v0.14.2

v0.14.2

Features

• Load pose estimation data directly from DANDI. Installable with the dandi extra: pip install vame-py[dandi] (PR #222)
• Gracefully stop an in-progress training run with vame.model.stop_training(config)(PR #220)
• Faster model training (PR #220)
• Project-wide reproducibility: project_random_state now seeds every RNG VAME draws from (Python, NumPy, Torch) across preprocessing, segmentation, community, UMAP, reports, and video/generative steps via the new vame.util.seed.seed_everything (PR #221)

Fixes

• Run the Gram-matrix SVD on CPU during training: SVD backward has no MPS autograd kernel and could silently produce wrong gradients on Apple Silicon (PR #220)

Docs

• Add a "Read from DANDI" how-to guide and example notebook under a new "How to" section (PR #222)
• Remove the standalone landing page (the site root now redirects to the docs introduction) and the unused blog (PR #219)

v0.14.1

v0.14.1

Fixes

• Roll back a partially-created project when initialization fails (PR #218)
• Speed up motif/community video clips generation (PR #218)
• Name copied/linked videos after their paired session and validate one video per pose estimation file (PR #218)
• Handle motifs absent from a session in the community motif report (PR #218)

v0.14.0

v0.14.0

Features

• Register pre-computed scalar features (speed, head direction, mask area, etc.) for model training via vame.io.add_extra_features and vame.validate_extra_features (PR #213)
• Accept movement-format .nc files as pose estimation input via source_software="movement", with strict schema validation (PR #213)
• Update pose loader to movement 0.16 unified loader interface (PR #212)

Fixes

• Fix Windows compatibility in project init and sample data download. (PR #209)

v0.13.0

v0.13.0

Features

• Read pose estimation data from NWB files (ndx-pose).

Fixes

• Bump imageio-ffmpeg pin from 0.5.1 to 0.6.0 so a prebuilt wheel (with the bundled ffmpeg binary) is available on Apple Silicon Python 3.11.

v0.12.1

v0.12.1

Fixes

• Add Apple Silicon (MPS) GPU support
• Centralize device detection in get_device() utility

v0.12.0

Fixes

• Correct dendogram motif ordering (Issue #175)

Features

• Optional to remove body parts from training (Issue #168)
• Motif thresholding plot (Issue #45)
• Uses project-wide seed in the model training and evaluation modules

Others

• Dropped support for MacOS 13 and older, due to dependencies incompatibility (pytorch, numpy and numba).

v0.11.0

Fixes

• Loose torch version torch>=2.3.1
• Updated pynwb==3.0.0
• Updated movement==0.6.1
• Updated ndx-pose==0.2.2
• Updated ndx-vame==0.2.2
• Remove deprecated verbose argument for ReduceLROnPlateau
• Fix gpu usage (Issue #173)
• Fix redundant inference when estimating motifs (Issue #176)
• Fix UMAP, now it uses all sessions (cohort) (Issue #177)
• Fix preprocessing (Issue #180)

Features

• Export Project results to NWB files (Issue #171)
• UMAP figures with proper labels and colors
• Interactive UMAP figures with Plotly
• tensorboard tracking during model training
• Batch processing speed up latent space embeddings
• Project-wide random state with project_random_state in config
• Creates pose_estimation_inspection function for quick inspection of useful information

v0.10.0

Fixes

• Environment file for fixing GPU not found on Windows

Features

• Better organization for Report and UMAP results
• Logs True by default for all functions

v0.9.0

Fixes

• Fix movement sets fps to None, which is not accepted by ds.to_netcdf (Issue #148)
• Re-order dataset dimensions to be in line with movement
• Fix imports for download sample data
• Fix get_version (Issue #164)
• Removed smoothing filter on community labels and save community fig as a .pdf
• Save pose segmentation data after each session instead of segmentation algorithm method (Issue #128)
• Fix IQR filter (Issue #158)
• Fix bug with motif usage counts in community function (PR #156)
• Removed redundant folder in the results folder (PR #156)

Features

• Update torch version to 2.5.1
• Update movement==0.5.0
• Update netCDF engine to netCDF4
• Add VAME package version to config (Issue #147)
• Improve test/train set split (Issue #146)
• re-scaling as an extra preprocessing step, optional with default value False
• visualization of the raw, threshold-cleaned, egoaligned, and IQR-cleaned (non-z-scored) time series example and X-vs-Y scatter plots that show how the data is being transformed
• Points cloud scatter plot
• Add tests for M1 (Issue #149)
• General improvements to save_state in some functions
• Community analysis, motif videos and community videos are now done automatically for all segmentation algorithms

v0.8.0

Fixes

• OS aware dependencies installation, specially important for pytorch

Features

• Initialize project and run pipeline without videos (Issue #129)
• Allow adding videos later on in the project (Issue #129)
• Run on Google colab tags in documentation pages (Issue #144)