soops = scoop output of parametric studies
Utilities to run parametric studies in parallel using dask, and to scoop the output files produced by the studies into a pandas dataframe.
Contents
- Installation
- Testing
- Example
- A Script
- Basic Run
- Parameterization
- Run Parametric Study
- Using Parametric Study Configuration Files
- Show Parameters Used in Each Output Directory
- Scoop Outputs of the Parametric Study
- Post-processing Plugins
- Notes
- Special Argument Values
- Generated Arguments
- Computed Arguments
- Find Runs with Given Parameters
- See Also
The latest release:
pip install soops
The source code of the development version in git:
git clone https://github.com/rc/soops.git cd soops pip install .
or the development version via pip:
pip install git+https://github.com/rc/soops.git
Install pytest:
pip install pytest
Install soops from sources (in the current directory):
pip install .
Run the tests (in any directory):
python -c "import soops; soops.test()"
Run tests in the source directory without installing soops:
export PYTHONPATH=. python -c "import soops; soops.test()" # or pytest soops/tests
Before we begin - TL;DR:
- Run a script in parallel with many combinations of parameters.
- Scoop all the results in many output directories into a big
DataFrame. - Work with the
DataFrame.
Suppose we have a script that takes a number of command line arguments. The actual arguments are not so important, neither what the script does. Nevertheless, to have something to work with, let us simulate the Monty Hall problem in Python.
For the first reading of the example below, it is advisable not to delve in details of the script outputs and code listings and just read the text to get an overall idea. After understanding the idea, return to the details, or just have a look at the complete example script.
This is our script and its arguments:
$ python soops/examples/monty_hall.py -h
usage: monty_hall.py [-h] [--switch] [--host {random,first}] [--num int]
[--repeat int] [--seed int] [--plot-opts dict-like] [-n]
[--silent]
output_dir
The Monty Hall problem simulator parameterized with soops.
https://en.wikipedia.org/wiki/Monty_Hall_problem
<snip>
positional arguments:
output_dir output directory
options:
-h, --help show this help message and exit
--switch if given, the contestant always switches the door,
otherwise never switches
--host {random,first}
the host strategy for opening doors
--num int the number of rounds in a single simulation [default:
100]
--repeat int the number of simulations [default: 5]
--seed int if given, the random seed is fixed to the given value
--plot-opts dict-like
matplotlib plot() options [default:
"linewidth=3,alpha=0.5"]
-n, --no-show do not call matplotlib show()
--silent do not print messages to screen
A run with the default parameters:
$ python soops/examples/monty_hall.py output monty_hall: num: 100 monty_hall: repeat: 5 monty_hall: switch: False monty_hall: host strategy: random monty_hall: elapsed: 0.004662119084969163 monty_hall: win rate: 0.25 monty_hall: elapsed: 0.0042096920078620315 monty_hall: win rate: 0.3 monty_hall: elapsed: 0.003894180990755558 monty_hall: win rate: 0.31 monty_hall: elapsed: 0.003928505931980908 monty_hall: win rate: 0.35 monty_hall: elapsed: 0.0035342529881745577 monty_hall: win rate: 0.31
produces some results:
Now we would like to run it for various combinations of arguments and their values, for example:
- --num=[100,1000,10000]
- --repeat=[10,20]
- --switch either given or not
- --seed either given or not, changing together with --seed
- --host=['random', 'first']
and then collect and analyze the all results. Doing this manually is quite tedious, but soops can help.
In order to run a parametric study, first we have to define a function describing the arguments of our script:
def get_run_info():
# script_dir is added by soops-run, it is the normalized path to
# this script.
run_cmd = """
{python} {script_dir}/monty_hall.py {output_dir}
"""
run_cmd = ' '.join(run_cmd.split())
# Arguments allowed to be missing in soops-run calls.
opt_args = {
'--num' : '--num={--num}',
'--repeat' : '--repeat={--repeat}',
'--switch' : '--switch',
'--host' : '--host={--host}',
'--seed' : '--seed={--seed}',
'--plot-opts' : '--plot-opts={--plot-opts}',
'--no-show' : '--no-show',
'--silent' : '--silent',
}
output_dir_key = 'output_dir'
is_finished_basename = 'wins.png'
return run_cmd, opt_args, output_dir_key, is_finished_basenameThe get_run_info() functions should provide four items:
- A command to run given as a string, with the non-optional arguments and
their values (if any) given as
str.format()keys. - A dictionary of optional arguments and their values (if any) given as
str.format()keys. - A special format key, that denotes the output directory argument of the command. Note that the script must have an argument allowing an output directory specification.
