* Let the zmq detector skip the input transform when the detector applies it
_transform_input converts the frame to the model's layout and dtype before
the detector sees it. For a remote detector that can apply the transform
itself, often on a GPU, that work is done twice over and the tensor is sent
in its most expensive form.
The detector now reports input_transform_offload in the model handshake. When
it does, Frigate sends the frame as captured (uint8 NHWC) and adds the model's
expected input_tensor and input_dtype to the request header, so the detector
can apply them. The model config is unchanged either way: the expected format
is relayed, not overridden.
The choice is read per frame rather than cached, because the detector redoes
its handshake on a timeout or socket error, so restarting or upgrading it
while Frigate runs already works and keeps working here.
Negotiated rather than configured, for two reasons. The detector is what knows
whether it can apply the transform, and a user who merely believed it could
would send unconverted frames to a model that cannot take them. It also needs
no config surface, which the device string format has no room for.
On a 640x640 region, measured on a Ryzen 9 5900XT, the host-side cost of
handing the detector a frame drops from about 4 ms to under 0.1 ms per
detection, and the payload from 4.9 MB to 1.2 MB. Nearly all of it is
tobytes(): np.transpose returns a view and astype keeps its strides by
default, so the transpose is deferred to serialization, where it becomes a
strided gather over the whole float32 tensor. The YUV to RGB conversion is
unaffected and stays on the host.
* Test that the handshake decides whether the input transform is applied
Parameterized over the three handshake values against the header the detector
ends up sending: reported, the frame goes over as captured uint8 NHWC with the
model's input_tensor and input_dtype declared; not reported or absent, the
existing transform runs and those fields are omitted.
A second case drives a real reconnect -- a timed out request, then a handshake
from an upgraded detector -- and checks the next frame follows the new answer
rather than the one from startup.
---------
Co-authored-by: Eric <eric@localhost>
