TP: Distributed Optimization and Games

1. Create a private account for the Inria nef cluster here [Please do it as soon as possible, as it may take time to create the account]:

   1. • for company : please choose "other"

      • for mail address : please use UCA email (no gmail, etc.)

      • expiration date : 2020-02-28

      • ssh public key : generate ssh key (Linux, Windows, macOS) and attach the public one

reminder - openssh format is needed for the key (windows/putty users will need to convert the key to openssh format - tutorial)

• • • usecase description : please indicate "UCA master of Data Science hands on"

    • account creation may require several days, a confirmation email will be sent once the account is created

2. Log in to the cluster once the account is created:

• • • ssh username@nef-frontal.inria.fr

    • ssh username@nef-devel.inria.fr

    • you may try to transfer your own local file from computer to the cluster directory by: scp localfile username@nef-frontal.inria.fr:~/

3. Reserve computing resources in the cluster and log into the reserved node:

• • • oarsub -l /nodes=1/core=2,walltime=1 -I (reserving 2 CPU cores from one node for one hour)

    • more variants for the commands

    • hardware details of the cluster

3. Get Pytorch prepared: mutiple ways , recommend Method 3.

• • • to check if Pytorch is well installed, from the terminal type:

      • • python

        • import torch

        • print(torch.__version__)

4. Python editor in cluster, recommend to use VIM and know the basic operation in advance.

5. If you have any question concerning the previous points, please contact me: chuan.xu@inria.fr.

6. Please install Pytorch on your own computer as well: "conda install pytorch torchvision cudatoolkit=10.1 -c pytorch"

1) To run Basic_Machine_Learning code using command: python Basic_Machine_Learning.py

2) If the torchvision package is not installed, please following the steps:

• Check the python version by "which python": the return should be "/misc/opt/conda3-5.0.1/bin/python". If not, redo the method 3

• Check the pip version by "which pip": the return should be "/misc/opt/conda3-5.0.1/bin/pip"

• Install torchvision by: pip install --user torchvision

• Then it should be ok, if not, please contact me by mail.

3) The paper on the Parameter Server Algorithm is available here. (See algorithm 1 in page 586)

4) The answers for the Ex I (5) is uploaded to here. The answer for Ex II will be uploaded to the same place on 17/02.

5) Answer to Ex II is updated here, which you could access directely by updating git directory in the cluster by: git pull origin master.

Then go to the /answers/ directory and run: ./PS.sh script to launch the program.