Parallel (super) computing with scikit-learn

I was granted an access to the super computer of the University of Strasbourg, which is perfect for tackling machine learning challenges where many models need to be tried.

Using scikit-learn, it is relatively easy to take advantage of the multiple cores of a PC with the n_jobs option. Parallel computing on a remote machine using the Slurm queuing system is not as straightforward.

I want to achieve the following:
* Use python 3 and be able to install any library
* Run jobs in parallel for any custom function

With a little search and testing, I came up with the following solution.

Connecting and setting up your environment

After registering, you will obtain a username and a password. If you are using Ubuntu like me, connecting through ssh is easy. Open a terminal and type

ssh -Y username@hpc-login.u-strasbg.fr

followed by your password. You can also use sshpass if you don’t want to type your password in blind mode.

Installing Anaconda

Anaconda is a widely used python package manager. It is pre-installed on the HPC and could be activated with load module python/Anaconda3. However, I was unable to add new libraries using the conda install command. Luckily, it can be installed without admin rights!
Simply do the following in the terminal:

wget "https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh"
bash Miniconda3-latest-Linux-x86_64.sh #installs in ~/conda
export PATH=$HOME/conda/bin:$PATH
conda install scikit-learn -y

you can add any python package you need with conda install.

Running tests in parallel

workflow

My workflow consists in developping on my PC and testing with little data. When I want to use the whole train data set, I can then transfer everything using scp.

For testing purpose, you can make a test_parallel directory on both your PC and the remote super computer.

mkdir test_parallel
cd ./test_parallel

Python code

We are going to use the standard “digits” data set included as an example in scikit-learn. We will not try to optimize for the perfect model, but only show how a custom function to classify this data set can be tested with several parameters in parallel.

As I said earlier, it is not as straightforward as calling the option n_jobs in most sklearn classes. We will need to use the package ipyparallel from the ipython project. Don’t forget to install it using conda install ipyparallel!

It requires functions to be pickle-able, so we will have to create two files: one containing your script script.py and one containing your functions some_funcs.py.

Note: unless you implement your own algorithm, using functions from the sklearn.model_selection rather than a custom function is certainly much better. For debugging and testing purpose, home-made functions can let you write information into files.

script.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Run it with:
    python3 script.py -p ipy_profile
where ipy_profile is the name of the ipython profile.
@author: hyamanieu
"""
import argparse
import logging
import os
import sys
from sklearn.externals.joblib import Parallel, parallel_backend
from sklearn.externals.joblib import register_parallel_backend
from sklearn.externals.joblib import delayed
from sklearn.externals.joblib import cpu_count
from sklearn.datasets import load_digits
from sklearn.model_selection import train_test_split
from ipyparallel import Client
from ipyparallel.joblib import IPythonParallelBackend
import numpy as np
import datetime
#module in the same directory
from some_funcs import SVC_rbf

FILE_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.append(FILE_DIR)

#prepare the logger
parser = argparse.ArgumentParser()
parser.add_argument("-p", "--profile", default="ipy_profile",
                 help="Name of IPython profile to use")
args = parser.parse_args()
profile = args.profile
logging.basicConfig(filename=os.path.join(FILE_DIR,profile+'.log'),
                    filemode='w',
                    level=logging.DEBUG)
logging.info("number of CPUs found: {0}".format(cpu_count()))
logging.info("args.profile: {0}".format(profile))

#prepare the engines
c = Client(profile=profile)
#The following command will make sure that each engine is running in
# the right working directory to access the custom function(s).
c[:].map(os.chdir, [FILE_DIR]*len(c))
logging.info("c.ids :{0}".format(str(c.ids)))
bview = c.load_balanced_view()
register_parallel_backend('ipyparallel', 
                          lambda : IPythonParallelBackend(view=bview))

#Get data
digits = load_digits()
#prepare it for the custom function
X_train, X_test, y_train, y_test = train_test_split(digits.data,
                                                    digits.target,
                                                    test_size=0.3)
#some parameters to test in parallel
param_space = {
    'C': np.logspace(-6, 6, 12),
}

gamma = 0.1

with parallel_backend('ipyparallel'):
    scores = Parallel(n_jobs=len(c))(delayed(SVC_rbf)(C,
                      gamma,
                      X_train, 
                      X_test, 
                      y_train, 
                      y_test, 
                      profile)
                               for C in param_space['C'])
    #write down the scores in a file.
with open('scores_rbf_digits.csv', 'w') as f:
    f.write('C,gamma,MAE\n')
    f.write("\n".join(','.join(str(c) for c in l) for l in scores))
    f.write('\n')

Make sure to import joblib functions from sklearn.external, not from joblib itself.

