The Python 3 wrapper: the adherer
module
Making the adherer module visible to Python (aka
installation)
The reference implementation is contained in single file
(adherer.py) included with the R
AdhereR package and whose location can be obtained using
the function getCallerWrapperLocation(full.path=TRUE) from
AdhereR (N.B. it is located in the directory where
the AdhereR package is installed on the system,
subdirectory wrappers/python3/adherer.py; for example, on
the example macos machine this is
/Library/Frameworks/R.framework/Versions/3.4/Resources/library/AdhereR/wrappers/python3/adherer.py).
In the future, as more wrappers are added, the argument
callig.platform will allow the selection of the desired
wrapper (now it is implicitely set to python3).
This file can be either:
- copied and placed in
Python’s “module search paths” (a list of directories comprising, in order, the directory contining the input script, the environment variablePYTHONPATH, and a default location; see here for details), in which cae it can be simply imported using the standardPythonsyntax (e.g.import adhererorimport adherer as ad), or - imported from its location by either:
- adding its directory to the
PYTHONPATHenvironment variable [the recommended solution], or - using various tricks described, for example, here.
- adding its directory to the
On the example macos machine, this can be achieved by
adding:
# Add AdhereR to PYTHONPATH
export PYTHONPATH=$PYTHONPATH:/Library/Frameworks/[...]/AdhereR/wrappers/python3to the .bash_profile file in the user’s home folder (if
this file does not exist, then it can be created using a text editor
such as nano; please note that the [...] are
for shortening the path and should be replaced by the actual path given
in full above). The process should be very similar on
Linux, while on MS Windows one should use the
system’s “Environment Variables” settings (for example, see here for details).
Please note that adherer needs pandas and
PIL, which can be installed, for example, using:
pip3 install pandas
pip3 install PillowNOTE: we will consistently use AdhereR to refer
to the R package, and adherer to refer to the
Python 3 module.
Importing the adherer module and initializations
Thus, the reference implementation is technically a
module called adherer that can be imported in
your code (we assume here the recommended solution, but see above for
other ways of doing it):
# Import adherer as ad:
import adherer as adWhen the adherer module is imported for the first time,
it runs the following initialization code:
- it tries to autodetect the location where
Ris installed on the system. More precisely, it looks forRscript(orRscript.exeon Windows) using several strategies, in order:which, followed by a set of standard locations onmacOS(/usr/bin/Rscript,/usr/local/bin/Rscript,/opt/local/bin/Rscriptand/Library/Frameworks/R.framework/Versions/Current/Resources/bin/Rscript) andLinux(/usr/bin/Rscript,/usr/local/bin/Rscript,/opt/local/bin/Rscriptand~/bin/Rscript) or a set of standard Windows Registry Keys onMS Windows(HKEY_CURRENT_USER\SOFTWARE\R-core\R,HKEY_CURRENT_USER\SOFTWARE\R-core\R32,HKEY_CURRENT_USER\SOFTWARE\R-core\R64,HKEY_LOCAL_MACHINE\SOFTWARE\R-core\R,HKEY_LOCAL_MACHINE\SOFTWARE\R-core\R32andHKEY_LOCAL_MACHINE\SOFTWARE\R-core\R64) which should containCurrent VersionandInstallPath(with the added complexity that 64 bits Windows hists both 32 and 64 bits regsistries). This procedure is inspired by the wayRStudiochecks forR: - if this process fails, a warning is thrown instructing the user to manually set the path using theset_rscript_path()function exposed by theadherermodule, and sets the internal variable_RSCRIPT_PATHtoNone(which insures that all future calls toAdhereRwill fail; - if the process succeeds, it checks if theAdhereRpackage is installed for the detectedRand has a correct version:- if this check fails, an appropriate warning is
thrown and
_RSCRIPT_PATHis set toNone; - if it succeeds, continue to step (2) below.
- if this check fails, an appropriate warning is
thrown and
- it tries to create a temporary directory (with
prefix
adherer-) with read and write access for the current user: - if this fails, it throws a warning instructing the user to manually set this to a directory with read & write access using theset_data_sharing_directory()function, and sets the internal variable_DATA_SHARING_DIRECTORYtoNone(ensuring that calls toAdhereRwill fail); - if it succeeds, the initialization code is considered to have finished successfully; also, on exit this temporary_DATA_SHARING_DIRECTORYis automatically deleted.
The class hierarchy
The adherer module tries to emulate the same philosophy
as the AhereR package, where various types of
CMAs (“continuous multiple-interval measures of medication
availability/gaps”) that implement different ways of computing adherence
encapsulate the data on which they were computed, the various relevant
parameter values used, as well as the results of the computation.
Here, we implemented this through the following class hierarchy (image generated with pyreverse, not showing the private attributes):
We will discuss now the main classes in turn.
