v2.0.0

PyPi release: https://pypi.org/project/osl-dynamics/2.0.0/

This is the first version to be included in FSL.

Changes:

• License changed from MIT to Apache-2.0 (needed to be changed for FSL).
• Option to stop HMM training early based on maximum fractional occupancy change.
• Removed WIP models and examples.

v1.4.5

PyPi release: https://pypi.org/project/osl-dynamics/1.4.5/

Changes:

• New method for automated bad segment removal.
• Standalone GLM module.
• Support for loading mat73 files is deprecated.
• Various bug fixes.

v1.4.4

PyPi release: https://pypi.org/project/osl-dynamics/1.4.4/

Changes:

• Coherence calculation with analysis.spectral.regression_spectra modified: fixed issue with nan.
• Changed feature of validation split method of the Data object to split within sessions.
• Option to save model checkpoint. This allows the user to resume training in a separate script.
• Added method to the Data object to align the sign of channels across sessions.
• Added new single-channel DyNeMo model (still a work in progress).

v1.4.3

PyPi release: https://pypi.org/project/osl-dynamics/1.4.3/

Changes:

• Tested installation against TensorFlow versions 2.11-2.15.
• Added functions to calculate partial coherence and partial directed coherence.
• Benchmarked the multitaper calculation against the HMM-MAR.
• Update docs.

v1.4.2

PyPi release: https://pypi.org/project/osl-dynamics/1.4.2/

Changes:

• Cost function calculation now takes the average over the time dimension by default.
• Option to calcalate Viterbi path with the HMM.
• Option to calculate cross multitaper spectra.
• Tweaked the scaling of subject/state-specific multitaper/welch spectra.
• Allow 3D parcellations in plotting.
• Removed source files for S/MAGE (haven't been supported for a while).

v1.4.1

PyPi release: https://pypi.org/project/osl-dynamics/1.4.1/

Changes:

• Update tutorials.
• Reduced default buffer size (for shuffling) from 100000 to 4000.
• Updated TINDA functions.

v1.4.0

PyPi release: https://pypi.org/project/osl-dynamics/1.4.0/

This release uses TensorFlow 2.11.1 and tensorflow-probability 0.19.

Changes:

• Updated installation instructions.
• Added M-DyNeMo config API wrapper.

v1.3.2

PyPi release: https://pypi.org/project/osl-dynamics/1.3.2/

This is the last release using TensorFlow 2.9.1 and tensorflow-probability 0.17 (next release will use newer versions).

Changes:

• Models:
  • Refactored M-DyNeMo.
  • Enhanced HIVE/DIVE and fixed bugs.
  • Added Simplified-DyNeMo.
  • Improved robustness of random_state_time_course_initialization.
• Option to select session and/or channels in the Data object.
• Other features/enhancements
  • Cleaned up messages printed to screen (suppressing external loggers).
  • Option to combine power map/connectivity network plots.
  • Added standalone HMM dual estimation function.
  • New function to plot HMM summary stats.

v1.3.1

PyPi release: https://pypi.org/project/osl-dynamics/1.3.1/

Changes:

• Models:
  • The efficiency of the model initialisation methods (random_subset_initialization, random_state_time_course_initialization) was improved (minimised the number of shuffles).
  • Methods was updated to ensure a TensorFlow (TFRecord) Dataset can be passed.
  • Improvements to H/DIVE:
    • Modification to the calculation the KL term in the loss.
    • Ability to pass multiple embeddings.
• Data object:
  • Option to pass arbitrary auxiliary inputs to models when creating datasets with the Data object.
  • Option to save/load TFRecord datasets (useful for training on very large datasets).
• Simulation classes:
  • random_seed argument was removed - this may cause old scripts to error due to the unexpected argument (can just be deleted in the script). The user can use osl_dynamics.utils.misc.set_random_seed to ensure scripts are deterministic now.
• Plotting:
  • Improved spatial map plotting to work with fMRI data (can now handle cifti files).

v1.3.0

PyPi release: https://pypi.org/project/osl-dynamics/1.3.0/

Changes:

• Models:
  • Subject embedding models finalised: HIVE and DIVE.
  • New HMM with a Poisson observation model.
• Data class:
  • New method to select channels.
  • No longer uses memory maps by default.
• Added decoding examples.