Struct ProtoData 

pub struct ProtoData {
    pub mask_coefficients: TensorDyn,
    pub protos: TensorDyn,
}

Raw prototype data for fused decode+render pipelines.

Holds post-NMS intermediate state before mask materialization, allowing the renderer to compute mask_coeff @ protos directly (e.g. in a GPU fragment shader) without materializing intermediate Array3<u8> masks.

Both fields are carried as [TensorDyn] so downstream consumers (Rust, C API, Python) get zero-copy typed access through the HAL’s shared tensor infrastructure. Dtype policy:

Source model	protos dtype	mask_coefficients dtype	protos.quantization
int8 quantized	[TensorDyn::I8]	[TensorDyn::F32] (dequantized)	Some(q)
f32	[TensorDyn::F32]	[TensorDyn::F32]	None
f16 (TensorRT fp16)	[TensorDyn::F16]	[TensorDyn::F16]	None
f64 (narrowed)	[TensorDyn::F32]	[TensorDyn::F32]	None

Quantization metadata lives on the proto tensor itself via edgefirst_tensor::Tensor::quantization — float tensors cannot carry quantization (compile-time gated on the IntegerType sealed trait).

TensorDyn is not Clone, so neither is ProtoData. Consumers that need to share the proto buffer across threads should use TensorDyn::clone_fd / dmabuf_clone to dup the backing fd.

Fields

mask_coefficients: TensorDyn

Per-detection mask coefficients, shape [num_detections, num_protos].

protos: TensorDyn

Prototype tensor, shape [proto_h, proto_w, num_protos].

Trait Implementations

impl Debug for ProtoData

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.