If you prefer using the GridSearchCV from sklearn rather than a custom function (strongly advised), you can replace the last bit with:

from sklearn.model_selection import GridSearchCV
svc_rbf = SVC(kernel='rbf',
              shrinking=False)
search = GridSearchCV(svc_rbf, 
                      param_space,
                      return_train_score=True,
                      n_jobs=len(c))
with parallel_backend('ipyparallel'):
    search.fit(X_train, y_train)
results = search.cv_results_
results = pd.DataFrame(results)
results.to_csv(os.path.join(FILE_DIR,'scores_rbf_digits.csv'))

some_funcs.py

import os
import datetime
from sklearn import metrics
from sklearn.svm import SVC
import numpy as np
now = datetime.datetime.now

FILE_DIR = os.path.dirname(os.path.abspath(__file__))

def SVC_rbf(C,gamma, X_train, X_test, y_train, y_test, profile):
    file_path = os.path.join(os.getcwd(),
                             '{0}_C{1:06}.log'.format(profile,C))
    #logging will not work from the HPC engines
    #need to write into a file manualy.
    with open(file_path, 'a+') as f:
        f.write('new task for C='+str(C)+' and gamma='+str(gamma)+'\n')

    t0 = now()
    with open(file_path, 'a+') as f:
        f.write('Start learning at {0}\n'.format(t0.isoformat()))
    svc_rbf_reg = SVC(kernel='rbf',
                       C=C,
                       gamma=gamma, 
                       shrinking=False)
    svc_rbf_reg.fit(X_train,y_train)
    predictions = svc_rbf_reg.predict(X_test)
    score = metrics.mean_absolute_error(y_test,predictions)

    ##uncomment if you want to make the function slower
    #with open(file_path, 'a+') as f:
    #    f.write('Compute complex matrix inverse')
    #    m = np.random.randn(10000,10000)
    #    np.linalg.inv(m)

    t1 = now()
    h = (t1-t0).total_seconds()//3600
    m = (t1-t0).total_seconds()//60 - h*60
    s = (t1-t0).total_seconds() -m*60 - h*60
    with open(file_path, 'a+') as f:
        f.write('Finished at {0} after '
                '{1}h {2}min {3:0.2f}s\n'.format(t1.isoformat(),
                                                 h,m,s))
        f.write('SVC rbf\n')
        f.write('C: {0}, gamma: {1}\n'.format(str(C),str(gamma)))
        f.write('MAE: {0}\n'.format(str(score)))


    return (C, gamma, score)

And voila!

Run your code locally

Let’s first test it on your own PC. You’ll need to open 1+n terminals, where n is the number of cores you can to use and run:

ipcontroller --profile=ipy_profile
ipengine --profile=ipy_profile (*n)

Alternatively, you can also run a cluster in a single terminal with

ipcluster start --profile=ipy_profile &

It will automatically start as many engines as there are cores. You can then stop it by typing:

ipcluster stop --profile=ipy_profile

You can find other options in the documentation. You can now run your python script with python3 script.py. If you used ipcontroller+ipengine(s), you should see the controller dispatching the jobs to the engines. Also, the log files will be created.

Run it on the remote HPC

The first thing you need to do is to copy both python files to the HPC:

#run this on your local computer
scp *.py username@hpc-login.u-strasbg.fr:~/test_parallel

Now you need to make a bash file to start the engines, your script, and define the needed resources in the header for Slurm.

#run this after ssh-ing into the HPC
cd ~/test_parallel
nano start_script.sh

You can then insert:

#! /bin/bash
#SBATCH -p public           #available partition, depends on your HPC 
#SBATCH -J ipy_engines      #job name
#SBATCH -n 4                # 4 cores, you can increase it
#SBATCH -N 1                # 1 node, you can increase it
#SBATCH -t 1:00:00         # Job is killed after 1h
#SBATCH --mail-type=END     # Sends an email when the job finishes
#SBATCH --mail-user=nickname@domain.com#your email address

#the following deletes all the ipython profiles older than 1 day.
find ~/.ipython/profile_job* -maxdepth 0 -type d -ctime +1 | xargs rm -r
#create a new ipython profile appended with the job id number
profile=job_${SLURM_JOB_ID}

echo "Creating profile_${profile}"
$HOME/conda/bin/ipython profile create ${profile}

$HOME/conda/bin/ipcontroller --ip="*" --profile=${profile} &
sleep 10

#srun: runs ipengine on each available core
srun $HOME/conda/bin/ipengine --profile=${profile} --location=$(hostname) &
sleep 25

echo "Launching job for script $1"
$HOME/conda/bin/python $1 -p ${profile}

I have allocated only 4 cores to compare the performance with my PC with similar resources.

Now exit (ctrl+X) and run it:

sbatch start_script.sh script.py

You can check where it is in the queue with queue -u username. It should not take that long: 4 cores and 1 hours computing time is rather little, it should be put on priority. The script should finish within a couple minutes only.

You can now compare the results on each by reading the *.log files! Once you’re done with testing, I would advise using GridSearchCV and run many more parameters than only 12! There will be a single log file and of course the results in a .csv.

Next

I am currently using the Alsacalcul HPC to find a solution from one of the ENS challenges, check them out! I will post my first solution soon, as well as how I combined working on the jupyter notebook and ran calculations on the HPC.

some sources I used:
https://services-numeriques.unistra.fr/les-services-aux-usagers/hpc.html
http://www.makeloft.org/2016/05/work-in-progress-parallel-ipython-from.html
http://twiecki.github.io/blog/2014/02/24/ipython-nb-cluster/
https://github.com/scikit-learn/scikit-learn/issues/7168 (ivdorelian 10 Aug 2016)
https://stackoverflow.com/questions/38601026/easy-way-to-use-parallel-options-of-scikit-learn-functions-on-hpc
image copyright owner: AlsaCalcul Services, http://alsacalcul.unistra.fr/

2 thoughts on “Parallel (super) computing with scikit-learn

  1. This was extremely helpful for me. Thank you for taking the time to make this tutorial.

Leave a Reply to hyamanieu Cancel reply

Your email address will not be published. Required fields are marked *