YA-WAMF 2.17.0

YA-WAMF 2.17.0

YA-WAMF 2.17.0 gives you a complete way to add sightings from your own photos or videos and makes
reclassification much safer for brief, distant, or partially tracked birds. It also keeps model and
accelerator choices honest for the exact image and hardware running your feeder.

What's new

• Add an observation from a photo or video. The new Observe → Add observation page retains
  the original media locally, classifies it through the normal full-frame, crop, or video path, and
  lets you confirm the species before it appears in history. Upload progress survives a browser
  reload, duplicate media is rejected, and failures can retry without another upload.
• Place uploaded sightings on the map. YA-WAMF reads valid EXIF GPS coordinates when available.
  If a photo has no location, you can place a pin, enter coordinates, leave the location empty, or
  clear it before saving.
• Use acceleration only after it is proved on this installation. The setup wizard, Settings,
  Model Manager, and automatic provider selection now share the same image-aware hardware evidence.
  Providers appear in measured preference order and are offered only when the current model, image,
  runtime, and visible hardware support them.

Smoother everyday use

• Reclassification handles fleeting birds more fairly. Video analysis looks for recurring,
  time-separated evidence instead of requiring a bird to fill most of a long visit. Nearby frames
  cannot manufacture extra votes, conflicting full-frame and crop results still abstain, and weak
  evidence cannot replace an existing identification.
• Cached media remains useful after Frigate retention expires. Reclassification prefers a
  playable retained recording or event clip. If video cannot produce a safe result, the same job
  visibly tries the best retained snapshot and explains when the identification stays unchanged.
• Model choices are simpler and safer. Four redundant comparison models no longer appear in
  current setup or Model Manager flows. If one was previously active, YA-WAMF moves to an installed
  ConvNeXt Large model when available or the bundled fallback, while leaving the old files in place
  for rollback.

Before you update

• Back up /config and /data, then update normally. This release adds idempotent database
  migrations for manual-observation drafts, optional observation coordinates, and retained video
  evidence diagnostics. YA-WAMF applies them automatically during startup; keep the container
  running until the startup progress view reports that the backend is ready.
• Keep the unsuffixed tag unless you intentionally use a smaller runtime image. latest and
  v2.17.0 remain the full compatibility build. Read the
  hardware-acceleration guide
  before switching to v2.17.0-cpu, v2.17.0-intel, or v2.17.0-cuda; change only the image tag
  so the existing /config, /data, models, and provider preference remain intact.

Good to know

• Uploaded media stays on your YA-WAMF storage. It is not dependent on Frigate retention.
  Location is optional, and coordinates are saved only when you confirm an EXIF value or map pin.
• Hardware validation is deliberately installation-specific. A GPU or NPU that is visible but
  has not passed the current model and runtime checks remains unavailable until validation succeeds.
  This avoids treating an older result from another image, artifact, or runtime as proof.
• The new video rule remains conservative. A source still needs three independent evaluated
  moments, at least two confident votes, and 60% agreement among confident votes. When trustworthy
  image representations disagree, YA-WAMF keeps the existing identification.

See the full changelog
for the complete technical record.