- A function
is_finished(pars, options), where pars is the dictionary of the actual values of the script arguments and options are soops-run options, see below. The dictionary contains the output directory argument of the script and the function should return True, whenever the results are already present in the given output directory. Instead of a function, a file name can be given, as in get_run_info() above. Then the existence of a file with the specified name means that the results are present in the output directory.
Putting get_run_info() into our script allows running a parametric study using soops-run:
$ soops-run -h
usage: soops-run [-h] [--dry-run] [-r {0,1,2}] [-n int]
[--run-function {subprocess.run,psutil.Popen,os.system}]
[-t float]
[--generate-pars dict-like: function=function_name,par0=val0,... or str]
[-c key1+key2+..., ...]
[--compute-pars dict-like: class=class_name,par0=val0,...]
[-s str] [--silent] [--shell] [-o path]
conf run_mod
Run parametric studies.
positional arguments:
conf a dict-like parametric study configuration or a study
configuration file name
run_mod the importable script/module with get_run_info()
options:
-h, --help show this help message and exit
--dry-run perform a trial run with no commands executed
-r {0,1,2}, --recompute {0,1,2}
recomputation strategy: 0: do not recompute, 1:
recompute only if is_finished() returns False, 2:
always recompute [default: 1]
-n int, --n-workers int
the number of dask workers [default: 2]
--run-function {subprocess.run,psutil.Popen,os.system}
function for running the parameterized command
[default: subprocess.run]
-t float, --timeout float
if given, the timeout in seconds; requires setting
--run-function=psutil.Popen
--generate-pars dict-like: function=function_name,par0=val0,... or str
if given, generate values of parameters using the
specified function; the generated parameters must be
set to @generate in the parametric study
configuration. Alternatively, a section key in a study
configuration file.
-c key1+key2+..., ..., --contract key1+key2+..., ...
list of option keys that should be contracted to vary
in lockstep
--compute-pars dict-like: class=class_name,par0=val0,...
if given, compute additional parameters using the
specified class
-s str, --study str study key when parameter sets are given by a study
configuration file
--silent do not print messages to screen
--shell run ipython shell after all computations
-o path, --output-dir path
output directory [default: output]
In our case (the arguments with no value (flags) can be specified either as
'@defined' or '@undefined'):
soops-run -r 1 -n 3 -c='--switch + --seed' -o output "python='python3', output_dir='output/study/%s', --num=[100,1000,10000], --repeat=[10,20], --switch=['@undefined', '@defined', '@undefined', '@defined'], --seed=['@undefined', '@undefined', 12345, 12345], --host=['random', 'first'], --silent=@defined, --no-show=@defined" soops/examples/monty_hall.py
This command runs our script using three dask workers (-n 3 option) and
produces a directory for each parameter set:
