github Jellman86/YetAnother-WhosAtMyFeeder v2.22.0
YA-WAMF 2.22.0

23 days ago

YA-WAMF 2.22.0 makes your feeder history easier to review and clarifies the difference between
hardware compatibility and species accuracy. Photos stay with your records under explicit storage
limits, while new model choices can be compared on your own hardware.

What's new

  • More local model choices, with species names built in. Model Manager adds DINOv2, newer
    RoPE models, regional NaFlex and Parallel models, and BioCLIP 2.5. Registered classifiers use
    the species catalogue rather than requiring a label text file. Regional coverage matters;
    compare coverage, sizes and tested hardware.
  • Setup that saves the model and provider you chose. The wizard offers a complete location
    pair for DINO, saves it before validation and keeps a passing manual provider choice. Continue
    saves the displayed preference. Reopening setup preserves unrelated Settings edits.
  • Seen visits and heard calls together. Explorer's optional Heard calls lane shows stored
    call bouts between camera visits, with spectrograms and playback. Matching calls attach to
    a seen visit. Guest sharing rules and visual-only filter restrictions still apply.
  • A clearer review of uncertain species. Needs a check shows the photographs behind a
    name, lets you correct selected visits and handles visits with many captures. A rename also
    teaches personalised identification; inspect the selected scope before applying it.
  • Count and review birds within one capture. Retained full frames can show separate bird boxes,
    crops and species suggestions. Correct or exclude individual birds without creating extra visits;
    counts remain estimates and do not identify a bird across frames.
  • A calmer view on phones and wide screens. Dashboard and Explorer use the available width,
    with clearer species names, responsive charts and adjustable text size. Visit groups keep the
    original captures accessible without counting every camera frame as another visit.

Smoother everyday use

  • A compatibility check no longer looks like 0% accuracy. Results show passing providers and
    measured inference time. Detection and Models refresh the workers' actual provider after a
    check, so a move to NPU does not leave an old CPU label on screen.
  • Photos follow retention controls. The old Media Cache and Snapshots switches are removed.
    Photographs and visit media are retained within age, per-species and storage limits; copies of
    Frigate's event clips remain optional. Unwritable media storage is reported as degraded health.
  • Less repeated video decoding. Photo updates can reuse the winning frame without changing
    the accepted species or counted-bird evidence. CPU, Intel and one NVIDIA configuration have
    been tested; sustained production backfill with live ingestion remains under review in
    #576.
  • Clearer failures and fresh support evidence. Backfill stops after repeated classifier
    failures, video retries keep their limits, and missing snapshots no longer pause unrelated
    inference work. Support bundles collect fresh server evidence; opening Settings no longer
    repeats old backfill completion messages.
  • Guest sharing follows the saved boundaries. History, audio, media and cached results respect
    the current visibility rules. Optional external services still receive only the data described
    in the privacy guide.

Before you update

  • Back up persistent data first. Preserve /config and /data, including configuration,
    database, models and favourite archive. This update includes migrations; see the
    upgrade and rollback guide before reverting to an older image.
  • Check the one-time retention migration. An installation that disabled either retired
    photo/cache switch has best-available snapshots disabled once; a previously disabled cache also
    disables full-visit clips once. Explicit environment overrides and later owner choices are
    preserved. Review Settings → Data and Settings → Connection after updating.
  • Your default classifier stays the same. New models are optional and need the matching
    application build. Downloading a new model asset does not add support to an older application.
    DINO is experimental; location is optional and sent to its local inference input, not embedded
    in redistributed weights. Hardware compatibility does not prove accuracy at your feeder.
  • Keep your image family unless you intend to change it. Unsuffixed tags remain the full
    compatibility image. Optional CPU, Intel and CUDA variants are described in
    Hardware Acceleration.
  • Additional-bird scanning is off by default, including on upgrade. Normal identification
    continues. Use Find more birds for a manual scan, or enable Automatically find additional
    birds
    under Settings → Data → Snapshot quality if you want the extra background work.
  • Standard text is one step smaller. The default is now 87.5% of your browser's size. Choose
    Large under Settings → Appearance → Text size for the previous size; the choice is saved
    separately on each device.

Good to know

  • The original #490 fault is still unconfirmed. Worker isolation and bounded same-model CPU
    recovery contain failures, but they do not prove the reporter's EVA-02/Intel GPU stall is gone.
    A healthy NPU run is evidence for that tested configuration only.
  • Video requires recordings. A Frigate event with a snapshot and has_clip: false can still
    classify; it cannot supply a missing event clip. Enable and retain recordings if you want video
    analysis or full-visit playback.
  • Garden-visitor history and full identification provenance are deferred. The proposals in
    #641 and
    #643 are not included as shipped features.

Thanks

Thanks to @Milirey for the detailed backfill/live-feed report and
support bundles in #490.
That evidence helped improve isolation and recovery while keeping the original hardware fault
open for confirmation.

Thanks to @aav7fl for the original multi-bird clips, screenshots,
diagnostics and follow-up testing in #481.
Their report informed the photo repair and the still-open decoding and retention follow-ups,
#576 and
#622.

Thank you as well to everyone listed in Contributors.
See the full changelog for the complete technical record.

Don't miss a new YetAnother-WhosAtMyFeeder release

NewReleases is sending notifications on new releases.