https://github.com/gweindel/hmp - v1.0.0-b1

Changelog

[1.0.0-b1]

Added

Refactored transform and model structure to fit/transform framework (qubixes)
Major documentation overhaul and improved installation instructions (GWeindel)
Created distribution and template classes (GWeindel)
Save submodels in cumulative for later analysis (qubixes)
Creating io module (GWeindel)

Changed

Restructured preprocessing (Maarten Schermer)
Rename levels to grouping (GWeindel)
Splitted the readmnedata into subfunctions (GWeindel)
Linting and code style improvements (qubixes, Jelle Treep, GWeindel)
Documentation improvements (qubixes, GWeindel)
Rendering website (GWeindel)

Fixed

Fixed tests and removed leftover (GWeindel)
Fixed few minor errors (GWeindel)
Fixed left-over from naming (GWeindel)
Fix visu import (GWeindel)
Small fix for timestodisplay (GWeindel)
Cleaned trial_data of redundant infos and adapted downstream (GWeindel)
Memory management improvement (GWeindel)

Contributors (2024-12 to 2025-07)

GWeindel
qubixes
Maarten Schermer
Jelle Treep

- Python
Published by GWeindel 11 months ago

https://github.com/gweindel/hmp - v0.5.0

What's Changed

Sdist fix by @qubixes in https://github.com/GWeindel/hmp/pull/158
Fixing save and load by @GWeindel in https://github.com/GWeindel/hmp/pull/159
Removing deprecated by @GWeindel in https://github.com/GWeindel/hmp/pull/160
Further deprec removal by @GWeindel in https://github.com/GWeindel/hmp/pull/161
Improving coverage by @GWeindel in https://github.com/GWeindel/hmp/pull/163

Full Changelog: https://github.com/GWeindel/hmp/compare/v0.4.2...v0.5.0

- Python
Published by GWeindel over 1 year ago

https://github.com/gweindel/hmp - v0.4.1

What's Changed

Change in simulation code, the simulated activity is now independent of the sampling frequency
plottopotimecourse() can now plot subsets of data (e.g. epoch_data.sel(epochs=[0,1]))
p_values in fit() are now optional, they're too slow to compute and only really useful in simulatino corner cases
Added rejection threshold to utlis.read_mne_data, it now removes epochs for which the electrode activity exceeds threshold but only from stim to RT (thus no effect if artefact is after RT)
Removed stacked epoch data from HMP model when not necessary, saves RAM by @rickdott in https://github.com/GWeindel/hmp/pull/148
Included Testing notebook & Started writing the testing workflow for every push on main. by @nmararo in https://github.com/GWeindel/hmp/pull/150

Full Changelog: https://github.com/GWeindel/hmp/compare/v0.4.0...v0.4.1

- Python
Published by GWeindel over 1 year ago

https://github.com/gweindel/hmp - v0.4.0

Major (breaking) changes:

HMP now consistently models (or simulate) the peak of events instead of the onset, this has a consequence on the estimation and the visualization of the results
The fit function has now a new parameters (pval) that now performs a statistical tests for the first event, if the event has a pvalue (based on zscores of distribution from 100 not maximized starting points) below pval accept, otherwise reject. Default is .1 to not change previous behavior, decrease if suspicion of false positive.
The new utils.centered_activityreplace utils.event_times and now does the heavy lifting for plotting channel activity by centering around an event (e.g. event=0 for stimulus, event =3 for event 3,...) with multiple options
Sped up the fit function by not evaluating every sample, and in the case of a rejected proposition, start from the latest position in time of that proposition.
The estimation at the end of the use of the fit function now turns off the location parameter (censoring in the likelihood for samples close to previous event) to get estimates for all samples
Default location (censoring in the likelihood for samples close to previous event) has now a default value = event_width_samples. A new method is being developed to completely get rid of the location (using correlation between events). It is for now experimental but can be used by setting location = 0 and location_corr_threshold to the desired maximum correlation between events
Implemented MCCA (still in the experimentation phase)

Minor:

Several options have been added to simulations, it is now possible to simulate non-sequential events, oscillatory events, events only present on some trials ([NB CHANGE] added possibility to simulate events with different relations (sequence), with different probability of occurence (proportions), with different durations (event_duration))
Simulations now better classify true vs test events
Simulations are now more efficient in terms of signal generation speed and memory usage
Removed default linear detrending in utils.read_mne_data as it is not needed and rather a bad default choice
Added the possibility to not cut at the RT for readmnedata() with the parameter ignore_rt
Updated computation of maximum possible number of events
Updated termination rule of the fit functions
Added option to extract raw likelihood for all time points and trials
Sped up centering operation in transform_data

Bug fix:

transform data was modifying the data inplace, mainly detrending all epochs, this had consequence on all the analyses using the data before transform_data

New Contributors

@rickdott made their first contribution in https://github.com/GWeindel/hmp/pull/126
@nmararo made their first contribution in https://github.com/GWeindel/hmp/pull/141

Full Changelog: https://github.com/GWeindel/hmp/compare/v0.2.0...v0.4.0

- Python
Published by GWeindel almost 2 years ago

https://github.com/gweindel/hmp - v0.2.0

Major changes

Possibility to fit condition-wise by sharing stages, topologies and varying the number of events across conditions (see Tutorial 4)
Estimates peak-to-peak latencies instead of onset-to-onset, this change does modify the duration of the first and last stage estimated by a magnitude of half an event width
The minimum duration for a stage has been changed from a location imputed to the PDF of stage duration to a censoring of the number of samples corresponding to that duration.
Added LOOCV method for any kind of fitted object
Bootstrap now also accept any model and will bootstrap on any dimension
Simulations can now be done on 'eeg'/'meg'/'eeg/meg' and at the desired sampling frequency

- Full support for an extended set of PDF: lognormal, Wald, Weibull, log-logistic, Chi

Minor changes

Added function to recover generating parameters and magnitudes
improved cluster plots
small fixes on the EM method
Re-introduced penalty on the channel contribution
Docstrings mostly present (except resample.py)
compute_topologies is faster and can also be done on a moving average of the size of an event

- Python
Published by GWeindel over 2 years ago