The CallAdhereRError exception class
Errors in the adherer code are signalled by throwing
CallAdhereRError exceptions (the red class shown in the
bottom right corner).
The base class CMA0
All classes that implement ways to compute adherence
(CMAs) are derived from CMA0.
CMA0 does not in itself compute any sort of adherence, but
instead provides the infrastructure for storing the
data, parameter values and results (including errors), and for
interacting with AdhereR. Please note that in the
“higher” CMAs, the class constructor
__init()__ implicitely performs the actual computation of
the CMA and saves the results (for CMA0 there
are no such computations and __init()__ only saves the
parameters internally)!
- storage of data and parameter values:
CMA0allows the user to set various parameters through the constructor__init()__, parameters that are stored for later use, printing, saving, and for facilitating easy reconstruction of all types of computations. By main groups, these are (please, see the manual entry for?CMA0in theAdhereRpackage and the full protocol in Appendix I as well):
datasetstores the primary data (as aPandastable with various columns) containing the actual events; must be given;id_colname,event_date_colname,event_duration_colname,event_daily_dose_colnameandmedication_class_colname: these give the names of the columns in thedatasettable containing important information about the events (the first three are required, the last two are optional);carryover_within_obs_window,carryover_into_obs_window,carry_only_for_same_medication,consider_dosage_change,medication_change_means_new_treatment_episode,maximum_permissible_gap,maximum_permissible_gap_unit: optional parameters defining the types of carry-over, changes and treatment episode triggers;followup_window_start_type,followup_window_startfollowup_window_start_unit,followup_window_duration_typefollowup_window_duration,followup_window_duration_unit,observation_window_start_type,observation_window_start,observation_window_start_unitobservation_window_duration_type,observation_window_duration,observation_window_duration_unit`: optional parameters defining the follow-up and observation windows;sliding_window_start_type,sliding_window_start,sliding_window_start_unit,sliding_window_duration_type,sliding_window_duration,sliding_window_duration_unit,sliding_window_step_duration_type,sliding_window_step_duration,sliding_window_step_unit,sliding_window_no_steps: optional parameters defining the sliding windows;cma_to_apply: optional parameter specifying which “simple”CMAis to be used when computing sliding windos and treatment episodes;date_format: optional parameter describing the format of column dates indataset(defaults to month/day/year);event_interval_colname,gap_days_colname: optional parameters allowing the user to change the names of the columns where these computed data are stored in the resuling table;force_na_cma_for_failed_patients,keep_window_start_end_dates,remove_events_outside_followup_window,keep_event_interval_for_all_events: optional parameters governing the content of the resuling table;parallel_backend,parallel_threads: these optional parameters control the parallelism of the computations (if any); see PARALLEL PROCESSING for details;suppress_warnings: should all the internal warning be shown?save_event_info: should this “advanced” info be also made available?na_symbol_numeric,na_symbol_string,logical_symbol_true,logical_symbol_false,colnames_dot_symbol,colnames_start_dot: these optional parameters allowAdhereRto adapt to “non-R” conventions concerning the data format for missing values, logicals and column names;path_to_rscript,path_to_data_directory: these parameters allow the user to override the_RSCRIPT_PATHand_DATA_SHARING_DIRECTORYvariables;print_adherer_messages: should the important messages be printed to the user as well?
- storage of, and access to, the results:
get_dataset(): returns the internally savedPandastabledataset;get_cma(): returns the computedCMA(if any);get_event_info(): returns the computed event information (if any);get_treatment_episodes(): returns the computed treatment episodes information (if any);get_computation_results(): return the results of the last computation (if any); more precisely, adictionarycontaining the numericcodereturned byAdhereRand the stringmessageswritten byAdhereRduring the computation;
computing event interval and treatment episode info: this can be done by explicitelly calling the
compute_event_int_gaps()andcompute_treatment_episodes()functions;plotting:
static plotting: this is realized by the
plot()function that takes several plotting-specific parameters:patients_to_plot: should a subset of the patients present in thedatasetbe plotted (by default, all will be)?save_to,save_as,width,height,quality,dpi: where should the plot be saved, in what format, dimentions and quality?duration,align_all_patients,align_first_event_at_zero,show_period,period_in_days: duration to plots and alignment of patients;show_legend,legend_x,legend_y,legend_bkg_opacity: legend parameters;cex,cex_axis,cex_lab: the relative size of various text elements;show_cma,print_cma,plot_cma,plot_cma_as_histogram,cma_plot_ratio,cma_plot_col,cma_plot_border,cma_plot_bkg,cma_plot_text: should the cma be shown and how?unspecified_category_label: implicit label of unlabelled categories?lty_event,lwd_event,pch_start_event,pch_end_event,show_event_intervals,col_na,col_continuation,lty_continuation,lwd_continuation: visual aspects of events and continuations;highlight_followup_window,followup_window_col,highlight_observation_window,observation_window_col,observation_window_density,observation_window_angle,show_real_obs_window_start,real_obs_window_density,real_obs_window_angle: visual appearance of the follow-up, obervation and “real observation” windows (the latter forCMAs that djust it);bw_plot: produce a grayscel plot?
