tempest-react-sdk 0.69.0 → 0.70.0

package/dist/vision/tasks/detectClassify.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"detectClassify.cjs","names":[],"sources":["../../../src/vision/tasks/detectClassify.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Run a fused detect→classify pipeline in the browser.\n *\n * The file this loads was built by the Python SDK's `ort_vision_sdk.compose`,\n * and it already contains both models plus the crop-and-resize bridge between\n * them. That matters far more in a tab than on a server: two models mean two\n * `.onnx` downloads, two WASM/WebGPU session initializations, and a per-crop\n * round trip through JavaScript to slice, resize and restack the regions before\n * the second model can see them. A fused pipeline has one download, one session\n * and no round trip — the crops are produced and consumed inside the graph.\n */\n\nimport type * as ort from \"onnxruntime-web\";\n\nimport { FusionError } from \"../core/exceptions\";\nimport { type ModelSource, type OrtSessionOptions, OrtSession } from \"../core/session\";\nimport { SpeedTimer } from \"../core/timing\";\nimport {\n    INPUT_IMAGE,\n    INPUT_PAD,\n    INPUT_SCALE,\n    INPUT_SOURCE,\n    OUTPUT_BOXES,\n    OUTPUT_CLASSES,\n    OUTPUT_NUM_DETECTIONS,\n    OUTPUT_PROBS,\n    OUTPUT_SCORES,\n    type FusionSpec,\n    readFusionSpec,\n} from \"../fusion\";\nimport { type ImageInput, loadImage } from \"../io/image\";\nimport { type LabelSpec, resolveLabels } from \"../labels\";\nimport { softmax, topK } from \"../postprocess/classification\";\nimport { toCHW, toFloat32, toFloat32Tensor } from \"../preprocess/image\";\nimport { LetterboxPipeline, zeroTensorData } from \"../preprocess/pipeline\";\nimport { Boxes, DetectClassifyResults } from \"../results\";\nimport {\n    BoundingBox,\n    RGBImage,\n    type ClassProbability,\n    type ClassificationResult,\n    type DetectionResult,\n} from \"../types\";\nimport { VisionTask, requireDetections } from \"./base\";\n\nexport interface DetectClassifyOptions extends OrtSessionOptions {\n    /**\n     * Class label spec for the **detection** stage — see {@link resolveLabels}.\n     * Defaults to the names recorded at fusion time, falling back to the COCO\n     * 80-class preset when the fusion recorded none.\n     */\n    readonly labels?: LabelSpec;\n    /**\n     * Class label spec for the **classification** stage. Defaults to the recorded\n     * names, falling back to generated `class_<id>` names.\n     */\n    readonly classifierLabels?: LabelSpec;\n    /**\n     * If `true`, a run that finds nothing throws {@link NoDetectionsError}\n     * instead of returning an empty envelope. Default `false`, because looking\n     * and finding nothing is a successful inference. Turn it on when an empty\n     * result means the surrounding pipeline should stop rather than carry on with\n     * zero rows. Can be overridden per `predict` call.\n     */\n    readonly raiseOnEmpty?: boolean;\n}\n\nexport interface DetectClassifyPredictOptions {\n    /**\n     * Drop detections scoring below this. The graph's own NMS threshold was fixed\n     * at fusion time and cannot be lowered here — this only filters further.\n     */\n    readonly confThreshold?: number;\n    /** If set, keep only detections whose detector `classId` is in this list. */\n    readonly classes?: readonly number[];\n    /** Truncate each detection's `classification.probabilities` to its top-k entries. */\n    readonly topK?: number;\n    /** Override the constructor's `raiseOnEmpty` setting for this call. */\n    readonly raiseOnEmpty?: boolean;\n}\n\n/**\n * Detector and classifier running as a single ONNX model.\n *\n * Everything the pipeline needs to know about itself — the resolution to\n * letterbox to, whether it wants the full-resolution image as well, whether its\n * classifier output still needs a softmax, the class names of both stages — was\n * written into the file at fusion time and is read back here. Nothing is\n * restated on the JavaScript side, so nothing can drift out of step with the\n * Python side that built it.\n *\n * @example\n * ```typescript\n * const pipeline = await DetectClassify.create(\"/models/pipeline.onnx\");\n * const result = (await pipeline.predict(\"/images/flock.jpg\"))[0];\n * for (const detection of result) {\n *   console.log(detection.name, detection.conf, detection.classification?.name);\n * }\n * ```\n */\nexport class DetectClassify extends VisionTask {\n    private constructor(\n        session: OrtSession,\n        private readonly _spec: FusionSpec,\n        private readonly _labels: readonly string[],\n        private readonly _names: Readonly<Record<number, string>>,\n        private readonly _classifierLabels: readonly string[],\n        private readonly _classifierNames: Readonly<Record<number, string>>,\n        private readonly _raiseOnEmpty: boolean,\n    ) {\n        super(session);\n    }\n\n    private _pipelineCache: LetterboxPipeline | null = null;\n\n    /**\n     * Run the model once on zero-filled inputs, paying one-time costs up front.\n     *\n     * Worth more here than on a single-stage task: a fused pipeline is two models\n     * plus the bridge in one graph, so the first inference compiles shaders for\n     * all of it. Calling this while a loading spinner is still up moves that cost\n     * somewhere the user is already waiting.\n     *\n     * @param runs How many warm-up inferences to run. One is enough for WASM;\n     *   WebGPU sometimes settles on the second.\n     */\n    async warmup(runs: number = 1): Promise<void> {\n        const [width, height] = this._spec.inputSize;\n        for (let i = 0; i < runs; i++) {\n            const feeds: Record<string, ort.Tensor> = {\n                [INPUT_IMAGE]: toFloat32Tensor(zeroTensorData(width, height), [\n                    1,\n                    3,\n                    height,\n                    width,\n                ]),\n            };\n            if (this._spec.needsSourceImage) {\n                feeds[INPUT_SOURCE] = toFloat32Tensor(zeroTensorData(width, height), [\n                    1,\n                    3,\n                    height,\n                    width,\n                ]);\n                feeds[INPUT_SCALE] = toFloat32Tensor(new Float32Array([1]), [1]);\n                feeds[INPUT_PAD] = toFloat32Tensor(new Float32Array([0, 0]), [2]);\n            }\n            await this._session.run(feeds);\n        }\n    }\n\n    /**\n     * The fused preprocessing pipeline, built on first use.\n     *\n     * Lazily, because constructing it allocates canvases: a pipeline built in an\n     * environment without a canvas implementation stays constructible, and only\n     * fails if it is actually asked to preprocess something.\n     */\n    private get _pipeline(): LetterboxPipeline {\n        if (this._pipelineCache === null) {\n            const [width, height] = this._spec.inputSize;\n            this._pipelineCache = new LetterboxPipeline(width, height);\n        }\n        return this._pipelineCache;\n    }\n\n    /**\n     * Load a fused pipeline and resolve both label spaces.\n     *\n     * @param model The fused `.onnx` — a URL, an `ArrayBuffer`, or bytes.\n     * @param options Label overrides plus the usual session options.\n     * @throws {@link FusionError} when the model carries no pipeline metadata,\n     *   i.e. it is a plain detector or classifier rather than something\n     *   `ort_vision_sdk.compose` produced.\n     */\n    static async create(\n        model: ModelSource,\n        options: DetectClassifyOptions = {},\n    ): Promise<DetectClassify> {\n        const session = await OrtSession.create(model, options);\n        const spec = readFusionSpec(session.metadata);\n        if (spec === null) {\n            throw new FusionError(\n                \"This model carries no fused-pipeline metadata, so DetectClassify cannot tell how to \" +\n                    \"drive it. Build one with ort_vision_sdk.compose.fuse_detect_classify, or load a \" +\n                    \"plain model with Detector/Classifier instead.