microsoft
/
recommenders
зеркало из https://github.com/microsoft/recommenders.git
1
0
Форкнуть 0
Код Задачи Пакеты Проекты Релизы Вики Активность
recommenders/SETUP.md

20 KiB
Исходник Ответственный История

Setup guide

This document describes how to setup all the dependencies to run the notebooks in this repository in following platforms:

Table of Contents

Compute environments

Depending on the type of recommender system and the notebook that needs to be run, there are different computational requirements. Currently, this repository supports Python CPU, Python GPU and PySpark.

Setup guide for Local or DSVM

There are different ways one may use the recommenders utilities. The most convenient one is probably by installing the recommenders package from PyPI.

Another way is to build a docker image and use the functions inside a docker container.

Another alternative is to run all the recommender utilities directly from a local copy of the source code. This requires installing all the necessary dependencies from Anaconda and PyPI. For instructions on how to do this, see this guide.

Requirements

  • A machine running Linux, MacOS or Windows

  • An optional requirement is Anaconda with Python version >= 3.6

    • This is pre-installed on Azure DSVM such that one can run the following steps directly. To setup on your local machine, Miniconda is a quick way to get started.

    Alternatively a virtual environment can be used instead of Anaconda.

  • Apache Spark (this is only needed for the PySpark environment).

Dependencies setup

As a pre-requisite to installing the dependencies, if using Conda, make sure that Anaconda and the package manager Conda are both up to date:

conda update conda -n root
conda update anaconda        # use 'conda install anaconda' if the package is not installed

If using venv, see these instructions.

NOTE the xlearn package has dependency on cmake. If one uses the xlearn related notebooks or scripts, make sure cmake is installed in the system. The easiest way to install on Linux is with apt-get: sudo apt-get install -y build-essential cmake. Detailed instructions for installing cmake from source can be found here.

NOTE the models from Cornac require installation of libpython i.e. using sudo apt-get install -y libpython3.6 or libpython3.7, depending on the version of Python.

NOTE PySpark v2.4.x requires Java version 8.

Install Java 8 on MacOS

To install Java 8 on MacOS using asdf:

brew install asdf
asdf plugin add Java
asdf install java adoptopenjdk-8.0.265+1
asdf global java adoptopenjdk-8.0.265+1
. ~/.asdf/plugins/java/set-java-home.zsh

Then, we need to set the environment variables PYSPARK_PYTHON and PYSPARK_DRIVER_PYTHON to point to the conda python executable.

Click on the following menus to see details:

Set PySpark environment variables on Linux or MacOS

If you use conda, to set these variables every time the environment is activated, you can follow the steps of this guide.

First, assuming that the environment is called reco_pyspark, get the path where the environment is installed:

RECO_ENV=$(conda env list | grep reco_pyspark | awk '{print $NF}')
mkdir -p $RECO_ENV/etc/conda/activate.d
mkdir -p $RECO_ENV/etc/conda/deactivate.d

Then, create the file $RECO_ENV/etc/conda/activate.d/env_vars.sh and add:

#!/bin/sh
RECO_ENV=$(conda env list | grep reco_pyspark | awk '{print $NF}')
export PYSPARK_PYTHON=$RECO_ENV/bin/python
export PYSPARK_DRIVER_PYTHON=$RECO_ENV/bin/python
unset SPARK_HOME

This will export the variables every time we do conda activate reco_pyspark. To unset these variables when we deactivate the environment, create the file $RECO_ENV/etc/conda/deactivate.d/env_vars.sh and add:

#!/bin/sh
unset PYSPARK_PYTHON
unset PYSPARK_DRIVER_PYTHON
Set PySpark environment variables on Windows

To set these variables every time the environment is activated, we can follow the steps of this guide. First, get the path of the environment reco_pyspark is installed:

for /f "delims=" %A in ('conda env list ^| grep reco_pyspark ^| awk "{print $NF}"') do set "RECO_ENV=%A"

Then, create the file %RECO_ENV%\etc\conda\activate.d\env_vars.bat and add:

@echo off
for /f "delims=" %%A in ('conda env list ^| grep reco_pyspark ^| awk "{print $NF}"') do set "RECO_ENV=%%A"
set PYSPARK_PYTHON=%RECO_ENV%\python.exe
set PYSPARK_DRIVER_PYTHON=%RECO_ENV%\python.exe
set SPARK_HOME_BACKUP=%SPARK_HOME%
set SPARK_HOME=
set PYTHONPATH_BACKUP=%PYTHONPATH%
set PYTHONPATH=

This will export the variables every time we do conda activate reco_pyspark. To unset these variables when we deactivate the environment, create the file %RECO_ENV%\etc\conda\deactivate.d\env_vars.bat and add:

@echo off
set PYSPARK_PYTHON=
set PYSPARK_DRIVER_PYTHON=
set SPARK_HOME=%SPARK_HOME_BACKUP%
set SPARK_HOME_BACKUP=
set PYTHONPATH=%PYTHONPATH_BACKUP%
set PYTHONPATH_BACKUP=

Using a virtual environment

It is straightforward to install the recommenders package within a virtual environment. However, setting up CUDA for use with a GPU can be cumbersome. We thus recommend setting up Nvidia docker and running the virtual environment within a container, as the most convenient way to do this.

