Improvements¶

• Better checking of scripts and reporting of errors found during checking.
• Better handling of long file names/paths.
• Better error reporting when trying to write to a file that is open elsewhere (e.g., Excel).

Bug Fixes¶

• Correct initialization of some random variables (e.g., Bernoulli)
• Copied DATA_FIELDS to solutions such that non-standard columns are recognized rather than looking for default column names.
• Passed PREPROCESS code to Grph scripts so that preprocessed input observations/covariates are loaded correctly.
• Removed erroneous point at dose events in spaghetti plots.
• Corrected the way that string columns are processed in auto-generated code so that conditions on string fields are evaluated correctly.

Improvements¶

• Simplified the launch interface - e.g., to “popy run …” rather than “popy_run” - to reduce the number of additional scripts bundled with PoPy.
• Retired the “imconv” and “edit” tools, trusting the user to use their preferred tools for image conversion and text file editing.
• Retired the CPPODE solver in favour of the superior CPPLSODA solver.

Bug Fixes¶

• Added measures to make PoPy more tolerant of long path names that exceed the Windows limit of 260 characters.
• Fixed errors when a single epoch contains more than one occasion without a reset separating occasions.

Improvements¶

• Scatter plots implemented for rapid diagnostics of model fits.
• Tabular outputs now available (similar to Nonmem’s $TABLE directive).
• Conditional Weighted Residuals added as a diagnostics measure.
• Sequential variables can now be defined as an arbitrary list.
• Likelihoods exported on a per-row basis for detailed analysis.
• Improvements to efficiency and reliability when using parallel processing.
• Simplifications to the output data structures, with fewer subfolders and better consistency in file locations.
• Added option to specify parallelism by % of CPU load rather than an absolute number of processors.

Bug Fixes¶

• Random effects with a fixed variance of zero are now correctly treated as constants, and any associated fixed effects treated as unassociated fixed effects.
• Tabulated fixed effects were previously incorrect at times, but are now outputted correctly.
• Fixed an indexing bug when updating individuals’ parameter values.
• Clearer numbering of populations in multi-population simulations to avoid confusion by both users and computers.
• Fixed bug in normal distribution objects after reseeding with a known value.
• Corrected the way that flows of the form “d[X] -= x + y” are evaluated.
• Improved the way variable substitutions are handled in the PREDICTIONS block.
• Fixed bugs arising when trying to write files to folders that don’t (yet) exist.
• Corrected the way that analytic solvers deal with parameter sensitivity.

Improvements¶

• Faster processing for models with lots (tens of thousands) of observation rows that have a normally distributed error model.
• Parameterized compartment definitions so that many compartments can be defined in a single line.
• Better feedback when calculations result in a not-a-number (NaN) that derails the fitting process.
• Assumes that any uninitialized d[X] are initialized to zero, so you can add to flows without having to explicitly initialize them first.
• Faster evaluation of normally-distributed likelihoods.
• Extended sequential() distribution now allows arbitrary splitting of tree levels, e.g., c[AMT] = sequential([0, 100, 200]).
• Better code parsing flags up errors in the script earlier in the process.

Bug Fixes¶

• Covariance matrix (OMEGA) values that were defined as being constant were being modified by the JOE fitter.
• Models containing fixed effects that are constant would sometimes crash due to an indexing bug in the subset of all fixed effects.
• Incorrect calculation in ds/dm when using mixed models (with both a closed-form compartmental model and ODEs).
• IMP under parallel processing tried to recompute proposal distributions on every process rather than just on the master process.
• IMP used an approximate Lambda when it should have been using an exact one (and vice versa).
• Parsing the PREPROCESS block would sometimes alter its contents, causing faults downstream.
• Failure due to long filenames avoided by using symbolic links and raising warnings rather than errors under recoverable conditions.
• Covariances (OMEGAs) were being updated when they were defined as constants in some methods (e.g. JOE).
• Constant fixed effects were causing indexing problems in some fitting algorithms.
• Output folders created whenever they are needed (which wasn’t always the case).
• Sim failed when a solution file had a different DATA_FIELDS content from the corresponding Gen file.
• Fixed effects “associated” with an r[X] that is effectively constant (e.g., mean=0 and variance=0) were not being updated at all, rather than being updated using alternative (e.g., gradient-based) methods.
• Incorrect covariance matrix when the model included no random effects.
• Tabulated f[X] values were all the same.
• Tutsum failed to copy the compartment diagram from Gensum.

Improvements¶

• More stable estimation when using forward sensitivity equations.
• Improvements to VPC generation
• Some improvements in speed during fitting, especially for situations where there are lots of bolus doses per subject.
• Better memory management when fitting in parallel.
• Quotes in column headers of the data file are now permitted.

Bug Fixes¶

• Fixed various bugs with fitting in parallel

Improvements¶

• Reduced RAM usage when running PoPy in parallel using MPI. Previous version shared all subject data on all processors. New parallel code spreads subject data across all processors and avoids unnecessary duplication. Improves performance on large multi-core machines.
• Increased number of parameters that the ND fitting method and BOBYQA optimiser can process, previously restricted to 35, now increased to 200.
• Change the colour of the dosing lines in individual subject plots to yellow (formerly black), this enables more informative graphs for subjects who have a lot of separate doses.
• Improved the speed and accuracy of the r[X] optimisation using more analytic gradient results.
• Added ability to create prediction corrected (pred_corr) and prediction variance corrected (predvar_corr) VPC output via the “norm_method” parameter of a Vpc Script.
• Added ability to create stratified VPC output via the “split_field” parameter of a Vpc Script. A separate vpc is output for each unique value in a specified c[X] data variable.
• Made the MSim Script and Vpc Script code more RAM efficient.
• Made models with a large number of dosing events per subject run more efficiently.