\",\n            );\n        }\n\n        const labels = resolveLabels(options.labels ?? spec.detectorNames ?? \"coco\");\n        const classifierLabels = resolveLabels(options.classifierLabels ?? spec.classifierNames, {\n            numClasses: classifierClasses(session) ?? undefined,\n        });\n        return new DetectClassify(\n            session,\n            spec,\n            labels,\n            indexNames(labels),\n            classifierLabels,\n            indexNames(classifierLabels),\n            options.raiseOnEmpty ?? false,\n        );\n    }\n\n    /** The pipeline configuration recorded in the model at fusion time. */\n    get spec(): FusionSpec {\n        return this._spec;\n    }\n\n    /** The `[width, height]` the detection stage runs at. */\n    get inputSize(): readonly [number, number] {\n        return this._spec.inputSize;\n    }\n\n    /** Detector class labels indexed by class id. */\n    get labels(): readonly string[] {\n        return this._labels;\n    }\n\n    /** Detector class id → class name (matches Ultralytics' `model.names`). */\n    get names(): Readonly<Record<number, string>> {\n        return this._names;\n    }\n\n    /** Classifier class labels indexed by class id. */\n    get classifierLabels(): readonly string[] {\n        return this._classifierLabels;\n    }\n\n    /** Classifier class id → class name. */\n    get classifierNames(): Readonly<Record<number, string>> {\n        return this._classifierNames;\n    }\n\n    /**\n     * Alias for {@link predict} — call the pipeline like a torch `nn.Module`.\n     *\n     * Use as `pipeline.call(img)` since JavaScript class instances are not\n     * callable; for direct invocation, prefer `pipeline.predict(img)`.\n     */\n    async call(\n        image: ImageInput,\n        options: DetectClassifyPredictOptions = {},\n    ): Promise<DetectClassifyResults[]> {\n        return this.predict(image, options);\n    }\n\n    /**\n     * Run the pipeline on a single image.\n     *\n     * The returned envelope carries a {@link Speed} breakdown in `speed`. Its\n     * `inference` figure covers detection *and* classification, since the\n     * pipeline runs them as one graph and no boundary between them is observable\n     * from outside.\n     */\n    async predict(\n        image: ImageInput,\n        options: DetectClassifyPredictOptions = {},\n    ): Promise<DetectClassifyResults[]> {\n        const timer = new SpeedTimer();\n        const path = typeof image === \"string\" ? image : null;\n        const original = await loadImage(image);\n        timer.stage(\"load\");\n        const { feeds, scale, padLeft, padTop } = this._preprocess(original);\n        timer.stage(\"preprocess\");\n        const outputs = await this._session.run(feeds);\n        this._pipeline.release();\n        timer.stage(\"inference\");\n\n        const probsTensor = output(outputs, OUTPUT_PROBS);\n        const boxes = floats(outputs, OUTPUT_BOXES);\n        const scores = floats(outputs, OUTPUT_SCORES);\n        const classes = integers(outputs, OUTPUT_CLASSES);\n        const probs = probsTensor.data as Float32Array;\n        const reported = integers(outputs, OUTPUT_NUM_DETECTIONS)[0] ?? 0;\n        const rows = Math.min(reported, Math.floor(boxes.length / 4));\n        const classCount = probsTensor.dims[probsTensor.dims.length - 1] ?? 0;\n\n        const allowed = options.classes === undefined ? null : new Set(options.classes);\n        const floor = options.confThreshold ?? 0;\n        const detections: DetectionResult[] = [];\n        for (let row = 0; row < rows; row++) {\n            const classId = classes[row] ?? 0;\n            const confidence = scores[row] ?? 0;\n            if (confidence < floor || (allowed !== null && !allowed.has(classId))) continue;\n\n            const bbox = this._toOriginal(boxes, row, { scale, padLeft, padTop, original });\n            const cropped = crop(original, bbox);\n            detections.push(\n                detection(\n                    classId,\n                    this._names[classId] ?? `class_${classId}`,\n                    confidence,\n                    bbox,\n                    cropped,\n                    this._classify(\n                        probs.subarray(row * classCount, (row + 1) * classCount),\n                        cropped,\n                        options.topK,\n                    ),\n                ),\n            );\n        }\n\n        requireDetections(detections.length, {\n            raiseOnEmpty: options.raiseOnEmpty ?? this._raiseOnEmpty,\n            confThreshold: Math.max(floor, this._spec.confThreshold),\n            classes: options.classes,\n            path,\n        });\n\n        const origShape: readonly [number, number] = [original.height, original.width];\n        timer.stage(\"postprocess\");\n        return [\n            new DetectClassifyResults(\n                bulkBoxes(detections, origShape),\n                detections,\n                this._names,\n                this._classifierNames,\n                original,\n                origShape,\n                path,\n                timer.speed(),\n            ),\n        ];\n    }\n\n    /**\n     * Letterbox the image and build the graph's feeds.\n     *\n     * The detector input runs through {@link LetterboxPipeline}, which fuses the\n     * resize, the padding and the HWC-to-CHW float conversion into one\n     * `drawImage` plus one readback, and reuses its output buffer between frames.\n     * That buffer goes straight to ONNX Runtime, so `_pipeline.release()` must not\n     * be called until the run resolves.\n     *\n     * A pipeline fused with `cropSource: \"original\"` also takes the untouched\n     * image as a second input, plus the scale and padding of the letterbox — that\n     * is what lets the graph undo the letterbox transform internally and crop at\n     * native resolution instead of from the downscaled copy. That one is **not**\n     * letterboxed by definition, so it does not go through the fused path.\n     */\n    private _preprocess(image: RGBImage): {\n        feeds: Record<string, ort.Tensor>;\n        scale: number;\n        padLeft: number;\n        padTop: number;\n    } {\n        const [width, height] = this._spec.inputSize;\n        const boxed = this._pipeline.run(image);\n        const feeds: Record<string, ort.Tensor> = {\n            [INPUT_IMAGE]: toFloat32Tensor(boxed.data, [1, 3, height, width]),\n        };\n\n        if (this._spec.needsSourceImage) {\n            feeds[INPUT_SOURCE] = tensorOf(image);\n            feeds[INPUT_SCALE] = toFloat32Tensor(new Float32Array([boxed.scale]), [1]);\n            feeds[INPUT_PAD] = toFloat32Tensor(\n                new Float32Array([boxed.padLeft, boxed.padTop]),\n                [2],\n            );\n        }\n        return { feeds, scale: boxed.scale, padLeft: boxed.padLeft, padTop: boxed.padTop };\n    }\n\n    /**\n     * Map one letterboxed xyxy row back onto the original image.\n     *\n     * The graph always reports boxes in the detector's letterboxed pixel space,\n     * whichever crop source it was fused with, so both sources agree here.\n     */\n    private _toOriginal(\n        boxes: Float32Array,\n        row: number,\n        context: {\n            scale: number;\n            padLeft: number;\n            padTop: number;\n            original: RGBImage;\n        },\n    ): BoundingBox {\n        const { scale, padLeft, padTop, original } = context;\n        const at = (offset: number): number => boxes[row * 4 + offset] ?? 0;\n        const clampX = (value: number): number => Math.min(Math.max(value, 0), original.width);\n        const clampY = (value: number): number => Math.min(Math.max(value, 0), original.height);\n        return new BoundingBox(\n            clampX((at(0) - padLeft) / scale),\n            clampY((at(1) - padTop) / scale),\n            clampX((at(2) - padLeft) / scale),\n            clampY((at(3) - padTop) / scale),\n        );\n    }\n\n    /**\n     * Turn one row of the classifier output into a result object.\n     *\n     * @param row The output row for this detection.\n     * @param image The crop the row describes, carried so callers can display\n     *   what was classified.\n     * @param k Optional truncation of the probability list.\n     */\n    private _classify(\n        row: Float32Array,\n        image: RGBImage,\n        k: number | undefined,\n    ): ClassificationResult {\n        const scores = this._spec.applySoftmax ? softmax(row) : row;\n        const { indices, values } = topK(scores, k ?? null);\n        const probabilities: ClassProbability[] = [];\n        for (let i = 0; i < indices.length; i++) {\n            const classId = indices[i] ?? 0;\n            const probability = values[i] ?? 