interactive plotting: the
plot_interactive()function launches a Shiny-powered interactive plot using the system’s WEB browser; the only parameterpatient_to_plotmay specify which patient to show initially, as all the relevant parameters can be interactively altered ar run-time;
printing: the
__repr__()function implements a very simple printing mechanism showing theCMAtype and a summary of thedataset;calling
AdhereR: the private function_call_adherer()is the real workhorse that manages all the interaction with theRAdhereRpackage as described above. This function can take many parameters covering all thatAdhereRcan do, but it is not intended to be directly called by the end-user but instead to be internally called by various exposed functions such asplot(),compute_event_int_gaps()and__init()__. Roughly, after some checks, it creates the files needed for communication, callsAdhereR, analyses any errors, warnings and messages that it might have generated, and packs the results in a manageable format.
To preserve the generality of the interaction with
AdhereR, all the CMA classes define a private
static member _adherer_function which is the name of the
corresponding S3 class as implemented in
AdhereR.
Class CMA1 and its daughter classes CMA2,
CMA3 and CMA4
CMA1 is derived from CMA0 by redefining the
__init__() constructor to (a) take only a subset of
arguments relevant for the CMAs 1–4 (see the
AdhereR help for them), and (b) to internally call
_call_adherer() with these parameters. It checks if the
result of _call_adherer() signals an error, in which case
ir throws a CallAdhereRError exception, otherwise packing
the code, messages, cma and (possibly) event information in the
corresponding member variables for later access.
Due to the generic mechanism implemented by
_adherer_function, CMA2, CMA3 and
CMA4 are derived directly from CMA1 but only
redefine _adherer_function appropriately.
Class CMA5 and its daughter classes CMA6,
CMA7, CMA8 and CMA9
The same story applies here, with CMA5 being derived
from CMA0 and redefining __init__(), with
CMA6–CMA9 only using the
_adherer_function mechanism. Compared with
CMA1, CMA5 defines new required arguments
related to medication type and dosage.
Classes CMAPerEpisode and
CMASlidingWindow
Just like CMA1 and CMA5, these two require
specific parameters and are thus derived directly from CMA0
(but, in contrast, they don’t have their own derived classes).
Examples of use
Below we show some examples of using the Python 3
reference wrapper. We are using IPython from the Spyder 3
environment; the In [n]: represents the input
prompt, the ...: the continuation of the input on
the following line(s), and Out[n]: the produced output.
Basic usage
Importing adherer and checking autodetection
Python 3.6.5 (v3.6.5:f59c0932b4, Mar 28 2018, 05:52:31)
Type "copyright", "credits" or "license" for more information.
IPython 6.3.0 -- An enhanced Interactive Python.
In [1]: # Import adherer as ad:
...: import adherer as ad
In [2]: # Show the _DATA_SHARING_DIRECTORY (should be set automatically to a temporary location):
...: ad._DATA_SHARING_DIRECTORY.name
Out[2]: '/var/folders/kx/bphryt7j5tz1n_fcjk5809940000gn/T/adherer-05hdq6el'
In [3]: # Show the _RSCRIPT_PATH (should be dectedt automatically):
...: ad._RSCRIPT_PATH
Out[3]: '/usr/local/bin/Rscript'Everything seems fine!
Export the test dataset from R and import it in
Python
Let’s export the sample dataset med.events included in
AdhereR as a TAB-separated CSV file in a location for use
here (please note that this must be done from an R console,
such as from RStudio, and not from
Python!):
R version 3.4.3 (2017-11-30) -- "Kite-Eating Tree"
Copyright (C) 2017 The R Foundation for Statistical Computing
Platform: x86_64-apple-darwin15.6.0 (64-bit)
R is free software and comes with ABSOLUTELY NO WARRANTY.
You are welcome to redistribute it under certain conditions.
Type 'license()' or 'licence()' for distribution details.
Natural language support but running in an English locale
R is a collaborative project with many contributors.
Type 'contributors()' for more information and
'citation()' on how to cite R or R packages in publications.
Type 'demo()' for some demos, 'help()' for on-line help, or
'help.start()' for an HTML browser interface to help.
Type 'q()' to quit R.
