Module schema 

Schema v2 metadata types for EdgeFirst model configuration.

Schema v2 introduces a two-layer output model that separates the logical contract (what the model produces semantically) from the physical realization (what tensors the converter emitted). Each entry in the top-level SchemaV2::outputs array is a LogicalOutput. A logical output either IS a physical tensor (when the converter did not split it further) or contains a outputs array of PhysicalOutput children that realize it.

Example — YOLOv8 detection, flat (TFLite)

use edgefirst_decoder::schema::SchemaV2;

let json = r#"{
  "schema_version": 2,
  "outputs": [
    {
      "name": "boxes", "type": "boxes",
      "shape": [1, 64, 8400],
      "dshape": [{"batch": 1}, {"num_features": 64}, {"num_boxes": 8400}],
      "encoding": "dfl", "decoder": "ultralytics", "normalized": true,
      "dtype": "int8",
      "quantization": {"scale": 0.00392, "zero_point": 0, "dtype": "int8"}
    },
    {
      "name": "scores", "type": "scores",
      "shape": [1, 80, 8400],
      "dshape": [{"batch": 1}, {"num_classes": 80}, {"num_boxes": 8400}],
      "decoder": "ultralytics", "score_format": "per_class",
      "dtype": "int8",
      "quantization": {"scale": 0.00392, "zero_point": 0, "dtype": "int8"}
    }
  ]
}"#;

let schema: SchemaV2 = serde_json::from_str(json).unwrap();
assert_eq!(schema.schema_version, 2);
assert_eq!(schema.outputs.len(), 2);

Structs

InputSpec

Input tensor specification.

Letterbox

Letterbox geometry for resize: letterbox.

LogicalOutput

Logical output: the semantic contract the model exposes.

PhysicalOutput

Physical output: a concrete tensor produced by the converter.

Preprocessing

Input pre-processing declared by the model producer.

Quantization

Quantization parameters for a quantized tensor.

SchemaV2

Root of the edgefirst.json schema v2 metadata.

SemanticParams

Semantic segmentation decode fields of a logical output.

Upsample

Score-map upsampling.

Enums

Activation

Activation function applied to or required by a physical tensor.

BoxEncoding

Box encoding for boxes logical outputs.

DType

Quantized or floating-point data type of a model’s logical I/O.

DecodeMode

Semantic decode mode.

DecoderKind

Decoder framework for a logical output.

DecoderVersion

YOLO architecture version for Ultralytics decoders.

Interpolation

Resampling filter.

LetterboxPlacement

Letterbox placement of the scaled image.

LogicalType

Semantic type of a logical output.

MaskActivation

Activation over mask-classification mask logits.

NmsMode

HAL NMS mode.

Normalization

Input normalization: a named preset or explicit per-channel mean/std, both defined as (x - mean) / std on 0–255 values.

NormalizationPreset

Named normalization presets, matching the validation.normalization vocabulary.

PadSpace

Space a letterbox pad value is expressed in.

PhysicalType

Semantic type of a physical output.

ResizeMode

How a frame is fitted to the fixed model input.

ScoreFormat

Score format for scores logical outputs.

Stride

FPN stride. Square(s) means (s, s); Rect(sx, sy) supports non-square inputs.

UpsampleMethod

Score-map upsampling filter.

Constants

MAX_SUPPORTED_SCHEMA_VERSION

Highest schema_version this parser accepts. Files with a higher version are rejected rather than silently parsed against the wrong grammar.