0;\n            const className = this._classifierLabels[classId] ?? `class_${classId}`;\n            probabilities.push({\n                classId,\n                className,\n                probability,\n                cls: classId,\n                name: className,\n                conf: probability,\n            });\n        }\n        const top = probabilities[0] ?? {\n            classId: 0,\n            className: \"class_0\",\n            probability: 0,\n            cls: 0,\n            name: \"class_0\",\n            conf: 0,\n        };\n        return {\n            classId: top.classId,\n            className: top.className,\n            confidence: top.probability,\n            cls: top.classId,\n            name: top.className,\n            conf: top.probability,\n            image,\n            probabilities,\n        };\n    }\n}\n\n/**\n * Read the classifier stage's class count off the `probs` output shape.\n *\n * @param session The loaded pipeline session.\n * @returns The class count, or `null` when the graph leaves that axis dynamic\n *   or declares no `probs` output — in which case label resolution falls back\n *   to whatever the fusion recorded.\n */\nfunction classifierClasses(session: OrtSession): number | null {\n    const index = session.outputNames.indexOf(OUTPUT_PROBS);\n    if (index < 0) return null;\n    const shape = session.outputShapes[index];\n    if (shape === undefined || shape.length === 0) return null;\n    return shape[shape.length - 1] ?? null;\n}\n\n/**\n * Build a class id → name record from an ordered label list.\n *\n * @param labels Labels indexed by class id.\n * @returns The equivalent record.\n */\nfunction indexNames(labels: readonly string[]): Readonly<Record<number, string>> {\n    const names: Record<number, string> = {};\n    for (let i = 0; i < labels.length; i++) names[i] = labels[i] as string;\n    return names;\n}\n\n/**\n * Fetch an output tensor by name.\n *\n * @param outputs The run's results.\n * @param name The output's name in the pipeline contract.\n * @returns The tensor.\n * @throws {@link FusionError} when the graph does not carry that output, which\n *   means the file is not a pipeline this version can drive.\n */\nfunction output(outputs: Record<string, ort.Tensor>, name: string): ort.Tensor {\n    const tensor = outputs[name];\n    if (tensor === undefined) {\n        throw new FusionError(`The fused pipeline is missing its '${name}' output.`);\n    }\n    return tensor;\n}\n\n/**\n * Fetch a float output by name.\n *\n * @param outputs The run's results.\n * @param name The output's name in the pipeline contract.\n * @returns The output's data.\n * @throws {@link FusionError} when the graph does not carry that output.\n */\nfunction floats(outputs: Record<string, ort.Tensor>, name: string): Float32Array {\n    return output(outputs, name).data as Float32Array;\n}\n\n/**\n * Fetch an integer output by name, normalizing ORT's 64-bit representation.\n *\n * ONNX Runtime Web returns `int64` tensors as `BigInt64Array`, whose values do\n * not compare or index like numbers. Class ids and detection counts are always\n * small, so widening them to `number` here is lossless and keeps every caller\n * free of `BigInt` handling.\n *\n * @param outputs The run's results.\n * @param name The output's name in the pipeline contract.\n * @returns The output's values as plain numbers.\n * @throws {@link FusionError} when the graph does not carry that output.\n */\nfunction integers(outputs: Record<string, ort.Tensor>, name: string): number[] {\n    const data = output(outputs, name).data as BigInt64Array | Int32Array | Float32Array;\n    const values: number[] = [];\n    for (let i = 0; i < data.length; i++) values.push(Number(data[i]));\n    return values;\n}\n\n/**\n * Convert an image to the `[1, 3, H, W]` float32 tensor the graph expects.\n *\n * @param image The image to convert.\n * @returns The batched CHW tensor, scaled to `[0, 1]`.\n */\nfunction tensorOf(image: RGBImage): ort.Tensor {\n    const chw = toCHW(toFloat32(image), image.width, image.height, 3);\n    return toFloat32Tensor(chw, [1, 3, image.height, image.width]);\n}\n\n/**\n * Cut the box region out of the original image.\n *\n * @param image The source image.\n * @param bbox The box, in original-image pixel coordinates.\n * @returns The cropped region, or a zero-sized image for a box with no area.\n */\nfunction crop(image: RGBImage, bbox: BoundingBox): RGBImage {\n    const [rawX1, rawY1, rawX2, rawY2] = bbox.asIntXyxy();\n    const x1 = Math.max(0, rawX1);\n    const y1 = Math.max(0, rawY1);\n    const x2 = Math.min(image.width, rawX2);\n    const y2 = Math.min(image.height, rawY2);\n    if (x2 <= x1 || y2 <= y1) return new RGBImage(new Uint8Array(0), 0, 0);\n\n    const width = x2 - x1;\n    const height = y2 - y1;\n    const out = new Uint8Array(width * height * 3);\n    for (let row = 0; row < height; row++) {\n        const offset = ((y1 + row) * image.width + x1) * 3;\n        out.set(image.data.subarray(offset, offset + width * 3), row * width * 3);\n    }\n    return new RGBImage(out, width, height);\n}\n\n/**\n * Assemble one detection, filling the Ultralytics-style aliases.\n *\n * @param classId Detector class index.\n * @param className Detector class name.\n * @param confidence Detection score.\n * @param bbox Box in original-image coordinates.\n * @param croppedImage The region the box covers.\n * @param classification What the classification stage said about that region.\n * @returns The detection object.\n */\nfunction detection(\n    classId: number,\n    className: string,\n    confidence: number,\n    bbox: BoundingBox,\n    croppedImage: RGBImage,\n    classification: ClassificationResult,\n): DetectionResult {\n    return {\n        classId,\n        className,\n        confidence,\n        bbox,\n        cls: classId,\n        name: className,\n        conf: confidence,\n        box: bbox,\n        croppedImage,\n        classification,\n    };\n}\n\n/**\n * Assemble the bulk-array `Boxes` view from per-instance detections.\n *\n * @param detections The surviving detections.\n * @param origShape `[height, width]` of the original image.\n * @returns The bulk view, empty when nothing survived.\n */\nfunction bulkBoxes(\n    detections: readonly DetectionResult[],\n    origShape: readonly [number, number],\n): Boxes {\n    const xyxy = new Float32Array(detections.length * 4);\n    const cls = new Int32Array(detections.length);\n    const conf = new Float32Array(detections.length);\n    detections.forEach((entry, index) => {\n        const [x1, y1, x2, y2] = entry.bbox.xyxy;\n        xyxy[index * 4] = x1;\n        xyxy[index * 4 + 1] = y1;\n        xyxy[index * 4 + 2] = x2;\n        xyxy[index * 4 + 3] = y2;\n        cls[index] = entry.classId;\n        conf[index] = entry.confidence;\n    });\n    return new Boxes(xyxy, cls, conf, origShape);\n}\n"],"mappings":"wYAqGA,IAAa,EAAb,MAAa,UAAuB,EAAA,UAAW,CAGtB,MACA,QACA,OACA,kBACA,iBACA,cAPrB,YACI,EACA,EACA,EACA,EACA,EACA,EACA,EACF,CACE,MAAM,CAAO,EAPI,KAAA,MAAA,EACA,KAAA,QAAA,EACA,KAAA,OAAA,EACA,KAAA,kBAAA,EACA,KAAA,iBAAA,EACA,KAAA,cAAA,CAGrB,CAEA,eAAmD,KAanD,MAAM,OAAO,EAAe,EAAkB,CAC1C,GAAM,CAAC,EAAO,GAAU,KAAK,MAAM,UACnC,IAAK,IAAI,EAAI,EAAG,EAAI,EAAM,IAAK,CAC3B,IAAM,EAAoC,EACrC,EAAA,aAAc,EAAA,gBAAgB,EAAA,eAAe,EAAO,CAAM,EAAG,CAC1D,EACA,EACA,EACA,CACJ,CAAC,CACL,EACI,KAAK,MAAM,mBACX,EAAM,EAAA,cAAgB,EAAA,gBAAgB,EAAA,eAAe,EAAO,CAAM,EAAG,CACjE,EACA,EACA,EACA,CACJ,CAAC,EACD,EAAM,EAAA,aAAe,EAAA,gBAAgB,IAAI,aAAa,CAAC,CAAC,CAAC,EAAG,CAAC,CAAC,CAAC,EAC/D,EAAM,EAAA,WAAa,EAAA,gBAAgB,IAAI,aAAa,CAAC,EAAG,CAAC,CAAC,EAAG,CAAC,CAAC,CAAC,GAEpE,MAAM,KAAK,SAAS,IAAI,CAAK,CACjC,CACJ,CASA,IAAY,WAA+B,CACvC,GAAI,KAAK,iBAAmB,KAAM,CAC9B,GAAM,CAAC,EAAO,GAAU,KAAK,MAAM,UACnC,KAAK,eAAiB,IAAI,EAAA,kBAAkB,EAAO,CAAM,CAC7D,CACA,OAAO,KAAK,cAChB,CAWA,aAAa,OACT,EACA,EAAiC,CAAC,EACX,CACvB,IAAM,EAAU,MAAM,EAAA,WAAW,OAAO,EAAO,CAAO,EAChD,EAAO,EAAA,eAAe,EAAQ,QAAQ,EAC5C,GAAI,IAAS,KACT,MAAM,IAAI,EAAA,YACN,mNAGJ,EAGJ,IAAM,EAAS,EAAA,cAAc,EAAQ,QAAU,EAAK,eAAiB,MAAM,EACrE,EAAmB,EAAA,cAAc,EAAQ,kBAAoB,EAAK,gBAAiB,CACrF,WAAY,EAAkB,CAAO,GAAK,IAAA,EAC9C,CAAC,EACD,OAAO,IAAI,EACP,EACA,EACA,EACA,EAAW,CAAM,EACjB,EACA,EAAW,CAAgB,EAC3B,EAAQ,cAAgB,EAC5B,CACJ,CAGA,IAAI,MAAmB,CACnB,OAAO,KAAK,KAChB,CAGA,IAAI,WAAuC,CACvC,OAAO,KAAK,MAAM,SACtB,CAGA,IAAI,QAA4B,CAC5B,OAAO,KAAK,OAChB,CAGA,IAAI,OAA0C,CAC1C,OAAO,KAAK,MAChB,CAGA,IAAI,kBAAsC,CACtC,OAAO,KAAK,iBAChB,CAGA,IAAI,iBAAoD,CACpD,OAAO,KAAK,gBAChB,CAQA,MAAM,KACF,EACA,EAAwC,CAAC,EACT,CAChC,OAAO,KAAK,QAAQ,EAAO,CAAO,CACtC,CAUA,MAAM,QACF,EACA,EAAwC,CAAC,EACT,CAChC,IAAM,EAAQ,IAAI,EAAA,WACZ,EAAO,OAAO,GAAU,SAAW,EAAQ,KAC3C,EAAW,MAAM,EAAA,UAAU,CAAK,EACtC,EAAM,MAAM,MAAM,EAClB,GAAM,CAAE,QAAO,QAAO,UAAS,UAAW,KAAK,YAAY,CAAQ,EACnE,EAAM,MAAM,YAAY,EACxB,IAAM,EAAU,MAAM,KAAK,SAAS,IAAI,CAAK,EAC7C,KAAK,UAAU,QAAQ,EACvB,EAAM,MAAM,WAAW,EAEvB,IAAM,EAAc,