> library(AdhereR) # load the AdhereR package
> head(med.events) # see how the included med.events dataset looks like
PATIENT_ID DATE PERDAY CATEGORY DURATION
286 1 04/26/2033 4 medA 50
287 1 07/04/2033 4 medB 30
288 1 08/03/2033 4 medB 30
289 1 08/17/2033 4 medB 30
291 1 10/13/2033 4 medB 30
290 1 10/16/2033 4 medB 30
> write.table(med.events, file="~/Temp/med-events.csv", quote=FALSE, sep="\t", row.names=FALSE, col.names=TRUE) # save med.events as TAB-separated CSV file in a location (here, in the Temp folder)
> Now, back to Python:
In [4]: # Import Pandas as pd:
...: import pandas as pd
In [5]: # Load the test dataset
...: df = pd.read_csv('~/Temp/med-events.csv', sep='\t', header=0)
In [6]: # Let's look at first 6 rows (it should match the R output above except for the row names):
...: df.head(6)
Out[6]:
PATIENT_ID DATE PERDAY CATEGORY DURATION
0 1 04/26/2033 4 medA 50
1 1 07/04/2033 4 medB 30
2 1 08/03/2033 4 medB 30
3 1 08/17/2033 4 medB 30
4 1 10/13/2033 4 medB 30
5 1 10/16/2033 4 medB 30All good so far, the data was imported successfully as a
Pandas table.
Compute and plot test CMA
Now let’s compute CMA8 on these data in
Python:
In [7]: # Compute CMA8 as a test:
...: cma8 = ad.CMA8(df,
...: id_colname='PATIENT_ID',
...: event_date_colname='DATE',
...: event_duration_colname='DURATION',
...: event_daily_dose_colname='PERDAY',
...: medication_class_colname='CATEGORY')
...:
Adherer returned code 0 and said:
AdhereR 0.2.0 on R 3.4.3 started at 2018-06-04 22:27:10:
OK: the results were exported successfully (but there might be warnings and messages above worth paying attention to)!We can see that things went pretty well, as no exceptions were being
thrown and the message starts with a reassuring
Adherer returned code 0, followed by precisely what
AdhereR said:
AdhereR 0.2.0 on R 3.4.3 started at 2018-06-04 22:27:10:: first, self-identification (its own version andR’s version), followed by the date and time the processing was initated;OK: the results were exported successfully (but there might be warnings and messages above worth paying attention to)!, which means that basically all seems allright but that there might still be some messages or warning displayed above that could be informative or point to subtler issues.
Let’s see how these results look like:
In [8]: # Summary of cma8:
...: cma8
Out[8]: CMA object of type CMA8 (on 1080 rows).
In [9]: # The return value and messages:
...: cma8.get_computation_results()
Out[9]:
{'code': 0,
'messages': ['AdhereR 0.2.0 on R 3.4.3 started at 2018-06-04 22:27:10:\n',
'OK: the results were exported successfully (but there might be warnings and messages above worth paying attention to)!\n']}
In [10]: # The CMA (the first 6 rows out of all 100):
...: cma8.get_cma().head(6)
Out[10]:
PATIENT_ID CMA
0 1 0.947945
1 2 0.616438
2 3 0.994521
3 4 0.379452
4 5 0.369863
5 6 0.406849
In [11]: # Plot it (statically):
...: cma8.plot(patients_to_plot=['1', '2', '3'],
...: align_all_patients=True,
...: period_in_days=30,
...: cex=0.5)
...:
Adherer returned code 0 and said:
AdhereR 0.2.0 on R 3.4.3 started at 2018-06-05 13:22:36:
OK: the results were exported successfully (but there might be warnings and messages above worth paying attention to)!
Out[11]: The output produced in this case (Out[11]) consists of
the actual image plotted in the IPython console (thanks to
the PIL/Pillow package) and reproduced below:
Now, we turn again to R:
> # Compute the same CMA8 in R:
> cma8 <- CMA8(data=med.events,
+ ID.colname="PATIENT_ID",
+ event.date.colname="DATE",
+ event.duration.colname="DURATION",
+ medication.class.colname="CATEGORY",
+ event.daily.dose.colname="PERDAY"
+ )
>
> # The computed CMA:
> head(getCMA(cma8))
PATIENT_ID CMA
1 1 0.9479452
2 2 0.6164384
3 3 0.9945205
4 4 0.3794521
5 5 0.3698630
6 6 0.4068493
>
> # Plot the cma:
> plot(cma8,
+ patients.to.plot=c("1", "2", "3"),
+ align.all.patients=TRUE,
+ period.in.days=30,
+ cex=0.5)
>Again, the output is an image (here, shown in the “Plots” panel in
RStudio):
It can be seen that, except for the slightly different dimensions (and x/y ratio and quality) due to the actual plotting and exporting, the images show identical patterns.