Bug Fixes¶

• Fixed bug when passing string data into Cython functions.

Improvements¶

• Added new ND fitting method, the latest and most robust PoPy estimation approach.
• Improved fitting and lower ObjV using ND method.
• Upgraded base Python distribution of PoPy from version 3.5 to version 3.8. This allows us access to more up-to-date Python packages and functionality.
• Improved automatic compartment diagram generation with “in” and “out” nodes.
• Improved automatic checking of input scripts.

Bug Fixes¶

• Improvements in error catching and reporting when running PoPy in parallel using MPI.
• Fixed bug when using backslashes in DERIVATIVES section.

Improvements¶

• Added CPPLSODA C++ based implementation of LSODA ODE_SOLVER, which runs faster than SCIPY_ODEINT within PoPy and returns similar results.
• Improved PARALLEL performance of FOCE algorithm.
• Improved fitting of FOCE method to achieve lower ObjV in many models.
• Revised documentation.

Bug Fixes¶

• Bugs in gradient computation due to variable c[X] within individual data fixed.
• Fixed bugs in standard error computation, now more reliable.

Improvements¶

• Added implementation of an FOCE Fitting Method to complement the previous JOE Fitting Method. Note we now recommend running the JOE fitter, then the FOCE fitter in serial to get the best parameter fitting results in PoPy.
• New ~rectnorm() distribution for fitting to BLQ data, using JOE or FOCE methods without using the more computationally expensive LAPLACE objective value. See Generate BLQ observations and fit different error models.
• Also new BLQ related ~cennorm() distribution and ~truncnorm() distribution.
• New ~binomial() distribution likelihood.
• Improved the speed and quality of the noise f[X] parameter estimates for JOE fitting.
• New plots show the final f[X] population fit (same for all individuals, with r[X] zero) and the individual fit (i.e. empirical bayes estimates for r[X] optimised for each individual). Previously we plotted only the initial unoptimised f[X] fit and the final f[X] individual fits.
• The temporary functions generated by PoPy (based on the config file) are now written into a single module on disk. This single module compiles quicker using Cython, compared to having each temporary function in it’s own module.

Format Changes¶

• Added PREPROCESS section to multiple scripts, for example a Fit Script. This allows the user to easily filter and alter the data set, using simple Python code, before fitting a model, without changing the data file on disk.
• Added POSTPROCESS section for multiple scripts, for example a Gen Script. This allows the user to easily alter or filter a synthetically generated data set, before saving to disk.

Bug Fixes¶

• Fixed bug in PREDICTIONS section that did not handle multiple observations correctly in some instances, when some rows should be ignored due to Python ‘if’ statements.
• Fixed interpretation of ‘+inf’ and ‘-inf’ labels in EFFECTS.

Improvements¶

• Added ability to process scripts in PARALLEL. This speeds up computation time for models with a large number of individuals and takes advantage of multi-core computers.
• Created POSTPROCESS section to allow filtering of data create by a PoPy model. For example with a Gen Script
• Added ability to process multiple files in a directory with one command using ‘*’ syntax, see Running multiple scripts, Checking multiple scripts, Format multiple scripts, popy_edit_multi_scripts, Open multiple html files
• Speed improvements to core JOE algorithm, Standard Errors computation and plotting using Grph Script.
• MFit Script now outputs a summary of the fitting results for all populations to a single .csv file.
• Displaying the diagonal elements of f[X] variance matrices at each iteration. The off diagonal f[X] elements are only displayed at the end of fitting, due to space considerations.

Format Changes¶

• Simplified the structure of old ‘LEVEL_PARAMS’, replaced with EFFECTS which removes the redundant ‘split_field’ and ‘split_dict’ attributes.
• Renamed the ‘output_mode’ values in OUTPUT_SCRIPTS, using shorter names ‘gen_script_and_run’ -> ‘run’, ‘gen_script_only’ -> ‘create’, ‘no_output’ -> ‘none’.
• Renamed the ‘FINITE_DIFF’ method in COVARIANCE to ‘SANDWICH’.

Note scripts in v1.0.1 and v1.0.0 format can be automatically updated to v1.0.2 using the popy_format tool.

Bug Fixes¶

• Made float_format change format of all float values in output scripts and summary html outputs. Only applied to subset before.
• Fixed bug in iteration number when displaying ObjV over time Objective function at each iteration for simple PopPK example

Improvements¶

• The Standard Errors are now output in an easy to read format, instead of the whole hessian matrix.
• Changed the configuration file entry required to compute standard errors see COVARIANCE.
• Log files apart from the main log file are now stored in a ‘_log’ sub directory.
• The popy_run tool now asks the user if they want to proceed if the main log file already exist for a given script. Previously only the existence of the output folder was checked. The can still skip this check using the ‘-o’ option.

Bug Fixes¶

• Licence manager was giving a date error in North American time zones.
• Bug in configuration file parser for DERIVATIVES section when encountering round brackets (as opposed to expected curly brackets) in Dosing Functions.
• Documentation contents page indexing in PDF was incorrect.