# Start docker daemon if not running
sudo dockerd &
# Pull the image from the Nvidia docker hub (https://hub.docker.com/r/nvidia/cuda) that is suitable for your system
# E.g. for Ubuntu 18.04 do
sudo docker run --gpus all -it --rm nvidia/cuda:10.0-cudnn7-runtime-ubuntu18.04

# Within the container: 

apt-get -y update
apt-get -y install python3.6
apt-get -y install python3-pip
apt-get -y install python3.6-venv
apt-get -y install libpython3.6-dev
apt-get -y install cmake
apt-get install -y libgomp1 openjdk-8-jre
export JAVA_HOME=/usr/lib/jvm/java-8-openjdk-amd64

python3.6 -m venv --system-site-packages /venv
source /venv/bin/activate
pip install --upgrade pip
pip install --upgrade setuptools

export PYSPARK_DRIVER_PYTHON=/venv/bin/python
export PYSPARK_PYTHON=/venv/bin/python

pip install recommenders[all]

If you prefer to use virtual-env instead of venv, you may follow the above steps, except you will need to replace the line

apt-get -y install python3.6-venv

with

python3.6 -m pip install --user virtualenv

and the line

python3.6 -m venv --system-site-packages /venv

with

python3.6 -m virtualenv /venv

Register the environment as a kernel in Jupyter

We can register our conda or virtual environment to appear as a kernel in the Jupyter notebooks. After activating the environment (my_env_name) do

python -m ipykernel install --user --name my_env_name --display-name "Python (my_env_name)"

If you are using the DSVM, you can connect to JupyterHub by browsing to https://your-vm-ip:8000.

Troubleshooting for the DSVM

  • We found that there can be problems if the Spark version of the machine is not the same as the one in the conda file. You can use the option --pyspark-version to address this issue.

  • When running Spark on a single local node it is possible to run out of disk space as temporary files are written to the user's home directory. To avoid this on a DSVM, we attached an additional disk to the DSVM and made modifications to the Spark configuration. This is done by including the following lines in the file at /dsvm/tools/spark/current/conf/spark-env.sh.

SPARK_LOCAL_DIRS="/mnt"
SPARK_WORKER_DIR="/mnt"
SPARK_WORKER_OPTS="-Dspark.worker.cleanup.enabled=true, -Dspark.worker.cleanup.appDataTtl=3600, -Dspark.worker.cleanup.interval=300, -Dspark.storage.cleanupFilesAfterExecutorExit=true"
  • Another source of problems is when the variable SPARK_HOME is not set correctly. In the Azure DSVM, SPARK_HOME is by default /dsvm/tools/spark/current. We need to unset it:
unset SPARK_HOME
  • Java 11 might produce errors when running the notebooks. To change it to Java 8:
sudo apt install openjdk-8-jdk
export JAVA_HOME="/usr/lib/jvm/java-8-openjdk-amd64"
  • We found that there might be conflicts between the current MMLSpark jars available in the DSVM and the ones used by the library. In that case, it is better to remove those jars and rely on loading them from Maven or other repositories made available by MMLSpark team.
cd /dsvm/tools/spark/current/jars
sudo rm -rf Azure_mmlspark-0.12.jar com.microsoft.cntk_cntk-2.4.jar com.microsoft.ml.lightgbm_lightgbmlib-2.0.120.jar

Setup guide for Azure Databricks

Requirements

  • Databricks Runtime version >= 4.3 (Apache Spark 2.3.1, Scala 2.11) and <= 5.5 (Apache Spark 2.4.3, Scala 2.11)
  • Python 3

An example of how to create an Azure Databricks workspace and an Apache Spark cluster within the workspace can be found from here. To utilize deep learning models and GPUs, you may setup GPU-enabled cluster. For more details about this topic, please see Azure Databricks deep learning guide.

Installation from PyPI

The recommenders package can be installed with core dependencies for utilities and CPU-based algorithms. This is done from the Libraries link at the cluster, selecting the option to import a library and selecting PyPI in the menu.
For installations with more dependencies, see the steps below.

Dependencies setup

You can setup the repository as a library on Databricks either manually or by running an installation script. Both options assume you have access to a provisioned Databricks workspace and cluster and that you have appropriate permissions to install libraries.

Quick install

This option utilizes an installation script to do the setup, and it requires additional dependencies in the environment used to execute the script.

To run the script, following prerequisites are required:

  • Setup CLI authentication for Azure Databricks CLI (command-line interface). Please find details about how to create a token and set authentication here. Very briefly, you can install and configure your environment with the following commands.

    conda activate reco_pyspark
databricks configure --token

  • Get the target cluster id and start the cluster if its status is TERMINATED.