Interactive plotting
We will initate now an interactive plotting session from
Python:
In [12]: cma8.plot_interactive()The output is represented by an interactive session in the default
browser; below is a screenshot of this session in Firefox:
The interactive session ends by pressing the “Exit” button in the
browser (and optinally also closing the browser tab/window), at which
point the usual text output is provided to Python and a
True value signalling success is returned:
Adherer returned code 0 and said:
AdhereR 0.2.0 on R 3.4.3 started at 2018-06-05 18:01:28:
OK: the results were exported successfully (but there might be warnings and messages above worth paying attention to)!
Out[12]: TruePlease note that it does not matter how the interactive session is
started, as it only needs access to the base CMA0 object
and, more precisely, the raw dataset; all relevant parameters, including
the CMA type can be changed interactively (this is why the CMA shown in
the screenshot is CMA1 even if the function
cma8.plot_interactive() was initiated from a
CMA8 object).
From a Jupyter Notebook
AdhereR is very easy to use from within a Jupyter Notebook (the only tricky bit
being making sure that adherer module is visible to the
Python kernel, but this is explained in detail in section
Making
the adherer module visible to Python (aka
installation)).
A full example is provided in the
jupyter_notebook_python3_wrapper.ipynb file accompanying
the package in the same directory as the adherer module,
but the code is reproduced below for convenience:
# import the adherer python module (see above about how to make it findable)
import adherer
# load the test dataset:
import pandas
df = pandas.read_csv('./test-dataset.csv', sep='\t', header=0)
# Change the column names:
df.rename(columns={'ID': 'patientID',
'DATE': 'prescriptionDate',
'PERDAY': 'quantityPerDay',
'CLASS': 'medicationType',
'DURATION': 'prescriptionDuration'},
inplace=True)
# check the file was read correctly
type(df)
# HTML printing of data frames
import rpy2.ipython.html
rpy2.ipython.html.init_printing()
# print it
df
# create a CMA7 object
cma7 = adherer.CMA7(df,
id_colname='patientID',
event_date_colname='prescriptionDate',
event_duration_colname='prescriptionDuration',
event_daily_dose_colname='quantityPerDay',
medication_class_colname='medicationType',
followup_window_start=230,
followup_window_duration=705,
observation_window_start=41,
observation_window_duration=100,
date_format="%m/%d/%Y",
suppress_warnings=True)
# print it for checking
cma7
# show the plot directly in the notebook
xWhile this is very easy and transparent, some people might prefer to
use the more generic mechanism provided by rpy2, which,
in principle, allows the transparent call of R code from
Python. In fact, AdhereR seems to play very
nicely with rpy2, even within a Jupyter Notebook, as shown by the code
below (the full notebook is available in the
jupyter_notebook_python3_py2.ipynb file accompanying the
package in the same directory as the adherer module):
# import ryp2 (please note that you might need to install it as per https://rpy2.github.io/doc.html)
import rpy2
# check that all is ok
rpy2.__path__
# import R's "AdhereR" package
from rpy2.robjects.packages import importr
adherer = importr('AdhereR')
# access the internal R session
import rpy2.robjects as robjects
# access the med.events dataset
med_events = robjects.r['med.events']
# check its type
type(med_events)
# HTML printing of data frames
import rpy2.ipython.html
rpy2.ipython.html.init_printing()
# print it
med_events
# make some AdhereR functions available to Python
CMA7 = robjects.r['CMA7']
getCMA = robjects.r['getCMA']
# create a CMA7 object
cma7 = CMA7(med_events,
ID_colname="PATIENT_ID",
event_date_colname="DATE",
event_duration_colname="DURATION",
event_daily_dose_colname="PERDAY",
medication_class_colname="CATEGORY",
followup_window_start=230,
followup_window_duration=705,
observation_window_start=41,
observation_window_duration=100,
date_format="%m/%d/%Y")
# print it for checking
cma7
# print the estimated CMAs
getCMA(cma7)
# plot it
# this is some clunky code involving do.call and named lists because
# plot has lots of arguments with . that cannot be autmatically handeled by rpy2
# the idea is to use a TaggedList that associated values and argument names
import rpy2.rlike.container as rlc # for TaggedList
rcall = robjects.r['do.call'] # do.call()
grdevices = importr('grDevices') # R graphics device
# the actual plotting
grdevices.jpeg(file="./cma7plot.jpg", width=512, height=512)
rcall("plot",
rlc.TaggedList([cma7,
robjects.IntVector([1,2,3]),
False,
True],
tags=('cma',
'patients.to.plot',
'show.legend',
'align.all.patients')))
grdevices.dev_off()Arguably, the code is clunkier and, at least in this approach, needs an extra step for showing the plot:
Include the CMA7 plot using the Markdown syntax (there are several alternatives: https://stackoverflow.com/questions/32370281/how-to-embed-image-or-picture-in-jupyter-notebook-either-from-a-local-machine-o):
Parallel processing (locally and on different machines)
AdhereR uses R’s parallel processing
capacities to split expensive computations and distribute them across
multiple CPUs/cores in a single computer or even across a network of
computers. As an example, we will compute here CMA1 across
sliding windows on the whole dataset, first in R and then
in Python 3.
