later), and there might be some dynamic changes in these address:ports (for example, the Hadoop Web UI may not always be available at cdh02:8088 due to load balancing - also to be detailed later).
Add Anaconda binaries to your PATH
export PATH=/etc/anaconda3/bin:$PATH
On the cluster, the following environment variables should be set:
# For Yarn, so that Spark knows where it runs
export HADOOP_CONF_DIR=/etc/hadoop/conf
# For Yarn, so Spark knows which version to use (and we want Anaconda to be used, so we have access to numpy, pandas, and so forth)
export PYSPARK_PYTHON=/etc/anaconda3/bin/python
export PYSPARK_DRIVER_PYTHON=/etc/anaconda3/bin/python
Simply connect on port 8000 locally (thanks to the SSH redirection), at 127.0.0.1:8000 in your browser.
For pushing data on the cluster, the best is to use SFTP, with scp command for example.
tar cvzf BigDataAnalytics_INFOH515.tgz BigDataAnalytics_INFOH515
scp -P 30 BigDataAnalytics_INFOH515.tgz [email protected]:/home/yourlogin
ssh -p 3128 -L 8000:jupyter:8000 -L 8888:hue:8888 -L 8088:cdh02:8088 [email protected]
tar xvzf BigDataAnalytics_INFOH515.tgz
The notebook otherwise allows you to upload files ('upload' button at the top right corner), which you can use as a quick workaround to get a notebook on the cluster.
- Upload notebook from
Check_Setup/Demo_RDD_cluster.ipynb - Run all cells
Follow instructions in Check_Setup/Demo_RDD_cluster.ipynb to have access to Spark UI.
If you get the following message
@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
@ WARNING: REMOTE HOST IDENTIFICATION HAS CHANGED! @
@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
Open .ssh/known_hosts, delete all entries starting with bigdata.ulb.ac.be, and retry connection.