$ ls output/study/ 000-7a6b546a625c2d37569346a286f2b2b6/ 024-6f9810a492faf793b80de2ec32dec4b1/ 001-1daf48cede910a9c7c700fb78ce3aa2d/ 025-a4d05c2889189c4e086f9d6f56e1ba1d/ 002-57c1271f4b9cbe00742e3c97e0c14e24/ 026-67a251e1c40f65bae8bbf621c4e1a987/ 003-2f828633fa9eefa8eb8b40873882247d/ 027-9e3d30603d2b382256f62fdf17bc23ae/ 004-24f370388496173d8e1d7a9e574262e0/ 028-6ff18af0333367a65ed131d210078653/ 005-7893091a6fedc4ccdf7d73d803a91687/ 029-54d77d99e74402a043af583ac1e14c4e/ 006-70132dc423f26c78f1d2e33f0607820c/ 030-4bad1e59de5b446e80a621fdfb5fb127/ 007-7e5ecb11154e4c402caa51878e283e63/ 031-d65b7afd4d43b3159b580cf6c974a26c/ 008-201e1ab3e47d3b994f2d6532859ac301/ 032-cd83aafc620d81b994f005c6a7b1d2c4/ 009-35105e72d8ec2ddfd8adc8ffa8c1f088/ 033-e065bfc2596f3b285877e36578d77cce/ 010-ff68ea026e0efba0e4c2a71d64e12f2c/ 034-0533ff015142c967f86b365076fcee18/ 011-217e45abc1d2b188b0755fc6a550dfe9/ 035-f127408b640dae1de6acc9bce1b68669/ 012-d6adcade17e2d7d843cbd8e14aebf76a/ 036-56654b678decdd2d77ecc07ead326ad7/ 013-cdff71cb542f8159ff5c5a023c91f61c/ 037-d3d16497570cb3f934e73c3f0c519822/ 014-551f32ba477c7e8e8fad0769ac793d3c/ 038-5b3b21be9e6dbbd5c7d8e031bd621717/ 015-856ad0b4ee0273da8cd8ad3cf222077b/ 039-d11e877087ec25fe2c8062708687204c/ 016-7eb991928b39b40c98e7cb7970d0f15b/ 040-5cf056a63f2e10ee78d599e097eb4d0e/ 017-9a3f4b32f5ba30ec173dd651c9810c6e/ 041-ca696dc0edbe70890f2dcbcfcf99fe47/ 018-9067a6dbbb4afaf285f5c9101fa5fa73/ 042-9962ccd67846d21245580de2c5e83bcc/ 019-03a0123bd55725fdabec32e0aeff9d44/ 043-18503a94bf6398644e2a32d3a93e9450/ 020-266ed9d092128d8e3c3c2f78669a0425/ 044-6c46f7a9e9cd0b50d914d6e2a188a64d/ 021-00a156df6ccecab8d35c5bdc5ddb6c0e/ 045-0af51ef33a80a99ac38bfbac10fea9b2/ 022-91f0d18a4d9cd2e6721d937c9de4dbe9/ 046-746823fee6450a294869dc9ca7396e15/ 023-e3edef5a83fe941c75df4257ac056ca5/ 047-f9046e62d8da3159dfcdebcf687092f3/
The directory names consist of an integer allowing an easy location and a MD5 hash of the run parameters. In each directory, there are four files:
$ ls output/study/000-7a6b546a625c2d37569346a286f2b2b6/ options.txt output_log.txt soops-parameters.csv wins.png
three just like in the basic run above, and soops-parameters.csv, where the run parameters (mostly command line arguments) are stored by soops-run. For convenience, parameters of all runs are collected in all_parameters.csv in the soops-run output directory (output by default), using the data in all soops-parameters.csv files found.
Our example script also stores the values of command line arguments in
options.txt for possible re-runs and inspection:
$ cat output/study/000-7a6b546a625c2d37569346a286f2b2b6/options.txt
command line
------------
"soops/examples/monty_hall.py" "output/study/000-7a6b546a625c2d37569346a286f2b2b6" "--num=100" "--repeat=10" "--host=random" "--no-show" "--silent"
options
-------
host: random
num: 100
output_dir: output/study/000-7a6b546a625c2d37569346a286f2b2b6
plot_opts: {'linewidth': 3, 'alpha': 0.5}
repeat: 10
seed: None
show: False
silent: True
switch: False
Instead of providing the parameter sets on the command line, a study configuration file can be used. Then the same parametric study as above can be run using:
soops-run -r 1 -n 3 -c='--switch + --seed' --study=study -o output soops/examples/studies.cfg soops/examples/monty_hall.py
where soops/examples/studies.cfg contains:
[study] python='python3' output_dir='output/study/%s' --num=[100,1000,10000] --repeat=[10,20] --switch=['@undefined', '@defined', '@undefined', '@defined'] --seed=['@undefined', '@undefined', 12345, 12345] --host=['random', 'first'] --silent=@defined --no-show=@defined
Several studies can be stored in a single file, see soops/examples/studies.cfg. See also the docstring of soops/examples/monty_hall.py for more examples.