Single thread on the local machine
The default mode of computation uses just a single CPU/core on the local machine.
R
> # Sliding windows with CMA1 (single thread, locally):
> cma1w.1l <- CMA_sliding_window(CMA.to.apply="CMA1",
+ data=med.events,
+ ID.colname='PATIENT_ID',
+ event.date.colname='DATE',
+ event.duration.colname='DURATION',
+ event.daily.dose.colname='PERDAY',
+ medication.class.colname='CATEGORY',
+ sliding.window.duration=30,
+ sliding.window.step.duration=30,
+ parallel.backend="none",
+ parallel.threads=1)
> head(getCMA(cma1w.1l))
PATIENT_ID window.ID window.start window.end CMA
1 1 1 2033-04-26 2033-05-26 NA
2 1 2 2033-05-26 2033-06-25 NA
3 1 3 2033-06-25 2033-07-25 NA
4 1 4 2033-07-25 2033-08-24 2.142857
5 1 5 2033-08-24 2033-09-23 NA
6 1 6 2033-09-23 2033-10-23 10.000000
>Python 3
In [13]: # Sliding windows with CMA1 (single thread, locally):
...: cma1w_1l = ad.CMASlidingWindow(dataset=df,
...: cma_to_apply="CMA1",
...: id_colname='PATIENT_ID',
...: event_date_colname='DATE',
...: event_duration_colname='DURATION',
...: event_daily_dose_colname='PERDAY',
...: medication_class_colname='CATEGORY',
...: sliding_window_duration=30,
...: sliding_window_step_duration=30,
...: parallel_backend="none",
...: parallel_threads=1)
...:
Adherer returned code 0 and said:
AdhereR 0.2.0 on R 3.4.3 started at 2018-06-06 15:19:07:
OK: the results were exported successfully (but there might be warnings and messages above worth paying attention to)!
In [14]: cma1w_1l.get_cma().head(6)
Out[14]:
PATIENT_ID window.ID window.start window.end CMA
0 1 1 04/26/2033 05/26/2033 NaN
1 1 2 05/26/2033 06/25/2033 NaN
2 1 3 06/25/2033 07/25/2033 NaN
3 1 4 07/25/2033 08/24/2033 2.142857
4 1 5 08/24/2033 09/23/2033 NaN
5 1 6 09/23/2033 10/23/2033 10.000000Multi-threaded on the local machine
If the local machine has multiple CPUs/cores (even with
hyperthreading), it might make sense to use them for lengthy
computations. AdhereR can use several backends (as provided
by the parallel package in R), of which the
most used are “multicore” (preffered on Linux and
macOS but currently not available on Windows)
and “SNOW” (on all three OS’s). AdhereR is smart enough to
use “SNOW” on Windows even if “multicore” was
requested.
R
> # Sliding windows with CMA1 (two threads, multicore, locally):
> cma1w.2ml <- CMA_sliding_window(CMA.to.apply="CMA1",
+ data=med.events,
+ ID.colname='PATIENT_ID',
+ event.date.colname='DATE',
+ event.duration.colname='DURATION',
+ event.daily.dose.colname='PERDAY',
+ medication.class.colname='CATEGORY',
+ sliding.window.duration=30,
+ sliding.window.step.duration=30,
+ parallel.backend="multicore", # <--- multicore
+ parallel.threads=2)
> head(getCMA(cma1w.2ml))
PATIENT_ID window.ID window.start window.end CMA
1 1 1 2033-04-26 2033-05-26 NA
2 1 2 2033-05-26 2033-06-25 NA
3 1 3 2033-06-25 2033-07-25 NA
4 1 4 2033-07-25 2033-08-24 2.142857
5 1 5 2033-08-24 2033-09-23 NA
6 1 6 2033-09-23 2033-10-23 10.000000
>
> cma1w.2sl <- CMA_sliding_window(CMA.to.apply="CMA1",
+ data=med.events,
+ ID.colname='PATIENT_ID',
+ event.date.colname='DATE',
+ event.duration.colname='DURATION',
+ event.daily.dose.colname='PERDAY',
+ medication.class.colname='CATEGORY',
+ sliding.window.duration=30,
+ sliding.window.step.duration=30,
+ parallel.backend="snow", # <--- SNOW
+ parallel.threads=2)
> head(getCMA(cma1w.2sl))
PATIENT_ID window.ID window.start window.end CMA
1 1 1 2033-04-26 2033-05-26 NA
2 1 2 2033-05-26 2033-06-25 NA
3 1 3 2033-06-25 2033-07-25 NA
4 1 4 2033-07-25 2033-08-24 2.142857
5 1 5 2033-08-24 2033-09-23 NA
6 1 6 2033-09-23 2033-10-23 10.000000
> Python 3
In [15]: # Sliding windows with CMA1 (two threads, multicore, locally):
...: cma1w_2ml = ad.CMASlidingWindow(dataset=df,
...: cma_to_apply="CMA1",
...: id_colname='PATIENT_ID',
...: event_date_colname='DATE',
...: event_duration_colname='DURATION',
...: event_daily_dose_colname='PERDAY',
...: medication_class_colname='CATEGORY',
...: sliding_window_duration=30,
...: sliding_window_step_duration=30,
...: parallel_backend="multicore", # <--- multicore
...: parallel_threads=2)
...:
Adherer returned code 0 and said:
AdhereR 0.2.0 on R 3.4.3 started at 2018-06-07 11:44:49:
OK: the results were exported successfully (but there might be warnings and messages above worth paying attention to)!
