Skip to content

Latest commit

 

History

History
118 lines (98 loc) · 13.7 KB

CHANGELOG.md

File metadata and controls

118 lines (98 loc) · 13.7 KB

Changelog

v0.1.8 (18.10.2021)

  • Since a recent update of numba, coefficients are computed incorrectly when parallelizing. This version disables the numba parallelization.

v0.1.7 (07.05.2021)

  • Consistency fixes of the example scripts to comply with the paper

v0.1.6 (23.04.2020)

  • Changed: Now under BSD 3-Clause License
  • Changed: When the data has more than one trial, we now require the user to choose which coefficient method to use (ts or sm) in mre.coefficients() and mre.full_analysis(). We showed that the resulting time scale that one finds can differ severely between the two methods. If unsure, compare results from both. We explain the difference in the paper and print some recommendation from the toolbox.
  • Changed: Due to above, method is now the second positional argument (this might break scripts that gave steps, dt, or dtunit as positonal arguments). Call mre.coefficients(data, 'ts') or as before via keyword mre.coefficients(data, method='ts')
  • Fixed: Typo that caused full_analysis() to crash when calling the consistency check.
  • Fixed: Workaround to prevent a memory leak when calling full_analysis() repeatedly. Always set showoverview=False when using full_analysis() in for loops. We now temporarily set matplotlib.rcParams['interactive'] = showoverview to avoid opening a new figure every time. This should make the panel and showoverview argument feel more consistent. The same workaround can be used in your custom scripts when using the OutputHandler (that also opens figures): Nest the loop inside a with matplotlib.rc_context(rc={'interactive': False}): (or adjust your rc parameters) to avoid figures.
  • Fixed: Various small bugs
  • New: coefficients has a new keyword argument knownmean to provide a known mean activity. If provdied, it will be used as the expectation value of the activity instead of calculating the mean as an approximation (both, in stationarymean and trialseparated method). This allows for custom estimates but, for instance, m>1 will not be detectable as the covariance cannot diverge when the same (time independent) expectation value is used for <a_{t}> and <a_{t+k}>. As one example, knownmean=0 restrains the fitted line (with slope r_k) to go through the origin (0,0). See Zierenberg et al., in press.

v0.1.5 (24.09.2019)

  • Changed: One-file spaghetti code was separated into submodules.
  • Fixed: stationarymean method for coefficients should work for m>1 (Note that this is a non-standard case. A detailed discussion will follow.)
  • New: Optional Numba dependency to parallelize and precompile the computation of the correlation coefficients. To install numby along with mrestimator, pip install -U mrestimator[numba]
  • New: Uploading pre-release versions to pypi. To switch run pip install -U --pre mrestimator[full] and to go back to stable pip install mrestimator==0.1.5.
  • New: Basic unit tests. python -m unittest mrestimator.test_suite

v0.1.4 (05.02.2019)

  • Changed: full_analysis() argument fitfunctions renamed to fitfuncs to be consistent with fit() and coefficients()
  • Changed: full_analysis() was rewritten, now only has three required arguments: data, dt and kmax, where kmax can be substituted by steps or tmax.
  • Changed: concerning the seed argument for various functions: all functions take either seed=None (no reseeding), seed='random' (reseeding to a random value - causing irreproducible resaults) or to a fixed value seed=int(yourseed). Per default, analysis functions - full_analysis(), fit() and coefficients() - produce same results by seeding to a fixed value each call. (only confidence intervals are affected by seeding) Per default, simulate_branching() and simulate_subsampling() seed to random.
  • Fixed: when calling branching process with subp and providing a seed, the subsampling no longer reseeds the rng device. (hence every call produces the same outcome, as expected)
  • Fixed: simulate_subsampling() now returns np arrays of correct dimensions
  • New: full_analysis() now shows a warning in the overview panel if consistency checks fail (so far only one).
  • New: Version number is printed into the overview panel of full_analysis() and into saved meta data

v0.1.3 (16.01.2019)

This is a bugfix version in preparation for the wrapper rewrite in 0.1.4.

  • Changed: If no steps are provided to coefficients(), the default maxstep is (for now) 1/10 of the trial length. (Was hard coded to 1500 before)
  • Changed: Default logs are less verbose to be clearer. The new function mre._enable_detailed_logging() enables fully detailed output to console and logfile. This also calls the two new switches, see next point. mre._enable_detailed_logging() also enables console display of runtime warnings that are usually only printed into the log.
  • Fixed: Crash due to logfiles. If the toolbox was used by more than one user on one machine, the logfile created in the temporary directory could not be overwritten by other users. We now try to set file permissions of the logfile and target directory to 777 if they are not subfolders of the user folder. Also, per default, each user gets their own directory /tmp/mre_username. Logfilehandler is now rotating and creates a maximum of 10 logfiles, 50mb each.
  • Fixed: full_analysis() no longer crashes with substracttrialaverage=True when the provided input is of integer type.
  • Fixed: fit() now returns a (mostly empty) FitResult when no fit converged instead of raising an exception. Helps with scripts that run multiple fits. The returned FitResult works with the OutputHandler in default settings and a note about the failed fit is added to the description and meta data.
  • Fixed: Calling coefficients() with custom steps e.g. steps=np.arange(0,100,5) is more robust and does not crash due to steps < 1. Incorrect entries are replaced.
  • Fixed: OutputHandler now has a deconstructor that closes the matplotlib figure if it was not provided as an arugment. Hence, opening many handlers (e.g. by reassigning a variable in a loop o = mre.OutputHandler()) does not keep the figure after reusing the variable. This used to cause a warning: More than 20 figures have been opened.
  • New: Enable logging of function arguments to console and logfile with mre._log_locals = True. Enable logging of stack traces to logfile via mre._log_trace = True. (Avoiding the console printout of stack traces on exceptions is not feasible at the moment). Per default, both options are False.