Use soops-info to explain which parameters were used in the given output
directories:
$ soops-info -h
usage: soops-info [-h] [-e dirname [dirname ...]] [--shell] run_mod
Get parametric study configuration information.
positional arguments:
run_mod the importable script/module with get_run_info()
optional arguments:
-h, --help show this help message and exit
-e dirname [dirname ...], --explain dirname [dirname ...]
explain parameters used in the given output
directory/directories
--shell run ipython shell after all computations
$ soops-info soops/examples/monty_hall.py -e output/study/000-7a6b546a625c2d37569346a286f2b2b6/ info: output/study/000-7a6b546a625c2d37569346a286f2b2b6/ info: finished: True info: * --host: random info: * --no-show: @defined info: * --num: 100 info: * --plot-opts: @undefined info: * --repeat: 10 info: * --seed: @undefined info: * --silent: @defined info: * --switch: @undefined info: * python: python3 info: output_dir: output/study/000-7a6b546a625c2d37569346a286f2b2b6 info: script_dir: examples
A * denotes a parameter used in the parameterization of the example script, other parameters are employed by soops-run.
In order to use soops-scoop to scoop/collect outputs of our parametric
study, a new function needs to be defined:
import soops.scoop_outputs as sc
def get_scoop_info():
info = [
('options.txt', partial(
sc.load_split_options,
split_keys=None,
), True),
('output_log.txt', scrape_output),
]
return infoThe function for loading the 'options.txt' files is already in soops. The
third item in the tuple, if present and True, denotes that the output contains
input parameters that were used for the parameterization. This allows getting
the parameterization in post-processing plugins, see below
the plot_win_rates() function.
The function to get useful information from 'output_log.txt' needs to be
provided:
def scrape_output(filename, rdata=None):
out = {}
with open(filename, 'r') as fd:
repeat = rdata['repeat']
for ii in range(4):
next(fd)
elapsed = []
win_rate = []
for ii in range(repeat):
line = next(fd).split()
elapsed.append(float(line[-1]))
line = next(fd).split()
win_rate.append(float(line[-1]))
out['elapsed'] = np.array(elapsed)
out['win_rate'] = np.array(win_rate)
return outThen we are ready to run soops-scoop:
$ soops-scoop -h
usage: soops-scoop [-h] [-s column[,column,...]]
[--filter filename[,filename,...]] [--no-plugins]
[--use-plugins name[,name,...] | --omit-plugins
name[,name,...]] [-p module] [--plugin-args dict-like]
[--results filename] [--no-csv] [-r] [--write] [--shell]
[--debug] [-o path]
scoop_mod directories [directories ...]
Scoop output files.
positional arguments:
scoop_mod the importable script/module with get_scoop_info()
directories results directories. On "Argument list too long"
system error, enclose the directories matching pattern
in "", it will be expanded using glob.glob().
options:
-h, --help show this help message and exit
-s column[,column,...], --sort column[,column,...]
column keys for sorting of DataFrame rows
--filter filename[,filename,...]
use only DataFrame rows with given files successfully
scooped
--no-plugins do not call post-processing plugins
--use-plugins name[,name,...]
use only the named plugins (no effect with --no-
plugins)
--omit-plugins name[,name,...]
omit the named plugins (no effect with --no-plugins)
-p module, --plugin-mod module
if given, the module that has get_plugin_info()
instead of scoop_mod
--plugin-args dict-like
optional arguments passed to plugins given as
plugin_name={key1=val1, key2=val2, ...}, ...
