chastesweep 0.10

Parameter Sweeper for Chaste

Chaste Parameter Sweeper

A tool to assist with parameter sweeping for Chaste on the HPC (SGE & Slurm). The package provides parameter expansion, generation of task array batch script, individual job runner and a c++ main() function generator.

The parameter sweeping code was brought in from the pscan package and converted to python 2.x to ensure it works with the rest of Chaste's build dependencies.

While originally created for Chaste, it can also be applied to any other application that correctly accepts the parameters in the correct format (i.e. your program should accept parameters in the format of --varname value).

Installation

By pip

Install by using:

pip install chastesweep

From repository

# Clone the respository then install it
git clone https://github.com/twinkarma/chastesweep.git
cd chastesweep
python setup.py install

General Usage

After installing the chastesweep package, create a python script for defining your sweep parameters e.g. we create a sweep.py file:

import numpy as np
from chastesweep import ParamSweeper

# We'll use 3 parameters in this example, param0, param1 and param2
# we create a dictionary and declare what each parameter values are
p = {}
p['param0'] = np.linspace(0, 10, 5)
p['param1'] = np.linspace(0.1, -0.5, 5)
p['param2'] = [10, 20, 30]

# We must also declare the execuatable that will run with the paramters
exec_cmd = "my_executable.sh"

# And the output directory 
output_dir = "sweep_results"

# We then create the parameter sweeper and have it generate the batch file in the output directory
sweeper = ParamSweeper()
sweeper.generate_batch_output(output_dir=output_dir,
                              exec_cmd=exec_cmd,
                              parameters=p,
                              scheduler=ParamSweeper.SGE,
                              batch_params=["-M [e-mail]", "-m bes"])

We then run sweep.py:

python sweep.py

In the output directory sweep_results, you will see the params.json, runsimulation.py and batch.sge.sh file.

• params.json Contains an expanded list of parameters that will be explored.
• batch.sge.sh Batch script containing a task array for running through all of the parameters to be explored.
• runsimulation.py Simulation runner script for running individual instances of the simulation.

You can submit this file to the SGE scheduler to start your parameter sweeping task:

qsub sweep_results/batch.sge.sh

If we're running a SLURM scheduler, simply change the scheduler=ParamSweeper.SGE to scheduler=ParamSweeper.SLURM. You'll also want to change your batch_params parameters as they'll likely be different from SGE. Your batch submit command will also change to:

sbatch sweep_results/batch.slurm.sh

Expanding parameters

The following is a demonstration of how parameters can be expanded. All examples also apply to generate_batch_output.

To only get the parameter expansion, call the expand_parameters function:

import numpy as np
from chastesweep import ParamSweeper

# We'll use 3 parameters in this example, param0, param1 and param2
# we create a dictionary and declare what each parameter values are
p = {}
p['param0'] = np.linspace(0, 10, 5)
p['param1'] = np.linspace(0.1, -0.5, 5)
p['param2'] = [10, 20, 30]
# We can then call the exp
sweeper = ParamSweeper()
results = sweeper.expand_parameters(parameters=p)

Joining Parameters

By default, parameters are combinatorially expanded. To prevent this expansion between some parameters, the parameters can be joined together by providing a join list. For example if you'd like to join param0 and param1 in the above example:

# We create a join list
join_lists = [['param0', 'param1']]
# Add the join list to the call parameters
results = sweeper.expand_parameters(parameters=p, joint_lists=join_lists)

You'll see that our results list has 15 items instead of 75 if we didn't include the join list.

Globally running each parameter set more than once

You can set the default_repeats to specify globally how many iterations of each parameter set will run, e.g. the value of 2:

results = sweeper.expand_parameters(parameters=p, default_repeats=2)

Result in 150 simulation runs.