EAAO,EAAS,EAAA,YAAY,EAC1C,EAAQ,EAAO,EAAS,EAAA,YAAY,EACpC,EAAS,EAAO,EAAS,EAAA,aAAa,EACtC,EAAU,EAAS,EAAS,EAAA,cAAc,EAC1C,EAAQ,EAAY,KACpB,EAAW,EAAS,EAAA,gBAA8B,CAAC,CAAC,IAAM,EAC1D,EAAO,KAAK,IAAI,EAAU,KAAK,MAAM,EAAM,OAAS,CAAC,CAAC,EACtD,EAAa,EAAY,KAAK,EAAY,KAAK,OAAS,IAAM,EAE9D,EAAU,EAAQ,UAAY,IAAA,GAAY,KAAO,IAAI,IAAI,EAAQ,OAAO,EACxE,EAAQ,EAAQ,eAAiB,EACjC,EAAgC,CAAC,EACvC,IAAK,IAAI,EAAM,EAAG,EAAM,EAAM,IAAO,CACjC,IAAM,EAAU,EAAQ,IAAQ,EAC1B,EAAa,EAAO,IAAQ,EAClC,GAAI,EAAa,GAAU,IAAY,MAAQ,CAAC,EAAQ,IAAI,CAAO,EAAI,SAEvE,IAAM,EAAO,KAAK,YAAY,EAAO,EAAK,CAAE,QAAO,UAAS,SAAQ,UAAS,CAAC,EACxE,EAAU,EAAK,EAAU,CAAI,EACnC,EAAW,KACP,EACI,EACA,KAAK,OAAO,IAAY,SAAS,IACjC,EACA,EACA,EACA,KAAK,UACD,EAAM,SAAS,EAAM,GAAa,EAAM,GAAK,CAAU,EACvD,EACA,EAAQ,IACZ,CACJ,CACJ,CACJ,CAEA,EAAA,kBAAkB,EAAW,OAAQ,CACjC,aAAc,EAAQ,cAAgB,KAAK,cAC3C,cAAe,KAAK,IAAI,EAAO,KAAK,MAAM,aAAa,EACvD,QAAS,EAAQ,QACjB,MACJ,CAAC,EAED,IAAM,EAAuC,CAAC,EAAS,OAAQ,EAAS,KAAK,EAE7E,OADA,EAAM,MAAM,aAAa,EAClB,CACH,IAAI,EAAA,sBACA,EAAU,EAAY,CAAS,EAC/B,EACA,KAAK,OACL,KAAK,iBACL,EACA,EACA,EACA,EAAM,MAAM,CAChB,CACJ,CACJ,CAiBA,YAAoB,EAKlB,CACE,GAAM,CAAC,EAAO,GAAU,KAAK,MAAM,UAC7B,EAAQ,KAAK,UAAU,IAAI,CAAK,EAChC,EAAoC,EACrC,EAAA,aAAc,EAAA,gBAAgB,EAAM,KAAM,CAAC,EAAG,EAAG,EAAQ,CAAK,CAAC,CACpE,EAUA,OARI,KAAK,MAAM,mBACX,EAAM,EAAA,cAAgB,EAAS,CAAK,EACpC,EAAM,EAAA,aAAe,EAAA,gBAAgB,IAAI,aAAa,CAAC,EAAM,KAAK,CAAC,EAAG,CAAC,CAAC,CAAC,EACzE,EAAM,EAAA,WAAa,EAAA,gBACf,IAAI,aAAa,CAAC,EAAM,QAAS,EAAM,MAAM,CAAC,EAC9C,CAAC,CAAC,CACN,GAEG,CAAE,QAAO,MAAO,EAAM,MAAO,QAAS,EAAM,QAAS,OAAQ,EAAM,MAAO,CACrF,CAQA,YACI,EACA,EACA,EAMW,CACX,GAAM,CAAE,QAAO,UAAS,SAAQ,YAAa,EACvC,EAAM,GAA2B,EAAM,EAAM,EAAI,IAAW,EAC5D,EAAU,GAA0B,KAAK,IAAI,KAAK,IAAI,EAAO,CAAC,EAAG,EAAS,KAAK,EAC/E,EAAU,GAA0B,KAAK,IAAI,KAAK,IAAI,EAAO,CAAC,EAAG,EAAS,MAAM,EACtF,OAAO,IAAI,EAAA,YACP,GAAQ,EAAG,CAAC,EAAI,GAAW,CAAK,EAChC,GAAQ,EAAG,CAAC,EAAI,GAAU,CAAK,EAC/B,GAAQ,EAAG,CAAC,EAAI,GAAW,CAAK,EAChC,GAAQ,EAAG,CAAC,EAAI,GAAU,CAAK,CACnC,CACJ,CAUA,UACI,EACA,EACA,EACoB,CACpB,IAAM,EAAS,KAAK,MAAM,aAAe,EAAA,QAAQ,CAAG,EAAI,EAClD,CAAE,UAAS,UAAW,EAAA,KAAK,EAAQ,GAAK,IAAI,EAC5C,EAAoC,CAAC,EAC3C,IAAK,IAAI,EAAI,EAAG,EAAI,EAAQ,OAAQ,IAAK,CACrC,IAAM,EAAU,EAAQ,IAAM,EACxB,EAAc,EAAO,IAAM,EAC3B,EAAY,KAAK,kBAAkB,IAAY,SAAS,IAC9D,EAAc,KAAK,CACf,UACA,YACA,cACA,IAAK,EACL,KAAM,EACN,KAAM,CACV,CAAC,CACL,CACA,IAAM,EAAM,EAAc,IAAM,CAC5B,QAAS,EACT,UAAW,UACX,YAAa,EACb,IAAK,EACL,KAAM,UACN,KAAM,CACV,EACA,MAAO,CACH,QAAS,EAAI,QACb,UAAW,EAAI,UACf,WAAY,EAAI,YAChB,IAAK,EAAI,QACT,KAAM,EAAI,UACV,KAAM,EAAI,YACV,QACA,eACJ,CACJ,CACJ,EAUA,SAAS,EAAkB,EAAoC,CAC3D,IAAM,EAAQ,EAAQ,YAAY,QAAQ,EAAA,YAAY,EACtD,GAAI,EAAQ,EAAG,OAAO,KACtB,IAAM,EAAQ,EAAQ,aAAa,GAEnC,OADI,IAAU,IAAA,IAAa,EAAM,SAAW,EAAU,KAC/C,EAAM,EAAM,OAAS,IAAM,IACtC,CAQA,SAAS,EAAW,EAA6D,CAC7E,IAAM,EAAgC,CAAC,EACvC,IAAK,IAAI,EAAI,EAAG,EAAI,EAAO,OAAQ,IAAK,EAAM,GAAK,EAAO,GAC1D,OAAO,CACX,CAWA,SAAS,EAAO,EAAqC,EAA0B,CAC3E,IAAM,EAAS,EAAQ,GACvB,GAAI,IAAW,IAAA,GACX,MAAM,IAAI,EAAA,YAAY,sCAAsC,EAAK,UAAU,EAE/E,OAAO,CACX,CAUA,SAAS,EAAO,EAAqC,EAA4B,CAC7E,OAAO,EAAO,EAAS,CAAI,CAAC,CAAC,IACjC,CAeA,SAAS,EAAS,EAAqC,EAAwB,CAC3E,IAAM,EAAO,EAAO,EAAS,CAAI,CAAC,CAAC,KAC7B,EAAmB,CAAC,EAC1B,IAAK,IAAI,EAAI,EAAG,EAAI,EAAK,OAAQ,IAAK,EAAO,KAAK,OAAO,EAAK,EAAE,CAAC,EACjE,OAAO,CACX,CAQA,SAAS,EAAS,EAA6B,CAC3C,IAAM,EAAM,EAAA,MAAM,EAAA,UAAU,CAAK,EAAG,EAAM,MAAO,EAAM,OAAQ,CAAC,EAChE,OAAO,EAAA,gBAAgB,EAAK,CAAC,EAAG,EAAG,EAAM,OAAQ,EAAM,KAAK,CAAC,CACjE,CASA,SAAS,EAAK,EAAiB,EAA6B,CACxD,GAAM,CAAC,EAAO,EAAO,EAAO,GAAS,EAAK,UAAU,EAC9C,EAAK,KAAK,IAAI,EAAG,CAAK,EACtB,EAAK,KAAK,IAAI,EAAG,CAAK,EACtB,EAAK,KAAK,IAAI,EAAM,MAAO,CAAK,EAChC,EAAK,KAAK,IAAI,EAAM,OAAQ,CAAK,EACvC,GAAI,GAAM,GAAM,GAAM,EAAI,OAAO,IAAI,EAAA,SAAS,IAAI,WAAe,EAAG,CAAC,EAErE,IAAM,EAAQ,EAAK,EACb,EAAS,EAAK,EACd,EAAM,IAAI,WAAW,EAAQ,EAAS,CAAC,EAC7C,IAAK,IAAI,EAAM,EAAG,EAAM,EAAQ,IAAO,CACnC,IAAM,IAAW,EAAK,GAAO,EAAM,MAAQ,GAAM,EACjD,EAAI,IAAI,EAAM,KAAK,SAAS,EAAQ,EAAS,EAAQ,CAAC,EAAG,EAAM,EAAQ,CAAC,CAC5E,CACA,OAAO,IAAI,EAAA,SAAS,EAAK,EAAO,CAAM,CAC1C,CAaA,SAAS,EACL,EACA,EACA,EACA,EACA,EACA,EACe,CACf,MAAO,CACH,UACA,YACA,aACA,OACA,IAAK,EACL,KAAM,EACN,KAAM,EACN,IAAK,EACL,eACA,gBACJ,CACJ,CASA,SAAS,EACL,EACA,EACK,CACL,IAAM,EAAO,IAAI,aAAa,EAAW,OAAS,CAAC,EAC7C,EAAM,IAAI,WAAW,EAAW,MAAM,EACtC,EAAO,IAAI,aAAa,EAAW,MAAM,EAU/C,OATA,EAAW,SAAS,EAAO,IAAU,CACjC,GAAM,CAAC,EAAI,EAAI,EAAI,GAAM,EAAM,KAAK,KACpC,EAAK,EAAQ,GAAK,EAClB,EAAK,EAAQ,EAAI,GAAK,EACtB,EAAK,EAAQ,EAAI,GAAK,EACtB,EAAK,EAAQ,EAAI,GAAK,EACtB,EAAI,GAAS,EAAM,QACnB,EAAK,GAAS,EAAM,UACxB,CAAC,EACM,IAAI,EAAA,MAAM,EAAM,EAAK,EAAM,CAAS,CAC/C"}
+{"version":3,"file":"detectClassify.cjs","names":[],"sources":["../../../src/vision/tasks/detectClassify.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Run a fused detect→classify pipeline in the browser.\n *\n * The file this loads was built by the Python SDK's `ort_vision_sdk.compose`,\n * and it already contains both models plus the crop-and-resize bridge between\n * them. That matters far more in a tab than on a server: two models mean two\n * `.onnx` downloads, two WASM/WebGPU session initializations, and a per-crop\n * round trip through JavaScript to slice, resize and restack the regions before\n * the second model can see them. A fused pipeline has one download, one session\n * and no round trip — the crops are produced and consumed inside the graph.\n */\n\nimport type * as ort from \"onnxruntime-web\";\n\nimport { FusionError } from \"../core/exceptions\";\nimport { type ModelSource, type OrtSessionOptions, OrtSession } from \"../core/session\";\nimport { SpeedTimer } from \"../core/timing\";\nimport {\n    INPUT_IMAGE,\n    INPUT_PAD,\n    INPUT_SCALE,\n    INPUT_SOURCE,\n    OUTPUT_BOXES,\n    OUTPUT_CLASSES,\n    OUTPUT_NUM_DETECTIONS,\n    OUTPUT_MASKS,\n    OUTPUT_PROBS,\n    OUTPUT_SCORES,\n    type FusionSpec,\n    readFusionSpec,\n} from \"../fusion\";\nimport { type ImageInput, loadImageSource } from \"../io/image\";\nimport { type LabelSpec, resolveLabels } from \"../labels\";\nimport { softmax, topK } from \"../postprocess/classification\";\nimport { asFloat32Array } from \"../core/dtypes\";\nimport { toCHW, toFloat32, toFloat32Tensor } from \"../preprocess/image\";\nimport { LetterboxPipeline, zeroTensorData } from \"../preprocess/pipeline\";\nimport { Boxes, DetectClassifyResults } from \"../results\";\nimport type { RGBImage } from \"../types\";\nimport {\n    BoundingBox,\n    Mask,\n    type ClassProbability,\n    type ClassificationResult,\n    type DetectionResult,\n} from \"../types\";\nimport { VisionTask, cropToBox, requireDetections } from \"./base\";\n\nexport interface DetectClassifyOptions extends OrtSessionOptions {\n    /**\n     * Class label spec for the **detection** stage — see {@link resolveLabels}.\n     * Defaults to the names recorded at fusion time, falling back to the COCO\n     * 80-class preset when the fusion recorded none.\n     */\n    readonly labels?: LabelSpec;\n    /**\n     * Class label spec for the **classification** stage. Defaults to the recorded\n     * names, falling back to generated `class_<id>` names.\n     */\n    readonly classifierLabels?: LabelSpec;\n    /**\n     * If `true`, a run that finds nothing throws {@link NoDetectionsError}\n     * instead of returning an empty envelope. Default `false`, because looking\n     * and finding nothing is a successful inference. Turn it on when an empty\n     * result means the surrounding pipeline should stop rather than carry on with\n     * zero rows. Can be overridden per `predict` call.\n     */\n    readonly raiseOnEmpty?: boolean;\n}\n\nexport interface DetectClassifyPredictOptions {\n    /**\n     * Drop detections scoring below this. The graph's own NMS threshold was fixed\n     * at fusion time and cannot be lowered here — this only filters further.\n     */\n    readonly confThreshold?: number;\n    /** If set, keep only detections whose detector `classId` is in this list. */\n    readonly classes?: readonly number[];\n    /** Truncate each detection's `classification.probabilities` to its top-k entries. */\n    readonly topK?: number;\n    /** Override the constructor's `raiseOnEmpty` setting for this call. */\n    readonly raiseOnEmpty?: boolean;\n}\n\n/**\n * Detector and classifier running as a single ONNX model.\n *\n * Everything the pipeline needs to know about itself — the resolution to\n * letterbox to, whether it wants the full-resolution image as well, whether its\n * classifier output still needs a softmax, the class names of both stages — was\n * written into the file at fusion time and is read back here. Nothing is\n * restated on the JavaScript side, so nothing can drift out of step with the\n * Python side that built it.