--results filename results file name [default: <output_dir>/results.h5]
--no-csv do not save results as CSV (use only HDF5)
-r, --reuse reuse previously scooped results file
--write write results files even when results were loaded
using --reuse option
--shell run ipython shell after all computations
--debug automatically start debugger when an exception is
raised
-o path, --output-dir path
output directory [default: .]
as follows:
$ soops-scoop soops/examples/monty_hall.py output/study/ -s rdir -o output/study --no-plugins --shell
<snip>
Python 3.7.3 | packaged by conda-forge | (default, Jul 1 2019, 21:52:21)
Type 'copyright', 'credits' or 'license' for more information
IPython 7.13.0 -- An enhanced Interactive Python. Type '?' for help.
In [1]: df.keys()
Out[1]:
Index(['rdir', 'rfiles', 'host', 'num', 'output_dir', 'plot_opts', 'repeat',
'seed', 'show', 'silent', 'switch', 'elapsed', 'win_rate', 'time'],
dtype='object')
In [2]: df.win_rate.head()
Out[2]:
0 [0.32, 0.4, 0.38, 0.27, 0.31, 0.39, 0.25, 0.33...
1 [0.64, 0.67, 0.68, 0.67, 0.73, 0.62, 0.66, 0.7...
2 [0.32, 0.32, 0.32, 0.32, 0.32, 0.32, 0.32, 0.3...
3 [0.68, 0.68, 0.68, 0.68, 0.68, 0.68, 0.68, 0.6...
4 [0.28, 0.28, 0.35, 0.32, 0.29, 0.33, 0.29, 0.3...
Name: win_rate, dtype: object
In [3]: df.iloc[0]
Out[3]:
rdir ~/projects/soops/output/study/000-7a6b546a625c...
rfiles [options.txt, output_log.txt]
host random
num 100
output_dir output/study/000-7a6b546a625c2d37569346a286f2b2b6
plot_opts {'linewidth': 3, 'alpha': 0.5}
repeat 10
seed NaN
show False
silent True
switch False
elapsed [0.0031552709988318384, 0.0032349379907827824,...
win_rate [0.32, 0.4, 0.38, 0.27, 0.31, 0.39, 0.25, 0.33...
time 2021-02-07 14:34:30.202971
Name: 0, dtype: object
The DataFrame with the all results is saved in output/study/results.h5
for reuse.
It is also possible to define simple plugins that act on the resulting
DataFrame. First, define a function that will register the plugins:
def get_plugin_info():
from soops.plugins import show_figures
info = [plot_win_rates, show_figures]
return infoThe show_figures() plugin is defined in soops. The plot_win_rates()
plugin allows plotting the all results combined:
def plot_win_rates(df, data=None, colormap_name='viridis'):
import soops.plot_selected as sps
df = df.copy()
df['seed'] = df['seed'].where(df['seed'].notnull(), -1)
uniques = sc.get_uniques(df, [key for key in data.multi_par_keys
if key not in ['output_dir']])
output('parameterization:')
for key, val in uniques.items():
output(key, val)
selected = sps.normalize_selected(uniques)
styles = {key : {} for key in selected.keys()}
styles['seed'] = {'alpha' : [0.9, 0.1]}
styles['num'] = {'color' : colormap_name}
styles['repeat'] = {'lw' : np.linspace(3, 2,
len(selected.get('repeat', [1])))}
styles['host'] = {'ls' : ['-', ':']}
styles['switch'] = {'marker' : ['x', 'o'], 'mfc' : 'None', 'ms' : 10}
styles = sps.setup_plot_styles(selected, styles)
fig, ax = plt.subplots(figsize=(8, 8))
sps.plot_selected(ax, df, 'win_rate', selected, {}, styles)
ax.set_xlabel('simulation number')
ax.set_ylabel('win rate')
fig.tight_layout()
fig.savefig(os.path.join(data.output_dir, 'win_rates.png'))
return dataThen, running:
soops-scoop soops/examples/monty_hall.py output/study/ -s rdir -o output/study -r
reuses the output/study/results.h5 file and plots the combined results:
It is possible to pass arguments to plugins using --plugin-args option, as
follows:
soops-scoop soops/examples/monty_hall.py output/study/ -s rdir -o output/study -r --plugin-args=plot_win_rates={colormap_name='plasma'}
- The get_run_info(), get_scoop_info() and get_plugin_info() info function can be in different modules.