Variable number of iteration for each parameter set

Instead of setting the number of iterations globally, the number of iteration per parameter set can vary by declaring count functions:

# The default count function, runs 2 iterations of every parameter
default_count = lambda p: 2

# A big_count function that forces the iteration to 1 if param 2 is larger than 13
big_count = lambda p: 1 if p['param2'] >= 14 else None

# The order of counters makes a difference, the next counter in the sequence can overwrite previous counts
count_funcs = [default_count, big_count]

results = sweeper.expand_parameters(parameters=p, count_funcs=count_funcs)

Running the sweep locally

Sweep can also be run locally on your machine with the perform_serial_sweep function:

import numpy as np
from chastesweep import ParamSweeper
p = {}
p['a'] = np.linspace(0, 10, 5)
p['b'] = np.linspace(0.1, -0.5, 5)
p['c'] = [10, 20, 30]

sweeper = ParamSweeper()
exec_cmd = "my_executable.sh"
output_dir = "sweep_results"
sweeper.perform_serial_sweep(output_dir=output_dir, exec_cmd=exec_cmd, parameters=p)

Parameter Sweeping Tutorial (Sheffield HPC)

In this tutorial we will go through how to setup chaste for parameter sweeping on the HPC cluster. Run this tutorial directly on the cluster to be able to follow all examples including job submission.

1. Getting a compute node

When logging in to ShARC or Bessemer, you'll start off on a login node. You'll want to get an interactive session on a worker node to start development:

# On ShARC
qrshx

# Bessemer
srun --pty bash -i

You'll see the console change from login to node e.g. [username@sharc-login#] $ to [username@sharc-node###] $ in the case of ShARC.

2. Install chastesweep python package

On ShARC

It is recommended to create a virtual environment on ShARC:

# Load the python module
module load apps/python/anaconda3-4.2.0

# Create a new virtual python environment called my_chaste_env with python 2.7
conda create -n my_chaste_env python=2.7

# Activate the environment
source activate my_chaste_env

# Install the chastesweep package
pip install chastesweep

On Bessemer

Equivalent code on Bessemer

# Load the python module
module load Anaconda3/5.3.0

# Create a new virtual python environment called my_chaste_env with python 2.7
conda create -n my_chaste_env2 python=2.7 chastesweep

# Activate the environment
source activate my_chaste_env2

# Install the chastesweep package
pip install chastesweep

3. Development Environment, pulling a a Singularity image

ShARC and Bessemer supports the Singularity containerisation technology. The use of containers allow us to standardise the development environment and build dependencies on your local machine and your cluster.

Note: On your local machine, if you're not developing on Linux, it is also possible to use docker instead. You will want to install either Singularity or docker on your local machine.

To get the latest Singularity image of Chaste, run the following:

singularity pull docker://chaste/chaste-docker:latest

You should see a file chaste-docker_latest.sif.

Run the following to get into the image's bash shell:

singularity exec chaste-docker_latest.sif /bin/bash

You're now actually inside the image, try running the command cat /etc/lsb-release to see the current Ubuntu version e.g.:

cat /etc/lsb-release
DISTRIB_ID=Ubuntu
DISTRIB_RELEASE=19.04
DISTRIB_CODENAME=disco
DISTRIB_DESCRIPTION="Ubuntu 19.04"

You can run exit to exit from the image.

Singularity's exec command can be used to run any program from inside the image without needing to go into the interactive session. This is useful when we're running our tasks in a batch script. For example, to get the Ubuntu version as above:

singularity exec chaste-docker_latest.sif cat /etc/lsb-release

You will be using the image you've just pulled to do the building and running of the code using the exec command as shown above.

Note: As you will only be using the container for its build dependency, you will not need to modify the image during development and so should not need admin permission.

4. Preparing your code

You now need your own version of Chaste code. The Chaste repository can be found at https://github.com/Chaste/Chaste and this can be forked or cloned directly depending on your requirements. We will just clone directly in this example:

# Clone the Chaste source code into our home folder
git clone https://github.com/Chaste/Chaste.git ~/Chaste

# Go inside the folder
cd ~/Chaste

The above command clones Chaste code into the Chaste folder to your home directory. You will now need create a custom Chaste app or project with a main() function to be able to accept parameters.