    • You can get the cluster id from the databricks CLI with: 
      databricks clusters list
    • If required, you can start the cluster with: 
      databricks clusters start --cluster-id <CLUSTER_ID>`

The installation script has a number of options that can also deal with different databricks-cli profiles, install a version of the mmlspark library, overwrite the libraries, or prepare the cluster for operationalization. For all options, please see:

python tools/databricks_install.py -h

Once you have confirmed the databricks cluster is RUNNING, install the modules within this repository with the following commands.

cd Recommenders
python tools/databricks_install.py <CLUSTER_ID>

Note If you are planning on running through the sample code for operationalization here, you need to prepare the cluster for operationalization. You can do so by adding an additional option to the script run. <CLUSTER_ID> is the same as that mentioned above, and can be identified by running databricks clusters list and selecting the appropriate cluster.

python tools/databricks_install.py --prepare-o16n <CLUSTER_ID>

See below for details.

Manual setup

To install the repo manually onto Databricks, follow the steps:

  1. Clone the Microsoft Recommenders repository to your local computer.

  2. Zip the contents inside the Recommenders folder (Azure Databricks requires compressed folders to have the .egg suffix, so we don't use the standard .zip):

    cd Recommenders
zip -r Recommenders.egg .

  3. Once your cluster has started, go to the Databricks workspace, and select the Home button.

  4. Your Home directory should appear in a panel. Right click within your directory, and select Import.

  5. In the pop-up window, there is an option to import a library, where it says: (To import a library, such as a jar or egg, click here). Select click here.

  6. In the next screen, select the option Upload Python Egg or PyPI in the first menu.

  7. Next, click on the box that contains the text Drop library egg here to upload and use the file selector to choose the Recommenders.egg file you just created, and select Open.

  8. Click on the Create library. This will upload the egg and make it available in your workspace.

  9. Finally, in the next menu, attach the library to your cluster.

Confirm Installation

After installation, you can now create a new notebook and import the utilities from Databricks in order to confirm that the import worked.

import recommenders

Troubleshooting Installation on Azure Databricks

  • For the recommenders import to work on Databricks, it is important to zip the content correctly. The zip has to be performed inside the Recommenders folder, if you zip directly above the Recommenders folder, it won't work.

Prepare Azure Databricks for Operationalization

This repository includes an end-to-end example notebook that uses Azure Databricks to estimate a recommendation model using matrix factorization with Alternating Least Squares, writes pre-computed recommendations to Azure Cosmos DB, and then creates a real-time scoring service that retrieves the recommendations from Cosmos DB. In order to execute that notebook, you must install the Recommenders repository as a library (as described above), AND you must also install some additional dependencies. With the Quick install method, you just need to pass an additional option to the installation script.

Quick install

This option utilizes the installation script to do the setup. Just run the installation script with an additional option. If you have already run the script once to upload and install the Recommenders.egg library, you can also add an --overwrite option:

python tools/databricks_install.py --overwrite --prepare-o16n <CLUSTER_ID>

This script does all of the steps described in the Manual setup section below.

Manual setup

You must install three packages as libraries from PyPI:

  • azure-cli==2.0.56
  • azureml-sdk[databricks]==1.0.8
  • pydocumentdb==2.3.3

You can follow instructions here for details on how to install packages from PyPI.

Additionally, you must install the spark-cosmosdb connector on the cluster. The easiest way to manually do that is to:

  1. Download the appropriate jar from MAVEN. NOTE This is the appropriate jar for spark versions 2.3.X, and is the appropriate version for the recommended Azure Databricks run-time detailed above.
  2. Upload and install the jar by:
    1. Log into your Azure Databricks workspace
    2. Select the Clusters button on the left.
    3. Select the cluster on which you want to import the library.
    4. Select the Upload and Jar options, and click in the box that has the text Drop JAR here in it.
    5. Navigate to the downloaded .jar file, select it, and click Open.
    6. Click on Install.
    7. Restart the cluster.

Setup guide for Docker

A Dockerfile is provided to build images of the repository to simplify setup for different environments. You will need Docker Engine installed on your system.

Note: docker is already available on Azure Data Science Virtual Machine

See guidelines in the Docker README for detailed instructions of how to build and run images for different environments.

Example command to build and run Docker image with base CPU environment.

DOCKER_BUILDKIT=1 docker build -t recommenders:cpu --build-arg ENV="cpu" --build-arg VIRTUAL_ENV="conda" .
docker run -p 8888:8888 -d recommenders:cpu

You can then open the Jupyter notebook server at http://localhost:8888

Setup guide for making a release

The process of making a new release and publishing it to pypi is as follows:

First make sure that the tag that you want to add, e.g. 0.6.0, is added in recommenders.py/init.py. Follow the contribution guideline to add the change.

  1. Make sure that the code in main passes all the tests (unit and nightly tests).
  2. Create a tag with the version number: e.g. git tag -a 0.6.0 -m "Recommenders 0.6.0".
  3. Push the tag to the remote server: git push origin 0.6.0.
  4. When the new tag is pushed, a release pipeline is executed. This pipeline runs all the tests again (unit, smoke and integration), generates a wheel and a tar.gz which are uploaded to a GitHub draft release.
  5. Fill up the draft release with all the recent changes in the code.
  6. Download the wheel and tar.gz locally, these files shouldn't have any bug, since they passed all the tests.
  7. Publish the wheel and tar.gz to pypi: twine upload recommenders*