- The script that is being parameterized need not be a Python module - any executable which can be run from a command line can be used.
'@defined'denotes that a value-less argument is present.'@undefined'denotes that a value-less argument is not present.'@arange([start,] stop[, step,], dtype=None)'denotes values obtained by callingnumpy.arange()with the given arguments.'@linspace(start, stop, num=50, endpoint=True, dtype=None, axis=0)'denotes values obtained by callingnumpy.linspace()with the given arguments.'@generate'denotes an argument whose values are generated, in connection with--generate-parsoption, see below.
Argument sequences can be generated using a function with the help of
--generate-pars option. For example, the same results as above can be
achieved by defining a function that generates --switch and --seed
arguments values:
def generate_seed_switch(args, gkeys, dconf, options):
"""
Parameters
----------
args : Struct
The arguments passed from the command line.
gkeys : list
The list of option keys to generate.
dconf : dict
The parsed parameters of the parametric study.
options : Namespace
The soops-run command line options.
"""
seeds, switches = zip(*product(args.seeds, args.switches))
gconf = {'--seed' : list(seeds), '--switch' : list(switches)}
return gconfand then calling soops-run as follows:
soops-run -r 1 -n 3 -c='--switch + --seed' -o output/study2 "python='python3', output_dir='output/study2/%s', --num=[100,1000,10000], --repeat=[10,20], --switch=@generate, --seed=@generate, --host=['random', 'first'], --silent=@defined, --no-show=@defined" --generate-pars="function=generate_seed_switch, seeds=['@undefined', 12345], switches=['@undefined', '@defined']" soops/examples/monty_hall.py
Notice the special @generate values of --switch and --seed, and the
use of --generate-pars: all key-value pairs, except the function name, are
passed into :func:generate_seed_switch() in the args dict-like
argument.
The combined results can again be plotted using:
soops-scoop soops/examples/monty_hall.py output/study2/0* -s rdir -o output/study2/
By using --compute-pars option it is possible to define arguments depending
on other arguments values in a more general way than with --contract.
A callable class needs to be provided with the following structure:
class ComputePars:
def __init__(self, args, par_seqs, key_order, options):
"""
Called prior to the parametric study to pre-compute reusable data.
"""
pass
def __call__(self, all_pars):
"""
Called for each parameter set of the study.
"""
out = {}
return outFor very large parametric studies, it might be impractical to view all_parameters.csv directly when searching a directory of a run with given parameters. The soops-find script can be used instead:
$ soops-find -h
usage: soops-find [-h] [-q pandas-query-expression]
[--engine {numexpr,python}] [--shell]
directories [directories ...]
Find parametric studies with parameters satisfying a given query.
Option-like parameters are transformed to valid Python attribute names removing
initial dashes and replacing other dashes by underscores. For example
'--output-dir' becomes 'output_dir'.
positional arguments:
directories one or more root directories with sub-directories
containing parametric study results
options:
-h, --help show this help message and exit
-q pandas-query-expression, --query pandas-query-expression
pandas query expression applied to collected
parameters
--engine {numexpr,python}
pandas query evaluation engine [default: numexpr]
--shell run ipython shell after all computations
Without options, it loads all parameter sets found in given directories into a DataFrame and launches the ipython shell:
$ soops-find output/study
find: 48 parameter sets stored in `apdf` DataFrame
find: column names:
Index(['finished', 'host', 'no_show', 'num', 'plot_opts', 'repeat', 'seed',
'silent', 'switch', 'python', 'output_dir', 'script_dir'],
dtype='object')
Python 3.8.5 (default, Sep 4 2020, 07:30:14)
Type 'copyright', 'credits' or 'license' for more information
IPython 7.21.0 -- An enhanced Interactive Python. Type '?' for help.
In [1]:
The --query option can be used to limit the search, for example:
$ soops-find output/study -q "num==1000 & repeat==20 & seed==12345"