The the instructions on how to create Chaste executable apps and user project can be followed below:

• Make executable apps: https://chaste.cs.ox.ac.uk/trac/wiki/ChasteGuides/BuildingExecutableApps
• User projects: https://chaste.cs.ox.ac.uk/trac/wiki/ChasteGuides/UserProjects

For this example we'll go with the executable app route due to the simpler setup. We'll ask the tool to generate the .cpp file for us.

We'll call our app ParamSweep and will use three parameters for this example, named param0, param1 and param2. To automatically generate a template using the parameter sweeper tool, call:

chastesweep_genmain param0,param1,param2 apps/src/ParamSweep.cpp

The generated source code is printed on screen as well as created for you.

We can then try to build the project using cmake:

# Go into the build directory
cd build

# Generate makefiles using cmake from inside the Chaste image
singularity exec chaste-docker_latest.sif cmake ..

# Build our ParamSweep app from inside the Chaste image
singularity exec chaste-docker_latest.sif make ParamSweep

After the build has finished, we can now run our app:

singularity exec chaste-docker_latest.sif apps/ParamSweep

Your code should execute as is but you'll get errors stating that parameters are missing.

5. Passing parameters to the main function

The parameter sweeper follows the boost::program_options format, where each parameter is passed in the format of --param_name param_value e.g. for our ParamSweep app:

singularity exec chaste-docker_latest.sif apps/ParamSweep --output_dir "/outtput/path/1" --param0 0.2 --param1 3 --param2 -0.53 

Note: The output_dir parameter is always passed to your app and all outputs of the app should be enforced to only write to this directory.

6. Declaring parameters and generating the batch script

Now that we have an executable ready for accepting parameters, we'll move on to declaring the values of parameters to be explored and generating batch script for submitting a batch task. For this purpose, python API has been created, the class ParamSweeper has been created to facilitate this parameter sweeping process.

We create a python script named sweep.py for defining our parameter and creating a batch job:

import numpy as np
from chastesweep import ParamSweeper
p = {}
p['param0'] = np.linspace(0, 10, 5)
p['param1'] = np.linspace(0.1, -0.5, 5)
p['param2'] = [10, 20, 30]

exec_cmd = "~/Chaste/build/apps/ParamSweep"
output_dir = "~/Chaste/build/sweep_results"

# Generate batch for SGE
sweeper = ParamSweeper()
sweeper.generate_batch_output(output_dir=output_dir,
                              exec_cmd=exec_cmd,
                              parameters=p,
                              scheduler=ParamSweeper.SGE,
                              batch_params=["-M [e-mail]", "-m bes"])

Notice that we can additional parameters to the batch script by adding to batch_params. In our example we've set the scheduler to alert us of the job status -m bes and provided our e-mail -M [e-mail].

We then run the sweep.py script:

python sweep.py

In the output directory ~/Chaste/build/sweep_results, you will see the params.json, runsimulation.py and batch.sge.sh file.

• params.json Contains an expanded list of parameters that will be explored.
• batch.sge.sh Batch script containing a task array for running through all of the parameters to be explored.
• runsimulation.py Simulation runner script for running individual instances of the simulation.

You can submit this file to the SGE scheduler to start your parameter sweeping task:

qsub ~/Chaste/build/sweep_results/batch.sge.sh

If we're running on Bessemer or other cluster that uses SLURM scheduler, simply change the scheduler=ParamSweeper.SGE to scheduler=ParamSweeper.SLURM. You'll also want to change your batch_params parameters as they'll likely be different from SGE. Your batch submit command will also change to:

sbatch ~/Chaste/build/sweep_results/batch.slurm.sh