In [16]: cma1w_2ml.get_cma().head(6)
Out[16]:
PATIENT_ID window.ID window.start window.end CMA
0 1 1 04/26/2033 05/26/2033 NaN
1 1 2 05/26/2033 06/25/2033 NaN
2 1 3 06/25/2033 07/25/2033 NaN
3 1 4 07/25/2033 08/24/2033 2.142857
4 1 5 08/24/2033 09/23/2033 NaN
5 1 6 09/23/2033 10/23/2033 10.000000
In [17]: # Sliding windows with CMA1 (two threads, snow, locally):
...: cma1w_2sl = ad.CMASlidingWindow(dataset=df,
...: cma_to_apply="CMA1",
...: id_colname='PATIENT_ID',
...: event_date_colname='DATE',
...: event_duration_colname='DURATION',
...: event_daily_dose_colname='PERDAY',
...: medication_class_colname='CATEGORY',
...: sliding_window_duration=30,
...: sliding_window_step_duration=30,
...: parallel_backend="snow", # <--- SNOW
...: parallel_threads=2)
...:
Adherer returned code 0 and said:
AdhereR 0.2.0 on R 3.4.3 started at 2018-06-07 11:44:49:
OK: the results were exported successfully (but there might be warnings and messages above worth paying attention to)!
In [18]: cma1w_2sl.get_cma().head(6)
Out[18]:
PATIENT_ID window.ID window.start window.end CMA
0 1 1 04/26/2033 05/26/2033 NaN
1 1 2 05/26/2033 06/25/2033 NaN
2 1 3 06/25/2033 07/25/2033 NaN
3 1 4 07/25/2033 08/24/2033 2.142857
4 1 5 08/24/2033 09/23/2033 NaN
5 1 6 09/23/2033 10/23/2033 10.000000Parallel on remote machines over a network
Sometimes it is better to use one or more powerful machines over a
network to do very expensive computations, usually, a Linux
cluster from a Windows/macos laptop.
AdhereR leverages the power of R’s snow
package (as exposed through the parallel package) to
distribute workloads across a network of computing nodes. There are
several types of “Simple Network of Workstations” (snow),
described in the package’s manual. For example, one may use an already
existing MPI (Message
Passing Interface) cluster, but an even simpler setup (and the one
that we will illustrate here) involves a collection of machines
running Linux and connected to a network (local or even
over the Internet).
The machines are called workhorse1 and
workhorse2, have differen hardware configurations (both
sport quad-core i7 CPUs of different generations with 16Gb RAM) but run
the same version of Ubuntu 16.04 and R 3.4.2
(not a requirement, as the architecture can be seriously heterogeneous,
combining different OS’s and versions of R). These two
machines are connected to the same WiFi router (but they could be on
different networks or even across the Internet). The “master” is the
same macOS laptop used before, connected to the same WiFi
router (not a requirement).
As pre-requisites, the worker machines should allow SSH
access (for easiness, we use here passwordless SSH access from
the “master”; see for example here for this setup)
and should have the snow package installed in
R. Let’s assume that the username allowing ssh
into the workers is user, so that
laptop:~> ssh user@workhorse1works with no password needed. With these, we can distribute our processing to the two “workers” (two parallel threads for each, totalling 4 parallel threads):
R
> # Sliding windows with CMA1 (two remote machines with two threads each):
> # First, we need to specify the workers
> # This is a list of lists!