v0.1.2 (27.11.2018)

  • Changed: coefficients() with trialseparate method calculates rk differently (now strictly linear regression). This should enable m>1 estimates.
  • Changed: builtin fitfunctions now use absolute values of the amplitude of the exponential term
  • Changed: fits drawn above data (again), otherwise, they get hidden if data is noisy
  • Fixed: maximum steps k in coefficients cannot exceed half the trial length any longer. this could lead to strong fluctuations in r_k and fits would fail
  • Fixed: Crashes when providing custom fitfunctions to fit() due to unhandled request of default parameters
  • New: Rasterization of plots in the OutputHandler. Especially timeseries grow large quickly. Now, if OutputHandlers create their own figures/axes elements (ax-argument not given on construction) all elements with zorder<0 are rastered. Per default, add_ts() uses a zorder of -1 but add_coefficients() and add_fit() have values above one so they stay vectorized. Call ax.set_rasterization_zorder(0) on your custom ax axes element if you want the same effect on customized figures.
  • New: export as png option for OutputHandler.save_plot()

v0.1.1 (01.11.2018)

  • Changed: We reworked the structure of CoefficientResult to be more consistent. This is now a completely selfsimilar , where each child-entry has exactly the same structure as the parent. The new attributes trialcrs and bootstrapcrs replaced samples. Both are now lists containing again CoefficientResults, any (previously multidmensional) ndarrays are now 1d.
  • Changed: Per default, full_analysis() initialises the random number generator (used for bootstrapping) once per call and passes None to the seed arguments of lower functions so they do not reseed. We introduced the convention that seed=None tells that function to use the current state of the rng without seeding. (Added an auto option for seeding where needed)
  • Changed: All prints now use the logging module. Hopefully nothing broke :P.
  • Changed: Default log level to console is now 'INFO', and some logs that one could consider info go to 'DEBUG' to decrease the spam. Default loglevel to file is 'DEBUG' (logfile placed in the default temporary directory, which is also printed when loading the toolbox).
  • Changed: When providing no loglevel to full_analysis() it uses the currently set level of mre._logstreamhandler.
  • Fixed: When calling full_analysis() with one trial, a running average is shown instead of an empty plot.
  • New: Added quantiles (and standard errors) to fit results if bootstrapping. The new default option, numboot='auto' calculates 250 bootstrap samples for the exponential and exp+offset fit functions (which are decently fast) and skips error estimation for the builtin complex (and custom) fits.
  • New: Added function set_logfile(fname, loglevel='DEBUG') to change the path of the global logfile + level. This should allow running the toolbox in parallel, with a seperate logfile per thread and relatively silent/no console output when combining with mre._logstreamhandler.setLevel('ERROR') or calling full_analysis(..., loglevel='ERROR')
  • New: Undocumented way to change the respective loglevels is e.g. mre._logstreamhandler.setLevel('WARNING') for console and mre._logfilehandler.setLevel('DEBUG') for file
  • New: Added custom handler class that does not log 'None Type' Traces if log.exception() is called without a try statement

v0.1.0 (11.10.2018)

  • Changed: OutputHandlers set_xdata() now adjusts existing data and is slightly smarter. Now returns an array containing the indices where the x axis value is right for the provided data (wrt the existing context). See the example in the documentation.
  • Changed: When calling OutputHanlders add_coefficients() or add_ts(), the meta data and plot range will be extended using set_xdata. Trying to add duplicates only changes their style to the new provided values (without adding meta).
  • Changed: The parameters of simulate_branching() are different. activity is now a, m is no longer optional and it is possible to set a (time dependent) drive using h.
  • Fixed: Calling fit() with only one trial does not crash anymore due to missing uncertainties
  • Fixed: Calling fit() without specifying steps now uses the range used in coefficients().
  • New: added full_analysis(), the wrapper function to chain individual tasks together.
  • New: added simulate_subsampling()
  • New: Whenn adding time series to the OutputHandler in trial structure with more than one trial via add_ts(), they are drawn slightly transparent by default. Setting alpha overwrites this. add_ts does not use the new set_xdata() yet.
  • New: Versionbump so we have the last digit for bugfixes :)
  • New: Mr. Estimator came up with his logo.

v0.0.3 (19.09.2018)

  • Changed: Check for old numpy versions in fit()
  • Changed: Per default, fits are drawn solid (dashed) over the fitted (remaining) range
  • Fixed: Typos

(14.09.2018)

  • New: CoefficientResult constructor now has some default arguments. Still required: steps and coefficients. Also added the dt, dtunit attributes.
  • New: FitResult constructor now has some default arguments. Still required: 'tau, mre, fitfunc'. Also added the dt, dtunit, steps attributes.
  • New: fit() takes argument steps=(minstep, maxstep) to specify a custom fitrange. OutputHandler plots the fitted range opaque (excluded range has less alpha).
  • Changed: dt is no longer an argument for fit(). Setting dt (the step size) and its units dtunit is done via the equally named parameters of coefficients(). It is added to the CoefficientResult, so fit and the OutputHandler can rely on it.

(13.09.2018)

  • Renamed: module from mre to mrestimator, use import mrestimator as mre
  • Renamed: correlation_coefficients() to coefficients()
  • Renamed: correlation_fit() to fit()
  • Renamed: CorrelationFitResult to FitResult