Model-info reference

model-info is a small JSON file next to a compiled .mxr model. It replaces DL Streamer's model-proc and DeepStream's config_infer_*.txt preprocessing and post-processing fields. It is parsed by ModelInfo::load (include/amdva/model.hpp, src/model.cpp) and shared by all amd* inference elements and the SDK.

{
  "task_type": "detection",
  "adapter": "yolov8",
  "input": {
    "name": "images",
    "shape": [1, 3, 640, 640],
    "format": "NCHW",
    "channel_order": "RGB",
    "letterbox": true,
    "scale": 0.00392156862,
    "mean": [0, 0, 0],
    "std": [1, 1, 1],
    "padding": 114
  },
  "output": { "format": "cxcywh", "num_classes": 80, "top_k": 5, "keypoints": 17 },
  "labels": "coco80.txt"
}

Fields

Field	Default	Meaning
task_type	detection	detection, instance-segmentation, pose, classification, semantic-segmentation or raw
adapter (alias family)	yolov8 for detection, segmentation and pose; otherwise the task name	Output decoder. Detection adapters: yolov8, yolov8-seg, yolov8-pose. YOLOv5, SSD and custom parsers are not supported.
input.name	images	Model input tensor name
input.shape	required	[N,3,H,W]: N 1-16, 3 channels, H and W 2-4096. Static shapes only.
input.format	NCHW	Only NCHW is accepted
input.channel_order	RGB	RGB or BGR
input.letterbox	true; false for classification and semantic segmentation	Preserve aspect ratio and pad, or resize directly
input.scale, mean, std	1/255, [0,0,0], [1,1,1]	Applied as (pixel * scale - mean) / std per channel, in the channel order above. std must be positive.
input.padding	114	Letterbox pad value
input.normalize	none	Legacy; only [0, 1] is accepted. Use scale/mean/std instead.
output.num_classes	80	Class count, 1-65536. Must equal the number of lines in the labels file.
output.top_k	5	Classification results kept, 1-100
output.keypoints	17	Keypoints per person for pose, 1-64
output.format	informational	Decoded box layout of the model (cxcywh for YOLOv8)
labels	indices as strings	Text file, one label per line, relative to the JSON file

YOLO adapters expect decoded center/size boxes with class probabilities and no separate objectness channel ([B, 4+C, N]). Segmentation adds [B,M,H,W] prototypes; pose adds K x 3 keypoint channels. Classification expects logits [B,C]; semantic segmentation [B,C,H,W].

Compiling

amdva-compile-model model.onnx model.mxr --batch=4 --precision=fp16 \
  --model-info=model-info.json [--input-shape=NAME:D1,D2,...]

The compiler writes model.mxr.manifest.json (schema 2) containing the SHA256, GPU architecture, batch size and offload_copy=false. At load time the runtime refuses a model whose manifest is missing, whose hash differs, or that was built for a different GPU architecture. Rebuild when you change GPU or batch size. --precision=fp16 (default) changes internal computation only; external tensors keep their original types.

The elements require one float32 NCHW image input. Multi-input models and other tensor types (float16, int32, int64, uint8) are available through the C++ ModelExecutor API.

Mapping from other formats

Other format	Field here
DL Streamer model-proc input_preproc.resize, color_space	input.letterbox, input.channel_order
DL Streamer model-proc output_postproc.converter	adapter / task_type
DL Streamer labels	labels
DeepStream net-scale-factor	input.scale
DeepStream offsets	input.mean (in scaled units: offset * scale)
DeepStream model-color-format (0 RGB, 1 BGR)	input.channel_order
DeepStream maintain-aspect-ratio	input.letterbox
DeepStream num-detected-classes	output.num_classes
DeepStream labelfile-path	labels
DeepStream batch-size	element batch-size and amdva-compile-model --batch