\n *\n * @example\n * ```typescript\n * const pipeline = await DetectClassify.create(\"/models/pipeline.onnx\");\n * const result = (await pipeline.predict(\"/images/flock.jpg\"))[0];\n * for (const detection of result) {\n *   console.log(detection.name, detection.conf, detection.classification?.name);\n * }\n * ```\n */\nexport class DetectClassify extends VisionTask {\n    private constructor(\n        session: OrtSession,\n        private readonly _spec: FusionSpec,\n        private readonly _labels: readonly string[],\n        private readonly _names: Readonly<Record<number, string>>,\n        private readonly _classifierLabels: readonly string[],\n        private readonly _classifierNames: Readonly<Record<number, string>>,\n        private readonly _raiseOnEmpty: boolean,\n    ) {\n        super(session);\n    }\n\n    private _pipelineCache: LetterboxPipeline | null = null;\n\n    /**\n     * Run the model once on zero-filled inputs, paying one-time costs up front.\n     *\n     * Worth more here than on a single-stage task: a fused pipeline is two models\n     * plus the bridge in one graph, so the first inference compiles shaders for\n     * all of it. Calling this while a loading spinner is still up moves that cost\n     * somewhere the user is already waiting.\n     *\n     * @param runs How many warm-up inferences to run. One is enough for WASM;\n     *   WebGPU sometimes settles on the second.\n     */\n    async warmup(runs: number = 1): Promise<void> {\n        const [width, height] = this._spec.inputSize;\n        for (let i = 0; i < runs; i++) {\n            const feeds: Record<string, ort.Tensor> = {\n                [INPUT_IMAGE]: toFloat32Tensor(zeroTensorData(width, height), [\n                    1,\n                    3,\n                    height,\n                    width,\n                ]),\n            };\n            if (this._spec.needsSourceImage) {\n                feeds[INPUT_SOURCE] = toFloat32Tensor(zeroTensorData(width, height), [\n                    1,\n                    3,\n                    height,\n                    width,\n                ]);\n                feeds[INPUT_SCALE] = toFloat32Tensor(new Float32Array([1]), [1]);\n                feeds[INPUT_PAD] = toFloat32Tensor(new Float32Array([0, 0]), [2]);\n            }\n            await this._session.run(feeds);\n        }\n    }\n\n    /**\n     * The fused preprocessing pipeline, built on first use.\n     *\n     * Lazily, because constructing it allocates canvases: a pipeline built in an\n     * environment without a canvas implementation stays constructible, and only\n     * fails if it is actually asked to preprocess something.\n     */\n    private get _pipeline(): LetterboxPipeline {\n        if (this._pipelineCache === null) {\n            const [width, height] = this._spec.inputSize;\n            this._pipelineCache = new LetterboxPipeline(width, height);\n        }\n        return this._pipelineCache;\n    }\n\n    /**\n     * Load a fused pipeline and resolve both label spaces.\n     *\n     * @param model The fused `.onnx` — a URL, an `ArrayBuffer`, or bytes.\n     * @param options Label overrides plus the usual session options.\n     * @throws {@link FusionError} when the model carries no pipeline metadata,\n     *   i.e. it is a plain detector or classifier rather than something\n     *   `ort_vision_sdk.compose` produced.\n     */\n    static async create(\n        model: ModelSource,\n        options: DetectClassifyOptions = {},\n    ): Promise<DetectClassify> {\n        const session = await OrtSession.create(model, options);\n        const spec = readFusionSpec(session.metadata);\n        if (spec === null) {\n            throw new FusionError(\n                \"This model carries no fused-pipeline metadata, so DetectClassify cannot tell how to \" +\n                    \"drive it. Build one with ort_vision_sdk.compose.fuse_detect_classify, or load a \" +\n                    \"plain model with Detector/Classifier instead.\",\n            );\n        }\n\n        const labels = resolveLabels(options.labels ?? spec.detectorNames ?? \"coco\");\n        const classifierLabels = resolveLabels(options.classifierLabels ?? spec.classifierNames, {\n            numClasses: classifierClasses(session) ?? undefined,\n        });\n        return new DetectClassify(\n            session,\n            spec,\n            labels,\n            indexNames(labels),\n            classifierLabels,\n            indexNames(classifierLabels),\n            options.raiseOnEmpty ?? false,\n        );\n    }\n\n    /** The pipeline configuration recorded in the model at fusion time. */\n    get spec(): FusionSpec {\n        return this._spec;\n    }\n\n    /** The `[width, height]` the detection stage runs at. */\n    get inputSize(): readonly [number, number] {\n        return this._spec.inputSize;\n    }\n\n    /** Detector class labels indexed by class id. */\n    get labels(): readonly string[] {\n        return this._labels;\n    }\n\n    /** Detector class id → class name (matches Ultralytics' `model.names`). */\n    get names(): Readonly<Record<number, string>> {\n        return this._names;\n    }\n\n    /** Classifier class labels indexed by class id. */\n    get classifierLabels(): readonly string[] {\n        return this._classifierLabels;\n    }\n\n    /** Classifier class id → class name. */\n    get classifierNames(): Readonly<Record<number, string>> {\n        return this._classifierNames;\n    }\n\n    /**\n     * Alias for {@link predict} — call the pipeline like a torch `nn.Module`.\n     *\n     * Use as `pipeline.call(img)` since JavaScript class instances are not\n     * callable; for direct invocation, prefer `pipeline.predict(img)`.\n     */\n    async call(\n        image: ImageInput,\n        options: DetectClassifyPredictOptions = {},\n    ): Promise<DetectClassifyResults[]> {\n        return this.predict(image, options);\n    }\n\n    /**\n     * Run the pipeline on a single image.\n     *\n     * The returned envelope carries a {@link Speed} breakdown in `speed`. Its\n     * `inference` figure covers detection *and* classification, since the\n     * pipeline runs them as one graph and no boundary between them is observable\n     * from outside.\n     */\n    async predict(\n        image: ImageInput,\n        options: DetectClassifyPredictOptions = {},\n    ): Promise<DetectClassifyResults[]> {\n        const timer = new SpeedTimer();\n        const path = typeof image === \"string\" ? image : null;\n        const { image: original, canvas } = await loadImageSource(image);\n        timer.stage(\"load\");\n        const { feeds, scale, padLeft, padTop, reused } = this._preprocess(original, canvas);\n        timer.stage(\"preprocess\");\n        let outputs: Record<string, ort.Tensor>;\n        try {\n            outputs = await this._session.run(feeds);\n        } finally {\n            this._pipeline.release({ reused });\n        }\n        timer.stage(\"inference\");\n\n        const probsTensor = output(outputs, OUTPUT_PROBS);\n        const boxes = floats(outputs, OUTPUT_BOXES);\n        const scores = floats(outputs, OUTPUT_SCORES);\n        const classes = integers(outputs, OUTPUT_CLASSES);\n        const probs = asFloat32Array(probsTensor.data);\n        const reported = integers(outputs, OUTPUT_NUM_DETECTIONS)[0] ?? 0;\n        const masks = this._spec.hasMasks ? floats(outputs, OUTPUT_MASKS) : null;\n        const [cropWidth, cropHeight] = this._spec.cropSize;\n        const rows = Math.min(reported, Math.floor(boxes.length / 4));\n        const classCount = probsTensor.dims[probsTensor.dims.length - 1] ?? 0;\n\n        const allowed = options.classes === undefined ? null : new Set(options.classes);\n        const floor = options.confThreshold ?? 0;\n        const detections: DetectionResult[] = [];\n        for (let row = 0; row < rows; row++) {\n            const classId = classes[row] ?? 0;\n            const confidence = scores[row] ?? 0;\n            if (confidence < floor || (allowed !