Struct SchemaV2 

pub struct SchemaV2 {
    pub schema_version: u32,
    pub input: Option<InputSpec>,
    pub outputs: Vec<LogicalOutput>,
    pub nms: Option<NmsMode>,
    pub decoder_version: Option<DecoderVersion>,
}

Root of the edgefirst.json schema v2 metadata.

All fields except SchemaV2::schema_version are optional, so third-party integrations can include only the sections relevant to their use case.

Fields

schema_version: u32

Schema version. Always 2 for v2 metadata.

input: Option<InputSpec>

Input tensor specification (shape, named dims, camera adaptor).

Required for decoders that need to know the model input resolution: DFL dist2bbox scaling, box normalization against input dimensions, and per-scale anchor grid generation.

outputs: Vec<LogicalOutput>

Logical outputs describing the model’s output tensors and their semantic roles.

nms: Option<NmsMode>

HAL NMS mode. Omitted for end-to-end models with embedded NMS.

decoder_version: Option<DecoderVersion>

YOLO architecture version for Ultralytics decoders.

Values: yolov5, yolov8, yolo11, yolo26. yolo26 is end-to-end (embedded NMS).

Implementations

impl SchemaV2

pub fn parse_json(s: &str) -> DecoderResult<Self>

Parse schema metadata from a JSON string.

Auto-detects the schema version from the schema_version field. Absent or 1 → legacy v1 metadata converted to v2 in memory. 2 → parsed as v2 directly. Any version higher than MAX_SUPPORTED_SCHEMA_VERSION is rejected with DecoderError::NotSupported.

pub fn parse_yaml(s: &str) -> DecoderResult<Self>

Parse schema metadata from a YAML string.

Same version-detection logic as SchemaV2::parse_json.

pub fn parse_file(path: impl AsRef<Path>) -> DecoderResult<Self>

Parse schema metadata from a file, auto-detecting JSON vs YAML from the file extension. Unknown extensions are parsed as JSON first then as YAML as a fallback.

pub fn from_json_value(value: Value) -> DecoderResult<Self>

Parse from an already-deserialized serde_json::Value. Useful for callers that have already done the initial deserialization step.

pub fn from_v1(v1: &ConfigOutputs) -> DecoderResult<Self>

Convert a legacy v1 ConfigOutputs to an equivalent v2 SchemaV2 in memory.

The conversion preserves:

• Output order and types (mapped to their v2 LogicalType).
• Quantization (v1 QuantTuple(scale, zp) → v2 Quantization with a single scalar scale/zero_point and unspecified dtype).
• dshape, shape, anchors, normalized fields.
• Root-level nms and decoder_version.

Fields v1 does not carry (tensor dtype, per-channel quant, box encoding, score format, activation metadata, stride on physical children) are left as None. The v2 decoder is expected to infer these from the runtime tensor type and the legacy decoder dispatch rules.

impl SchemaV2

pub fn to_legacy_config_outputs(&self) -> DecoderResult<ConfigOutputs>

Downconvert a v2 schema to a legacy ConfigOutputs for the v1 decoder dispatch path.

Each LogicalOutput maps to one ConfigOutput variant selected by LogicalType; per-tensor scalar quantization becomes a QuantTuple(scale, zp); and decoder, anchors, normalized are copied verbatim.

This conversion does not reject logical outputs that also declare physical children (per-scale FPN splits, or channel sub-splits such as ARA-2 boxes_xy / boxes_wh). The returned legacy config captures the logical-level metadata, while the physical-to-logical merge is handled separately by the DecodeProgram that DecoderBuilder::build compiles alongside this legacy config.

Returns DecoderError::NotSupported when the schema uses features the v1 decoder cannot express at the logical level:

• Per-channel quantization arrays on a logical output.
• encoding: dfl on a flat logical output (no physical children). DFL combined with per-scale children is handled by the merge path (see [crate::decoder::merge::DecodeProgram]) which decodes the distribution before producing the merged post-decode (1, 4, total_anchors) tensor the legacy decoder consumes.

pub fn validate(&self) -> DecoderResult<()>

Validate the schema against the rules the HAL enforces.

Rules checked:

• schema_version in [1, MAX_SUPPORTED_SCHEMA_VERSION].
• Physical children, when present, carry non-empty name fields so tensor binding by name is unambiguous.
• All physical children of a given logical output have pairwise distinct shapes (shape-based binding safety, per HailoRT spec).
• For a boxes logical output with encoding: dfl, every physical child shape has a num_features (or last) dimension divisible by 4 (the box-coordinate count).
• Per-scale splits carry stride on every child.
• Mixed per-scale + channel-sub-split decompositions are rejected (the spec permits only one merge strategy per logical).
• end2end models (decoder_version=yolo26 with detections output) do not also declare per-scale split children on that output.

Trait Implementations

impl Clone for SchemaV2

fn clone(&self) -> SchemaV2

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for SchemaV2

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

Formats the value using the given formatter.

impl Default for SchemaV2

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for SchemaV2

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl PartialEq for SchemaV2

fn eq(&self, other: &SchemaV2) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl Serialize for SchemaV2

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.

impl StructuralPartialEq for SchemaV2