tobil's picture
Publish selected model-only bundle for Omatrack 1.8.3
27ebca0 verified
Raw History Blame Contribute Delete
4.36 kB
{
"schema": "gauge-detector-center-v1",
"model_file": "gauge-detector.onnx",
"model_sha256": "99bdd483402444e8c799452d9885c1184980a41f2cd80d4af4d7a6b3ae00abc6",
"model_bytes": 966419,
"contract_file": "contract.json",
"contract_sha256": "87a4cec9e0466250b635e561c74be81f2ce5f15fbd79804a1be639c3719a69d3",
"checkpoint_sha256": "42fca6a5cad028bd21c7a05195c93c02597f8ffe2c5280875a5f1254029c0f83",
"selected_epoch": 7,
"frozen_config_sha256": "a58dcdd47804b38eb2dfeccce12f43f570628bf82117ccdf62975b16a0244280",
"parameters": 237142,
"onnx_opset": 17,
"operators": [
"Add",
"Cast",
"Concat",
"Constant",
"Conv",
"Div",
"Gather",
"GlobalAveragePool",
"HardSigmoid",
"HardSwish",
"Mul",
"Relu",
"Resize",
"Shape",
"Slice",
"Sub",
"Unsqueeze"
],
"onnx_version": "1.19.1",
"ort_version": "1.24.2",
"status": "experimental_candidate_not_independently_accepted",
"input": {
"name": "image",
"dtype": "float32",
"layout": "NCHW",
"shape": [
1,
3,
384,
640
],
"channels": "RGB",
"range": [
0,
1
],
"spatial_multiple": 32,
"dynamic_spatial": true,
"resize": "stretch_full_source_half_pixel_bilinear_no_antialias",
"pixel_formula": "sx=(dx+0.5)*source_width/dest_width-0.5; sy analogous; floor coordinates, clamp each neighbor index; interpolate raw uint8 RGB in float32 then divide by 255; no intermediate uint8 rounding",
"normalization": "divide_255_only; ImageNet normalization is inside the graph"
},
"outputs": [
{
"name": "center_logits",
"channels": 1
},
{
"name": "size_logits",
"channels": 2,
"order": [
"width",
"height"
]
},
{
"name": "offset_logits",
"channels": 2,
"order": [
"x",
"y"
]
},
{
"name": "representation_logits",
"channels": 4,
"labels": [
"digits",
"bar",
"wheel",
"needle"
]
},
{
"name": "semantic_logits",
"channels": 5,
"labels": [
"unknown",
"gear",
"stint_lap",
"brake",
"throttle"
]
}
],
"decode": {
"score_threshold": 0.35,
"local_max_kernel": 3,
"local_max_rule": "sigmoid(center) >= every in-bounds 3x3 neighbor; equal maxima all retained before top-k",
"candidate_limit": 100,
"sort": "descending score then ascending flat y*grid_width+x index",
"center": "((x+sigmoid(offset_x))/grid_width,(y+sigmoid(offset_y))/grid_height)",
"size": "(sigmoid(width_logit),sigmoid(height_logit)) normalized full-source width/height",
"box": "center +/- size/2; clip each coordinate to [0,1]; discard zero-area",
"nms": "class-agnostic greedy suppress IoU strictly > 0.4",
"nms_iou": 0.4,
"max_detections": 32,
"representation": "argmax; lowest index wins ties",
"semantic": "softmax argmax; lowest index wins ties; emit unknown if max probability < 0.8",
"semantic_threshold": 0.8
},
"limitations": [
"scores are not calibrated probabilities",
"unknown semantic has positive supervision only for explicitly observed non-Tier1 fields in four TRAIN images; ambiguous fields and generic lap-counter ownership remain masked; no open-set certification",
"all synthetic semantic loss fully masked; no renderer-intent labels",
"real weak windows use positive MIL only, no extent or offset loss or background negatives; four pixel-reviewed TRAIN images provide stronger geometry with explicit ignored regions",
"bar includes segmented indicators, no fill direction or scale predicted",
"visibility/readability/support remain unknown",
"detector boxes never imply arbitrary crop reader support",
"confirmed user/reviewed configuration required before extraction",
"only two known real layouts and four reviewed TRAIN background assets; repeated views are not new source diversity"
],
"licensing": "torchvision BSD-3-Clause backbone; task weights redistribution license unspecified; no public upload",
"required_metadata_keys": [
"schema",
"model_file",
"model_sha256",
"contract_file",
"contract_sha256",
"status",
"input",
"outputs",
"decode",
"limitations"
]
}