== null && !allowed.has(classId))) continue;\n\n            const bbox = this._toOriginal(boxes, row, { scale, padLeft, padTop, original });\n            const cropped = cropToBox(original, bbox);\n            detections.push(\n                detection(\n                    classId,\n                    this._names[classId] ?? `class_${classId}`,\n                    confidence,\n                    bbox,\n                    cropped,\n                    this._classify(\n                        probs.subarray(row * classCount, (row + 1) * classCount),\n                        cropped,\n                        options.topK,\n                    ),\n                    masks === null ? null : maskToBox(masks, row, cropWidth, cropHeight, cropped),\n                ),\n            );\n        }\n\n        requireDetections(detections.length, {\n            raiseOnEmpty: options.raiseOnEmpty ?? this._raiseOnEmpty,\n            confThreshold: Math.max(floor, this._spec.confThreshold),\n            classes: options.classes,\n            path,\n        });\n\n        const origShape: readonly [number, number] = [original.height, original.width];\n        timer.stage(\"postprocess\");\n        return [\n            new DetectClassifyResults(\n                bulkBoxes(detections, origShape),\n                detections,\n                this._names,\n                this._classifierNames,\n                original,\n                origShape,\n                path,\n                timer.speed(),\n            ),\n        ];\n    }\n\n    /**\n     * Letterbox the image and build the graph's feeds.\n     *\n     * The detector input runs through {@link LetterboxPipeline}, which fuses the\n     * resize, the padding and the HWC-to-CHW float conversion into one\n     * `drawImage` plus one readback, and reuses its output buffer between frames.\n     * That buffer goes straight to ONNX Runtime, so `_pipeline.release()` must not\n     * be called until the run resolves.\n     *\n     * A pipeline fused with `cropSource: \"original\"` also takes the untouched\n     * image as a second input, plus the scale and padding of the letterbox — that\n     * is what lets the graph undo the letterbox transform internally and crop at\n     * native resolution instead of from the downscaled copy. That one is **not**\n     * letterboxed by definition, so it does not go through the fused path.\n     *\n     * @param image The decoded input.\n     * @param canvas The opaque canvas it was decoded on, when there is one — see\n     *   {@link LetterboxPipeline.run}.\n     */\n    private _preprocess(\n        image: RGBImage,\n        canvas: CanvasImageSource | null,\n    ): {\n        feeds: Record<string, ort.Tensor>;\n        scale: number;\n        padLeft: number;\n        padTop: number;\n        reused: boolean;\n    } {\n        const [width, height] = this._spec.inputSize;\n        const boxed = this._pipeline.run(image, canvas);\n        const feeds: Record<string, ort.Tensor> = {\n            [INPUT_IMAGE]: toFloat32Tensor(boxed.data, [1, 3, height, width]),\n        };\n\n        if (this._spec.needsSourceImage) {\n            feeds[INPUT_SOURCE] = tensorOf(image);\n            feeds[INPUT_SCALE] = toFloat32Tensor(new Float32Array([boxed.scale]), [1]);\n            feeds[INPUT_PAD] = toFloat32Tensor(\n                new Float32Array([boxed.padLeft, boxed.padTop]),\n                [2],\n            );\n        }\n        return {\n            feeds,\n            scale: boxed.scale,\n            padLeft: boxed.padLeft,\n            padTop: boxed.padTop,\n            reused: boxed.reused,\n        };\n    }\n\n    /**\n     * Map one letterboxed xyxy row back onto the original image.\n     *\n     * The graph always reports boxes in the detector's letterboxed pixel space,\n     * whichever crop source it was fused with, so both sources agree here.\n     */\n    private _toOriginal(\n        boxes: Float32Array,\n        row: number,\n        context: {\n            scale: number;\n            padLeft: number;\n            padTop: number;\n            original: RGBImage;\n        },\n    ): BoundingBox {\n        const { scale, padLeft, padTop, original } = context;\n        const at = (offset: number): number => boxes[row * 4 + offset] ?? 0;\n        const clampX = (value: number): number => Math.min(Math.max(value, 0), original.width);\n        const clampY = (value: number): number => Math.min(Math.max(value, 0), original.height);\n        return new BoundingBox(\n            clampX((at(0) - padLeft) / scale),\n            clampY((at(1) - padTop) / scale),\n            clampX((at(2) - padLeft) / scale),\n            clampY((at(3) - padTop) / scale),\n        );\n    }\n\n    /**\n     * Turn one row of the classifier output into a result object.\n     *\n     * @param row The output row for this detection.\n     * @param image The crop the row describes, carried so callers can display\n     *   what was classified.\n     * @param k Optional truncation of the probability list.\n     */\n    private _classify(\n        row: Float32Array,\n        image: RGBImage,\n        k: number | undefined,\n    ): ClassificationResult {\n        const scores = this._spec.applySoftmax ? softmax(row) : row;\n        const { indices, values } = topK(scores, k ?? null);\n        const probabilities: ClassProbability[] = [];\n        for (let i = 0; i < indices.length; i++) {\n            const classId = indices[i] ?? 0;\n            const probability = values[i] ?? 0;\n            const className = this._classifierLabels[classId] ?? `class_${classId}`;\n            probabilities.push({\n                classId,\n                className,\n                probability,\n                cls: classId,\n                name: className,\n                conf: probability,\n            });\n        }\n        const top = probabilities[0] ?? {\n            classId: 0,\n            className: \"class_0\",\n            probability: 0,\n            cls: 0,\n            name: \"class_0\",\n            conf: 0,\n        };\n        return {\n            classId: top.classId,\n            className: top.className,\n            confidence: top.probability,\n            cls: top.classId,\n            name: top.className,\n            conf: top.probability,\n            image,\n            probabilities,\n        };\n    }\n}\n\n/**\n * Read the classifier stage's class count off the `probs` output shape.\n *\n * @param session The loaded pipeline session.\n * @returns The class count, or `null` when the graph leaves that axis dynamic\n *   or declares no `probs` output — in which case label resolution falls back\n *   to whatever the fusion recorded.\n */\nfunction classifierClasses(session: OrtSession): number | null {\n    const index = session.outputNames.indexOf(OUTPUT_PROBS);\n    if (index < 0) return null;\n    const shape = session.outputShapes[index];\n    if (shape === undefined || shape.length === 0) return null;\n    return shape[shape.length - 1] ?? null;\n}\n\n/**\n * Build a class id → name record from an ordered label list.\n *\n * @param labels Labels indexed by class id.\n * @returns The equivalent record.\n */\nfunction indexNames(labels: readonly string[]): Readonly<Record<number, string>> {\n    const names: Record<number, string> = {};\n    for (let i = 0; i < labels.length; i++) names[i] = labels[i] as string;\n    return names;\n}\n\n/**\n * Fetch an output tensor by name.\n *\n * @param outputs The run's results.\n * @param name The output's name in the pipeline contract.\n * @returns The tensor.\n * @throws {@link FusionError} when the graph does not carry that output, which\n *   means the file is not a pipeline this version can drive.\n */\nfunction output(outputs: Record<string, ort.Tensor>, name: string): ort.Tensor {\n    const tensor = outputs[name];\n    if (tensor === undefined) {\n        throw new FusionError(`The fused pipeline is missing its '${name}' output.`);\n    }\n    return tensor;\n}\n\n/**\n * Fetch a float output by name.\n *\n * @param outputs The run's results.\n * @param name The output's name in the pipeline contract.\n * @returns The output's data.\n * @throws {@link FusionError} when the graph does not carry that output.\n */\nfunction floats(outputs: Record<string, ort.Tensor>, name: string): Float32Array {\n    return asFloat32Array(output(outputs, name).data);\n}\n\n/**\n * Fetch an integer output by name, normalizing ORT's 64-bit representation.\n *\n * ONNX Runtime Web returns `int64` tensors as `BigInt64Array`, whose values do\n * not compare or index like numbers. Class ids and detection counts are always\n * small, so widening them to `number` here is lossless and keeps every caller\n * free of `BigInt` handling.\n *\n * @param outputs The run's results.\n * @param name The output's name in the pipeline contract.\n * @returns The output's values as plain numbers.\n * @throws {@link FusionError} when the graph does not carry that output.