Skip to content

Latest commit

 

History

History
101 lines (97 loc) · 7.65 KB

CHANGELOG.md

File metadata and controls

101 lines (97 loc) · 7.65 KB
Raw

Changelog

  • 0.3 (dev): This is a major release, refactoring most of pysaliency.datasets including some breaking changes.

    • Feature: Scanpaths is a new class for storing scanpaths. It has a similar API to the old FixationTrains, but exclusively cares about scanpaths. For example, the length of a Scanpaths instance is the number of scanpaths, not the number of fixations. It is intended to make iterating over and working with scanpaths more convenient.
    • Feature: ScanpathFixations is a new subclass of Fixations intended for handling fixations that come from scanpaths. It is intended to replace the old FixationTrains class. ScanpathFixations has a scanpaths: Scanpaths attribute storing the source scanpaths (what used to be stored manually as train_xs etc in FixationTrains). Unlike FixationTrains, ScanpathFixations does not have any attributes that are not derived from the scanpaths. FixationTrains is now a deprecated subclass of ScanpathFixations which adds the old properties and constructor and allows for attributes which are neither scanpath attributes nor scanpath fixation attributes.
    • Feature: VariableLengthArray for inuititively handling data like scanpaths where each row can have a different length. Fixations.x_Hist, Fixations.y_hist, Scanpaths.xs etc are now instances of VariableLengthArray.
    • Fixations.lengths has been renamed to Fixations.scanpath_history_length to make it clear that it is the length of the scanpath history. Fixations.lengths is now a deprecated alias.
    • In general, naming convention for attriutes has been changed to use the plural form if the attribute is a list of values for each element (i.e., Scanpaths.xs) and the singular form if the attribute is a single value (i.e., Scanpaths.length, Fixations.x). This resulted in renaming Fixations.subjects to Fixations.subject. The old name is now a deprecated alias.
    • Bugfix: Compatibility with torch 2.2 and numpy 2.0
    • Bugfix!: The download location of the RARE2012 model changed. The new source code results in slightly different predictions.
    • Feature: The RARE2007 model is now available as pysaliency.external_models.RARE2007. It's execution requires MATLAB.
    • matlab scripts are now called with the -batch option instead of -nodisplay -nosplash -r, which should behave better.
    • Enhancement: preloaded stimulus ids are passed on to subsets of Stimuli and FileStimuli.
    • Feature: pysaliency.read_hdf5 now takes additional keyword arguments which are passed to the respective class methods. This allows, e.g., to load FileStimuli with caching disabled.

  • 0.2.22:

    • Enhancement: New Tutorial.
    • Bugfix: SaliencyMapModel.AUC failed if some images didn't have any fixations.
    • Feature: StimulusDependentSaliencyMapModel
    • Bugfix: The NUSEF dataset scaled some fixations not correctly to image coordinates. Also, we now account for some typos in the dataset source data.
    • Feature: CrossvalMultipleRegularizations and GeneralMixtureKernelDensityEstimator in baseline utils (names might change!)
    • Feature: DVAAwareScanpathModel
    • Feature: ShuffledBaselineModel is now much more efficient and able to handle large numbers of stimuli. hence, ShuffledSimpleBaselineModel is not necessary anymore and a deprecated alias to ShuffledBaselineModel
    • Feature: ShuffledBaselineModel can now compute predictions for very large numbers of stimuli without needing to have all individual predictions in memory due to a recursive reduce logsumexp implementation.
    • Feature: plotting.plot_scanpath to visualize scanpaths and saccades. WIP, expect the API to change!
    • Feature: DeepGaze I and DeepGazeIIE models
    • Feature: COCO Freeview dataset
    • Feature: optimize_for_information_gain(framework='torch', ...) now supports a cache_directory`, where intermediate steps are cached. This supports resuming crashed optimization runs.
    • Bugfix: fixed some edge cases in optimize_for_information_gain(framework='torch')
    • Feature: COCO Seach18 dataset
    • Feature: FixationTrains.train_lengths
    • Feature: FixationTrains.scanpath_fixation_attributes allows handling of per-fixation attributes on scanpath level, e.g. fixation durations. According attributes as in a Fixations instance are automatically created, e.g. for durations there will be an attribute durations and an attribute duration_hist. Also for scanpath_attributes (e.g., attributes applying to a whole scanpath, such as task) will also generate an attribute for each fixation to make this information available in Fixations instance.
    • Feature: scanpaths_from_fixations reconstructs a FixationTrains object from a Fixations instance
    • Bugfix: t_hist got replaced with y_hist in Fixations instances (but luckily not in FixationTrains instances)
    • Bugfix: torch code was broken due to changes in torch 1.11
    • Bugfix: SALICON dataset download did not work anymore
    • Bugfix: NUSEF datast links changed

  • 0.2.21:

    • Added new datasets: PASCAL-S and DUT-OMRON
    • Feature: FixedStimulusSizeModel and DVAAwareModel
    • Feature: Fixations finally support len()
    • Experimental feature: conditional_log_densities(stimuli, fixations) and conditional_saliency_maps(...). This is WIP to enable batch processing in models.
    • Fallback models for stimulus dependent models
    • MixtureScanpathModel
    • Reimplemented AUC for special case of only one positive sample, leading to substantial speedup
    • There is a new version of the CAT2000 train dataset which fixes some details in the processing. Since it changes the dataset, by default the old processing is used.
    • Feature: ShuffledSimpleBaselineModel. Baseline model to be used with ShuffledAUCSaliencyMapModel in cases where using ShuffledBaselineModel is not feasible.
    • pysaliency.get_toronto now returns a Fixations instance instead of FixationTrains since we don not have scanpath information.
    • pysaliency.baseline_utils.KDEGoldModel now supports a keyword argument grid_spacing which controls how densly the log density of the KDEModel is computed before it is linearly interpolated. This can substantially speed up computations on high resolution images.
    • Feature: pysaliency.precomputed_models.SaliencyMapModelFromArchive and ModelFromArchive for loading model predictions from ZIP, TAR and RAR files.
    • Bugfix: all matlab scripts where missing in the pip installation since the change to setuptools.

  • 0.2.20:

    • Stimuli now support attributes, just like Fixations. The CAT2000 train and test datasets now have the stimulus categories as attribute.
    • failure to download and setup a dataset will no longer result in leftover dataset files that keep pysaliency from trying again.
    • crossvalidation splits now support stratifying stimulus attributes
    • the MIT1003 dataset now also contains the history of fixation durations
    • FixationIndexDependentModel
    • Bugfix: The CC of a constant saliency map wrt to a nonconstant one now returns zero (instead of nan as previously).
    • Feature: Added keyword argument attributes to Fixations constructor
    • Feature: Provide KLDiv and SIM as functions that can be applied to saliency maps without need for a model.

  • 0.2.19:

    • added pytorch implementation for optimization of similarity metric as alternative to tensorflow implementation which still uses tensorflow 1.x
    • added pytorch implementation for saliency map processing as alternative to theano implementation.
    • removed obsolete dependency on openmp
    • made import of pytorch, theano and tensorflow optional
    • bugfixes in precomputed models for stimuli sets with nested directories