> # rep(,2) means that we generate two threads on each worker
> workers <- c(rep(list(list(host="workhorse1", # hostname (make sure this works from the "master", otherwise use the IP-address)
+ user="user", # the username that can ssh into the worker (passwordless highly recommended)
+ rscript="/usr/local/bin/Rscript", # the location of Rscript on the worker
+ snowlib="/usr/local/lib64/R/library/")), # the location of the snow package on the worker
+ 2),
+ rep(list(list(host="workhorse2",
+ user="user",
+ rscript="/usr/local/bin/Rscript",
+ snowlib="/usr/local/lib64/R/library/")),
+ 2));
>
> cma1w.2sw <- CMA_sliding_window(CMA="CMA1",
+ data=med.events,
+ ID.colname="PATIENT_ID",
+ event.date.colname="DATE",
+ event.duration.colname="DURATION",
+ event.daily.dose.colname="PERDAY",
+ medication.class.colname="CATEGORY",
+ carry.only.for.same.medication=FALSE,
+ consider.dosage.change=FALSE,
+ sliding.window.duration=30,
+ sliding.window.step.duration=30,
+ parallel.backend="snow",
+ parallel.threads=workers)
> head(getCMA(cma1w.2sw))
PATIENT_ID window.ID window.start window.end CMA
1 1 1 2033-04-26 2033-05-26 NA
2 1 2 2033-05-26 2033-06-25 NA
3 1 3 2033-06-25 2033-07-25 NA
4 1 4 2033-07-25 2033-08-24 2.142857
5 1 5 2033-08-24 2033-09-23 NA
6 1 6 2033-09-23 2033-10-23 10.000000
> Python 3
A quick for Python is that due to the communication
protocol between the wrapper and AdhereR, the specification
of the computer cluster must be a one-line string literally contaning
the R code defining it, string that will be verbatim parsed
and interpreted by AdhereR:
In [19]: # Sliding windows with CMA1 (two remote machines with two threads each):
...: # The workers are defined as *literal R code* this is verbatim sent to AdhereR for parsing and interpretation
...: # Please note, however, that this string should not contain line breaks (i.e., it should be a one-liner):
...: workers = 'c(rep(list(list(host="workhorse1", user="user", rscript="/usr/local/bin/Rscript", snowlib="/usr/local/lib64/R/library/")), 2), rep(list(list(host="workhorse2", user="user", rscript="/usr/local/bin/Rscript", snowlib="/usr/local/lib64/R/library/")), 2))'
In [20]: cma1w_2sw = ad.CMASlidingWindow(dataset=df,
...: cma_to_apply="CMA1",
...: id_colname='PATIENT_ID',
...: event_date_colname='DATE',
...: event_duration_colname='DURATION',
...: event_daily_dose_colname='PERDAY',
...: medication_class_colname='CATEGORY',
...: sliding_window_duration=30,
...: sliding_window_step_duration=30,
...: parallel_backend="snow",
...: parallel_threads=workers)
Adherer returned code 0 and said:
AdhereR 0.2.0 on R 3.4.3 started at 2018-06-07 13:22:21:
OK: the results were exported successfully (but there might be warnings and messages above worth paying attention to)!
In [21]: cma1w_2sw.get_cma().head(6)
Out[21]:
PATIENT_ID window.ID window.start window.end CMA
0 1 1 04/26/2033 05/26/2033 NaN
1 1 2 05/26/2033 06/25/2033 NaN
2 1 3 06/25/2033 07/25/2033 NaN
3 1 4 07/25/2033 08/24/2033 2.142857
4 1 5 08/24/2033 09/23/2033 NaN
5 1 6 09/23/2033 10/23/2033 10.000000Some caveats for over-the-network distributed computation
While this is a very good way to transparently distribute processing to more powerful nodes over a network, there are several (potential) issues one must be aware of:
it may be very hard to debug failures: failures of this might result from network issues, firewals blocking connections, incorrect
SSHsetup on the “workers” or errors in accesing the “workers” with the given user accounts; see, for examples, discussion here and here in case you need to solve such problems;latency over the network: starting the “workers” and especially transmitting the data to the “workers” and the results back to the “master” may take a non-negligible time, especially on slow networks (such as the Internet) and for large datasets; therefore, the best scenarios would involve relatively large computations (but not too large; see below) distributed to several nodes over a fast network;
you need to wait for the results: this process assumes that the “master” will wait for the “workers” to finish and return their results; thus, putting the “master” to sleep, shutting it down or disconnecting it from the network will probably result in not being able to collect the resuls back. If one needs very long computations (say 3+ hours), offline mobility or the network is unreliable, we would suggest setting up a separate compute process (that may itself parallelise computations) on the remote machines using, for example,
screen,nohupor a more specialised cluster management platform such as Son of a Grid Engine (SGE).
COMMENTS
Everything on a line following
///or#is considered a comment and ignored (except when included within quotes" "or' ').