v0.4.0

What's Changed

Added

• OutputQuantization::effective(qmode) helper that normalizes
  Kinara's per-qmode dequantization formulas into a single
  (scale, offset) pair. Only qmode 9 is currently supported; other
  modes return Error::UnsupportedQmode.
• InputPreprocess struct (core) / class (ara2-py) holding the
  per-channel image normalization parameters (mean, scale,
  bgr_to_rgb, aspect_resize, mirror, center_crop), queried via
  Model::input_preprocess(i) / model.input_preprocess(i).
• InputQuantization.qmode and InputQuantization.offset fields
  sourced from dv_model_input_preprocess_param::qmode and ::offset.
• Ara2Info metadata section (DvmMetadata.ara2) parsing the
  optional ara2.qmode field from edgefirst.json embedded in DVMs.
• Session.close() and Model.close() Python methods; both are
  idempotent and are called by the __exit__ path so with blocks
  now actually release resources.
• Error::UnsupportedQmode(i32) variant raised by dequantize() when
  a model uses a quantization mode other than 9.

Changed

• Model::dequantize() (core and Python) now uses the correct
  qmode-9 formula (raw - offset) * scale. Previously applied the
  qmode 0-3 formula raw / qn, which silently produced values off
  by several orders of magnitude on current production models.
• ara2-py::Model.set_input_tensor accepts any numpy array whose
  total byte length matches the tensor size — the buffer is obtained
  via tobytes() and memcpy'd verbatim. Callers no longer need to
  .view(np.uint8) before calling. Non-contiguous arrays are handled
  transparently (numpy makes a contiguous copy on the fly).
• ara2-py::Model.get_output_tensor returns a typed array
  (int8/uint8/int16/uint16/float32) reshaped to the
  tensor's declared (C, H, W) shape. Callers that relied on the
  legacy flat-uint8 return must either use the typed array directly
  or call .ravel().
• ara2-py::Session and ara2-py::Model now wrap Option<inner>
  internally. Method calls on a closed instance raise
  Ara2Error("session is closed") / "model is closed" instead of
  returning stale data.

Fixed

• Quantization: dequantize() produced grossly wrong values on
  every qmode-9 model (which is every production DVM today).
• Input dtype erasure: set_input_tensor rejected any array that
  wasn't uint8, forcing a .view(np.uint8) workaround in every
  consumer.
• Output dtype erasure: get_output_tensor returned uint8
  regardless of the tensor's actual signedness, forcing a manual
  .view(int8) on consumers.
• Input zero-point confusion: Consumers were reading
  InputQuantization.mean as an integer zero-point — but mean was
  the per-channel float normalization mean, not a quantization
  zero-point. The actual zero-point is now exposed as
  InputQuantization.offset.
• Missing close: Session.__exit__ and Model.__exit__ were
  no-ops; there was no way to deterministically release resources
  outside a context manager. close() methods plus a real __exit__
  body fix both cases.

Removed (BREAKING)

• OutputQuantization.scale field (was output_scale in the C
  struct; unused by every downstream consumer).
• InputQuantization.mean and InputQuantization.scale fields —
  moved to InputPreprocess.

Migration

Consumers of edgefirst-ara2 0.3.x must update:

0.3.x	0.4.0
int(iq.mean) as zero-point	iq.offset (true zero-point)
iq.mean, iq.scale (per-channel)	model.input_preprocess(i).mean, .scale
oq.scale (used with old dequantized = raw / qn formula)	oq.qn (used with new dequantized = (raw - offset) * qn formula — see Changed section)
model.set_input_tensor(0, arr.view(np.uint8))	model.set_input_tensor(0, arr)
raw.view(np.int8) after get_output_tensor	unnecessary — returned array is already typed
session.__exit__(None, None, None)	session.close()
model.__exit__(None, None, None)	model.close()

Non-qmode-9 DVMs now raise Ara2Error("unsupported quantization mode: qmode=N ...") from dequantize(). If you encounter this, file an issue with the model so qmode 0-3 support can be added with a test fixture.

---

crates.io: ara2 | ara2-sys
PyPI: edgefirst-ara2