\n */\nfunction integers(outputs: Record<string, ort.Tensor>, name: string): number[] {\n    const data = output(outputs, name).data as BigInt64Array | Int32Array | Float32Array;\n    const values: number[] = [];\n    for (let i = 0; i < data.length; i++) values.push(Number(data[i]));\n    return values;\n}\n\n/**\n * Convert an image to the `[1, 3, H, W]` float32 tensor the graph expects.\n *\n * @param image The image to convert.\n * @returns The batched CHW tensor, scaled to `[0, 1]`.\n */\nfunction tensorOf(image: RGBImage): ort.Tensor {\n    const chw = toCHW(toFloat32(image), image.width, image.height, 3);\n    return toFloat32Tensor(chw, [1, 3, image.height, image.width]);\n}\n\n/**\n * Assemble one detection, filling the Ultralytics-style aliases.\n *\n * @param classId Detector class index.\n * @param className Detector class name.\n * @param confidence Detection score.\n * @param bbox Box in original-image coordinates.\n * @param croppedImage The region the box covers.\n * @param classification What the classification stage said about that region.\n * @returns The detection object.\n */\nfunction detection(\n    classId: number,\n    className: string,\n    confidence: number,\n    bbox: BoundingBox,\n    croppedImage: RGBImage,\n    classification: ClassificationResult,\n    mask: Mask | null,\n): DetectionResult {\n    return {\n        classId,\n        className,\n        confidence,\n        bbox,\n        cls: classId,\n        name: className,\n        conf: confidence,\n        box: bbox,\n        croppedImage,\n        classification,\n        mask,\n    };\n}\n\n/**\n * Resample a crop-space mask onto the cropped image's own pixel grid.\n *\n * Exported for the test suite rather than for callers: it is not re-exported\n * from the package root, and a mask reaches user code already resampled, on\n * {@link DetectionResult.mask}.\n *\n * The graph computes masks at the resolution the segmenter was exported at,\n * which is rarely the size of the box in the original image. Resampling here is\n * what lets {@link DetectionResult.mask} carry the contract `Segmenter` already\n * produces, so a caller that handles one handles the other.\n *\n * Nearest-neighbour on purpose: the values are already thresholded, and\n * interpolating between them would invent edge pixels that are neither.\n *\n * @param masks The graph's `masks` output.\n * @param row Which detection row to read.\n * @param cropWidth Width of one mask in the graph's output.\n * @param cropHeight Height of one mask in the graph's output.\n * @param target The cropped image the mask must line up with.\n * @returns The mask on the crop's grid, or `null` when the box has no area.\n */\nexport function maskToBox(\n    masks: Float32Array,\n    row: number,\n    cropWidth: number,\n    cropHeight: number,\n    target: RGBImage,\n): Mask | null {\n    const { width, height } = target;\n    if (width < 1 || height < 1) return null;\n    const stride = cropWidth * cropHeight;\n    const source = masks.subarray(row * stride, (row + 1) * stride);\n    const data = new Uint8Array(width * height);\n    for (let y = 0; y < height; y++) {\n        const sourceRow = Math.min(Math.floor((y * cropHeight) / height), cropHeight - 1);\n        for (let x = 0; x < width; x++) {\n            const sourceColumn = Math.min(Math.floor((x * cropWidth) / width), cropWidth - 1);\n            data[y * width + x] =\n                (source[sourceRow * cropWidth + sourceColumn] ?? 0) > 0.5 ? 255 : 0;\n        }\n    }\n    return new Mask(data, width, height);\n}\n\n/**\n * Assemble the bulk-array `Boxes` view from per-instance detections.\n *\n * @param detections The surviving detections.\n * @param origShape `[height, width]` of the original image.\n * @returns The bulk view, empty when nothing survived.\n */\nfunction bulkBoxes(\n    detections: readonly DetectionResult[],\n    origShape: readonly [number, number],\n): Boxes {\n    const xyxy = new Float32Array(detections.length * 4);\n    const cls = new Int32Array(detections.length);\n    const conf = new Float32Array(detections.length);\n    detections.forEach((entry, index) => {\n        const [x1, y1, x2, y2] = entry.bbox.xyxy;\n        xyxy[index * 4] = x1;\n        xyxy[index * 4 + 1] = y1;\n        xyxy[index * 4 + 2] = x2;\n        xyxy[index * 4 + 3] = y2;\n        cls[index] = entry.classId;\n        conf[index] = entry.confidence;\n    });\n    return new Boxes(xyxy, cls, conf, origShape);\n}\n"],"mappings":"waAwGA,IAAa,EAAb,MAAa,UAAuB,EAAA,UAAW,CAGtB,MACA,QACA,OACA,kBACA,iBACA,cAPrB,YACI,EACA,EACA,EACA,EACA,EACA,EACA,EACF,CACE,MAAM,CAAO,EAPI,KAAA,MAAA,EACA,KAAA,QAAA,EACA,KAAA,OAAA,EACA,KAAA,kBAAA,EACA,KAAA,iBAAA,EACA,KAAA,cAAA,CAGrB,CAEA,eAAmD,KAanD,MAAM,OAAO,EAAe,EAAkB,CAC1C,GAAM,CAAC,EAAO,GAAU,KAAK,MAAM,UACnC,IAAK,IAAI,EAAI,EAAG,EAAI,EAAM,IAAK,CAC3B,IAAM,EAAoC,EACrC,EAAA,aAAc,EAAA,gBAAgB,EAAA,eAAe,EAAO,CAAM,EAAG,CAC1D,EACA,EACA,EACA,CACJ,CAAC,CACL,EACI,KAAK,MAAM,mBACX,EAAM,EAAA,cAAgB,EAAA,gBAAgB,EAAA,eAAe,EAAO,CAAM,EAAG,CACjE,EACA,EACA,EACA,CACJ,CAAC,EACD,EAAM,EAAA,aAAe,EAAA,gBAAgB,IAAI,aAAa,CAAC,CAAC,CAAC,EAAG,CAAC,CAAC,CAAC,EAC/D,EAAM,EAAA,WAAa,EAAA,gBAAgB,IAAI,aAAa,CAAC,EAAG,CAAC,CAAC,EAAG,CAAC,CAAC,CAAC,GAEpE,MAAM,KAAK,SAAS,IAAI,CAAK,CACjC,CACJ,CASA,IAAY,WAA+B,CACvC,GAAI,KAAK,iBAAmB,KAAM,CAC9B,GAAM,CAAC,EAAO,GAAU,KAAK,MAAM,UACnC,KAAK,eAAiB,IAAI,EAAA,kBAAkB,EAAO,CAAM,CAC7D,CACA,OAAO,KAAK,cAChB,CAWA,aAAa,OACT,EACA,EAAiC,CAAC,EACX,CACvB,IAAM,EAAU,MAAM,EAAA,WAAW,OAAO,EAAO,CAAO,EAChD,EAAO,EAAA,eAAe,EAAQ,QAAQ,EAC5C,GAAI,IAAS,KACT,MAAM,IAAI,EAAA,YACN,mNAGJ,EAGJ,IAAM,EAAS,EAAA,cAAc,EAAQ,QAAU,EAAK,eAAiB,MAAM,EACrE,EAAmB,EAAA,cAAc,EAAQ,kBAAoB,EAAK,gBAAiB,CACrF,WAAY,EAAkB,CAAO,GAAK,IAAA,EAC9C,CAAC,EACD,OAAO,IAAI,EACP,EACA,EACA,EACA,EAAW,CAAM,EACjB,EACA,EAAW,CAAgB,EAC3B,EAAQ,cAAgB,EAC5B,CACJ,CAGA,IAAI,MAAmB,CACnB,OAAO,KAAK,KAChB,CAGA,IAAI,WAAuC,CACvC,OAAO,KAAK,MAAM,SACtB,CAGA,IAAI,QAA4B,CAC5B,OAAO,KAAK,OAChB,CAGA,IAAI,OAA0C,CAC1C,OAAO,KAAK,MAChB,CAGA,IAAI,kBAAsC,CACtC,OAAO,KAAK,iBAChB,CAGA,IAAI,iBAAoD,CACpD,OAAO,KAAK,gBAChB,CAQA,MAAM,KACF,EACA,EAAwC,CAAC,EACT,CAChC,OAAO,KAAK,QAAQ,EAAO,CAAO,CACtC,CAUA,MAAM,QACF,EACA,EAAwC,CAAC,EACT,CAChC,IAAM,EAAQ,IAAI,EAAA,WACZ,EAAO,OAAO,GAAU,SAAW,EAAQ,KAC3C,CAAE,MAAO,EAAU,UAAW,MAAM,EAAA,gBAAgB,CAAK,EAC/D,EAAM,MAAM,MAAM,EAClB,GAAM,CAAE,QAAO,QAAO,UAAS,SAAQ,UAAW,KAAK,YAAY,EAAU,CAAM,EACnF,EAAM,MAAM,YAAY,EACxB,IAAI,EACJ,GAAI,CACA,EAAU,MAAM,KAAK,SAAS,IAAI,CAAK,CAC3C,QAAU,CACN,KAAK,UAAU,QAAQ,CAAE,QAAO,CAAC,CACrC,CACA,EAAM,MAAM,WAAW,EAEvB,IAAM,EAAc,EAAO,EAAS,EAAA,YAAY,EAC1C,EAAQ,EAAO,EAAS,EAAA,YAAY,EACpC,EAAS,EAAO,EAAS,EAAA,aAAa,EACtC,EAAU,EAAS,EAAS,EAAA,cAAc,EAC1C,EAAQ,EAAA,eAAe,EAAY,IAAI,EACvC,EAAW,EAAS,EAAA,gBAA8B,CAAC,CAAC,IAAM,EAC1D,EAAQ,KAAK,MAAM,SAAW,EAAO,EAAS,EAAA,YAAY,EAAI,KAC9D,CAAC,EAAW,GAAc,KAAK,MAAM,SACrC,EAAO,KAAK,IAAI,EAAU,KAAK,MAAM,EAAM,OAAS,CAAC,CAAC,EACtD,EAAa,EAAY,KAAK,EAAY,KAAK,OAAS,IAAM,EAE9D,EAAU,EAAQ,UAAY,IAAA,GAAY,KAAO,IAAI,IAAI,EAAQ,OAAO,EACxE,EAAQ,EAAQ,eAAiB,EACjC,EAAgC,CAAC,EACvC,IAAK,IAAI,EAAM,EAAG,EAAM,EAAM,IAAO,CACjC,IAAM,EAAU,EAAQ,IAAQ,EAC1B,EAAa,EAAO,IAAQ,EAClC,GAAI,EAAa,GAAU,IAAY,MAAQ,CAAC,EAAQ,IAAI,CAAO,EAAI,SAEvE,IAAM,EAAO,KAAK,YAAY,EAAO,EAAK,CAAE,QAAO,UAAS,SAAQ,UAAS,CAAC,EACxE,EAAU,EAAA,UAAU,EAAU,CAAI,EACxC,EAAW,KACP,EACI,EACA,KAAK,OAAO,IAAY,SAAS,IACjC,EACA,EACA,EACA,KAAK,UACD,EAAM,SAAS,EAAM,GAAa,EAAM,GAAK,CAAU,EACvD,EACA,EAAQ,IACZ,EACA,IAAU,KAAO,KAAO,EAAU,EAAO,EAAK,EAAW,EAAY,CAAO,CAChF,CACJ,CACJ,CAEA,EAAA,kBAAkB,EAAW,OAAQ,CACjC,aAAc,EAAQ,cAAgB,KAAK,cAC3C,cAAe,KAAK,IAAI,EAAO,KAAK,MAAM,aAAa,EACvD,QAAS,EAAQ,QACjB,MACJ,CAAC,EAED,IAAM,EAAuC,CAAC,EAAS,OAAQ,EAAS,KAAK,EAE7E,OADA,EAAM,MAAM,aAAa,EAClB,CACH,IAAI,EAAA,sBACA,EAAU,EAAY,CAAS,EAC/B,EACA,KAAK,OACL,KAAK,iBACL,EACA,EACA,EACA,EAAM,MAAM,CAChB,CACJ,CACJ,CAqBA,YACI,EACA,EAOF,CACE,GAAM,CAAC,EAAO,GAAU,KAAK,MAAM,UAC7B,EAAQ,KAAK,UAAU,IAAI,EAAO,CAAM,EACxC,EAAoC,EACrC,EAAA,aAAc,EAAA,gBAAgB,EAAM,KAAM,CAAC,EAAG,EAAG,EAAQ,CAAK,CAAC,CACpE,EAUA,OARI,KAAK,MAAM,mBACX,EAAM,EAAA,cAAgB,EAAS,CAAK,EACpC,EAAM,EAAA,aAAe,EAAA,gBAAgB,IAAI,aAAa,CAAC,EAAM,KAAK,CAAC,EAAG,CAAC,CAAC,CAAC,EACzE,EAAM,EAAA,WAAa,EAAA,gBACf,IAAI,aAAa,CAAC,EAAM,QAAS,EAAM,MAAM,CAAC,EAC9C,CAAC,CAAC,CACN,GAEG,CACH,QACA,MAAO,EAAM,MACb,QAAS,EAAM,QACf,OAAQ,EAAM,OACd,OAAQ,EAAM,MAClB,CACJ,CAQA,YACI,EACA,EACA,EAMW,CACX,GAAM,CAAE,QAAO,UAAS,SAAQ,YAAa,EACvC,EAAM,GAA2B,EAAM,EAAM,EAAI,IAAW,EAC5D,EAAU,GAA0B,KAAK,IAAI,KAAK,IAAI,EAAO,CAAC,EAAG,EAAS,KAAK,EAC/E,EAAU,GAA0B,KAAK,IAAI,KAAK,IAAI,EAAO,CAAC,EAAG,EAAS,MAAM,EACtF,OAAO,IAAI,EAAA,YACP,GAAQ,EAAG,CAAC,EAAI,GAAW,CAAK,EAChC,GAAQ,EAAG,CAAC,EAAI,GAAU,CAAK,EAC/B,GAAQ,EAAG,CAAC,EAAI,GAAW,CAAK,EAChC,GAAQ,EAAG,CAAC,EAAI,GAAU,CAAK,CACnC,CACJ,CAUA,UACI,EACA,EACA,EACoB,CACpB,IAAM,EAAS,KAAK,MAAM,aAAe,EAAA,QAAQ,CAAG,EAAI,EAClD,CAAE,UAAS,UAAW,EAAA,KAAK,EAAQ,GAAK,IAAI,EAC5C,EAAoC,CAAC,EAC3C,IAAK,IAAI,EAAI,EAAG,EAAI,EAAQ,OAAQ,IAAK,CACrC,IAAM,EAAU,EAAQ,IAAM,EACxB,EAAc,EAAO,IAAM,EAC3B,EAAY,KAAK,kBAAkB,IAAY,SAAS,IAC9D,EAAc,KAAK,CACf,UACA,YACA,cACA,IAAK,EACL,KAAM,EACN,KAAM,CACV,CAAC,CACL,CACA,IAAM,EAAM,EAAc,IAAM,CAC5B,QAAS,EACT,UAAW,UACX,YAAa,EACb,IAAK,EACL,KAAM,UACN,KAAM,CACV,EACA,MAAO,CACH,QAAS,EAAI,QACb,UAAW,EAAI,UACf,WAAY,EAAI,YAChB,IAAK,EAAI,QACT,KAAM,EAAI,UACV,KAAM,EAAI,YACV,QACA,eACJ,CACJ,CACJ,EAUA,SAAS,EAAkB,EAAoC,CAC3D,IAAM,EAAQ,EAAQ,YAAY,QAAQ,EAAA,YAAY,EACtD,GAAI,EAAQ,EAAG,OAAO,KACtB,IAAM,EAAQ,EAAQ,aAAa,GAEnC,OADI,IAAU,IAAA,IAAa,EAAM,SAAW,EAAU,KAC/C,EAAM,EAAM,OAAS,IAAM,IACtC,CAQA,SAAS,EAAW,EAA6D,CAC7E,IAAM,EAAgC,CAAC,EACvC,IAAK,IAAI,EAAI,EAAG,EAAI,EAAO,OAAQ,IAAK,EAAM,GAAK,EAAO,GAC1D,OAAO,CACX,CAWA,SAAS,EAAO,EAAqC,EAA0B,CAC3E,IAAM,EAAS,EAAQ,GACvB,GAAI,IAAW,IAAA,GACX,MAAM,IAAI,EAAA,YAAY,sCAAsC,EAAK,UAAU,EAE/E,OAAO,CACX,CAUA,SAAS,EAAO,EAAqC,EAA4B,CAC7E,OAAO,EAAA,eAAe,EAAO,EAAS,CAAI,CAAC,CAAC,IAAI,CACpD,CAeA,SAAS,EAAS,EAAqC,EAAwB,CAC3E,IAAM,EAAO,EAAO,EAAS,CAAI,CAAC,CAAC,KAC7B,EAAmB,CAAC,EAC1B,IAAK,IAAI,EAAI,EAAG,EAAI,EAAK,OAAQ,IAAK,EAAO,KAAK,OAAO,EAAK,EAAE,CAAC,EACjE,OAAO,CACX,CAQA,SAAS,EAAS,EAA6B,CAC3C,IAAM,EAAM,EAAA,MAAM,EAAA,UAAU,CAAK,EAAG,EAAM,MAAO,EAAM,OAAQ,CAAC,EAChE,OAAO,EAAA,gBAAgB,EAAK,CAAC,EAAG,EAAG,EAAM,OAAQ,EAAM,KAAK,CAAC,CACjE,CAaA,SAAS,EACL,EACA,EACA,EACA,EACA,EACA,EACA,EACe,CACf,MAAO,CACH,UACA,YACA,aACA,OACA,IAAK,EACL,KAAM,EACN,KAAM,EACN,IAAK,EACL,eACA,iBACA,MACJ,CACJ,CAwBA,SAAgB,EACZ,EACA,EACA,EACA,EACA,EACW,CACX,GAAM,CAAE,QAAO,UAAW,EAC1B,GAAI,EAAQ,GAAK,EAAS,EAAG,OAAO,KACpC,IAAM,EAAS,EAAY,EACrB,EAAS,EAAM,SAAS,EAAM,GAAS,EAAM,GAAK,CAAM,EACxD,EAAO,IAAI,WAAW,EAAQ,CAAM,EAC1C,IAAK,IAAI,EAAI,EAAG,EAAI,EAAQ,IAAK,CAC7B,IAAM,EAAY,KAAK,IAAI,KAAK,MAAO,EAAI,EAAc,CAAM,EAAG,EAAa,CAAC,EAChF,IAAK,IAAI,EAAI,EAAG,EAAI,EAAO,IAAK,CAC5B,IAAM,EAAe,KAAK,IAAI,KAAK,MAAO,EAAI,EAAa,CAAK,EAAG,EAAY,CAAC,EAChF,EAAK,EAAI,EAAQ,IACZ,EAAO,EAAY,EAAY,IAAiB,GAAK,GAAM,IAAM,CAC1E,CACJ,CACA,OAAO,IAAI,EAAA,KAAK,EAAM,EAAO,CAAM,CACvC,CASA,SAAS,EACL,EACA,EACK,CACL,IAAM,EAAO,IAAI,aAAa,EAAW,OAAS,CAAC,EAC7C,EAAM,IAAI,WAAW,EAAW,MAAM,EACtC,EAAO,IAAI,aAAa,EAAW,MAAM,EAU/C,OATA,EAAW,SAAS,EAAO,IAAU,CACjC,GAAM,CAAC,EAAI,EAAI,EAAI,GAAM,EAAM,KAAK,KACpC,EAAK,EAAQ,GAAK,EAClB,EAAK,EAAQ,EAAI,GAAK,EACtB,EAAK,EAAQ,EAAI,GAAK,EACtB,EAAK,EAAQ,EAAI,GAAK,EACtB,EAAI,GAAS,EAAM,QACnB,EAAK,GAAS,EAAM,UACxB,CAAC,EACM,IAAI,EAAA,MAAM,EAAM,EAAK,EAAM,CAAS,CAC/C"}

package/dist/vision/tasks/detectClassify.js

@@ -1,17 +1,18 @@
 import { FusionError as e } from "../core/exceptions.js";
-import { BoundingBox as t, RGBImage as n } from "../types.js";
+import { BoundingBox as t, Mask as n } from "../types.js";
 import { SpeedTimer as r } from "../core/timing.js";
 import { Boxes as i, DetectClassifyResults as a } from "../results.js";
 import { resolveLabels as o } from "../labels.js";
-import { OrtSession as s } from "../core/session.js";
-import { loadImage as c } from "../io/image.js";
-import { INPUT_IMAGE as l, INPUT_PAD as u, INPUT_SCALE as d, INPUT_SOURCE as f, OUTPUT_BOXES as p, OUTPUT_CLASSES as m, OUTPUT_PROBS as h, OUTPUT_SCORES as g, readFusionSpec as _ } from "../fusion.js";
-import { toCHW as v, toFloat32 as y, toFloat32Tensor as b } from "../preprocess/image.js";
-import { LetterboxPipeline as x, zeroTensorData as S } from "../preprocess/pipeline.js";
-import { softmax as C, topK as w } from "../postprocess/classification.js";
-import { VisionTask as T, requireDetections as E } from "./base.js";
+import { asFloat32Array as s } from "../core/dtypes.js";
+import { OrtSession as c } from "../core/session.js";
+import { loadImageSource as l } from "../io/image.js";
+import { INPUT_IMAGE as u, INPUT_PAD as d, INPUT_SCALE as f, INPUT_SOURCE as p, OUTPUT_BOXES as m, OUTPUT_CLASSES as h, OUTPUT_MASKS as g, OUTPUT_PROBS as _, OUTPUT_SCORES as v, readFusionSpec as y } from "../fusion.js";
+import { toCHW as b, toFloat32 as x, toFloat32Tensor as S } from "../preprocess/image.js";
+import { LetterboxPipeline as C, zeroTensorData as w } from "../preprocess/pipeline.js";
+import { softmax as T, topK as E } from "../postprocess/classification.js";
+import { VisionTask as D, cropToBox as O, requireDetections as k } from "./base.js";
 //#region src/vision/tasks/detectClassify.ts
-var D = class n extends T {
+var A = class n extends D {
 	_spec;
 	_labels;
 	_names;
@@ -25,32 +26,32 @@ var D = class n extends T {
 	async warmup(e = 1) {
 		let [t, n] = this._spec.inputSize;
 		for (let r = 0; r < e; r++) {
-			let e = { [l]: b(S(t, n), [
+			let e = { [u]: S(w(t, n), [
 				1,
 				3,
 				n,
 				t
 			]) };
-			this._spec.needsSourceImage && (e[f] = b(S(t, n), [
+			this._spec.needsSourceImage && (e[p] = S(w(t, n), [
 				1,
 				3,
 				n,
 				t
-			]), e[d] = b(new Float32Array([1]), [1]), e[u] = b(new Float32Array([0, 0]), [2])), await this._session.run(e);
+			]), e[f] = S(new Float32Array([1]), [1]), e[d] = S(new Float32Array([0, 0]), [2])), await this._session.run(e);
 		}
 	}
 	get _pipeline() {
 		if (this._pipelineCache === null) {
 			let [e, t] = this._spec.inputSize;
-			this._pipelineCache = new x(e, t);
+			this._pipelineCache = new C(e, t);
 		}
 		return this._pipelineCache;
 	}
 	static async create(t, r = {}) {
-		let i = await s.create(t, r), a = _(i.metadata);
+		let i = await c.create(t, r), a = y(i.metadata);
 		if (a === null) throw new e("This model carries no fused-pipeline metadata, so DetectClassify cannot tell how to drive it. Build one with ort_vision_sdk.compose.fuse_detect_classify, or load a plain model with Detector/Classifier instead.");
-		let c = o(r.labels ?? a.detectorNames ?? "coco"), l = o(r.classifierLabels ?? a.classifierNames, { numClasses: O(i) ?? void 0 });
-		return new n(i, a, c, k(c), l, k(l), r.raiseOnEmpty ?? !1);
+		let s = o(r.labels ?? a.detectorNames ?? "coco"), l = o(r.classifierLabels ?? a.classifierNames, { numClasses: j(i) ?? void 0 });
+		return new n(i, a, s, M(s), l, M(l), r.raiseOnEmpty ?? !1);
 	}
 	get spec() {
 		return this._spec;
@@ -74,45 +75,51 @@ var D = class n extends T {
 		return this.predict(e, t);
 	}
 	async predict(e, t = {}) {
-		let n = new r(), i = typeof e == "string" ? e : null, o = await c(e);
+		let n = new r(), i = typeof e == "string" ? e : null, { image: o, canvas: c } = await l(e);
 		n.stage("load");
-		let { feeds: s, scale: l, padLeft: u, padTop: d } = this._preprocess(o);
+		let { feeds: u, scale: d, padLeft: f, padTop: p, reused: y } = this._preprocess(o, c);
 		n.stage("preprocess");
-		let f = await this._session.run(s);
-		this._pipeline.release(), n.stage("inference");
-		let _ = A(f, h), v = j(f, p), y = j(f, g), b = M(f, m), x = _.data, S = M(f, "num_detections")[0] ?? 0, C = Math.min(S, Math.floor(v.length / 4)), w = _.dims[_.dims.length - 1] ?? 0, T = t.classes === void 0 ? null : new Set(t.classes), D = t.confThreshold ?? 0, O = [];
-		for (let e = 0; e < C; e++) {
-			let n = b[e] ?? 0, r = y[e] ?? 0;
-			if (r < D || T !== null && !T.has(n)) continue;
-			let i = this._toOriginal(v, e, {
-				scale: l,
-				padLeft: u,
-				padTop: d,
+		let b;
+		try {
+			b = await this._session.run(u);
+		} finally {
+			this._pipeline.release({ reused: y });
+		}
+		n.stage("inference");
+		let x = N(b, _), S = P(b, m), C = P(b, v), w = F(b, h), T = s(x.data), E = F(b, "num_detections")[0] ?? 0, D = this._spec.hasMasks ? P(b, g) : null, [A, j] = this._spec.cropSize, M = Math.min(E, Math.floor(S.length / 4)), I = x.dims[x.dims.length - 1] ?? 0, B = t.classes === void 0 ? null : new Set(t.classes), V = t.confThreshold ?? 0, H = [];
+		for (let e = 0; e < M; e++) {
+			let n = w[e] ?? 0, r = C[e] ?? 0;
+			if (r < V || B !== null && !B.has(n)) continue;
+			let i = this._toOriginal(S, e, {
+				scale: d,
+				padLeft: f,
+				padTop: p,
 				original: o
-			}), a = P(o, i);
-			O.push(F(n, this._names[n] ?? `class_${n}`, r, i, a, this._classify(x.subarray(e * w, (e + 1) * w), a, t.topK)));
+			}), a = O(o, i);
+			H.push(L(n, this._names[n] ?? `class_${n}`, r, i, a, this._classify(T.subarray(e * I, (e + 1) * I), a, t.topK), D === null ? null : R(D, e, A, j, a)));
 		}
-		E(O.length, {
+		k(H.length, {
 			raiseOnEmpty: t.raiseOnEmpty ?? this._raiseOnEmpty,
-			confThreshold: Math.max(D, this._spec.confThreshold),
+			confThreshold: Math.max(V, this._spec.confThreshold),
 			classes: t.classes,
 			path: i
 		});
-		let k = [o.height, o.width];
-		return n.stage("postprocess"), [new a(I(O, k), O, this._names, this._classifierNames, o, k, i, n.speed())];
+		let U = [o.height, o.width];
+		return n.stage("postprocess"), [new a(z(H, U), H, this._names, this._classifierNames, o, U, i, n.speed())];
 	}
-	_preprocess(e) {
-		let [t, n] = this._spec.inputSize, r = this._pipeline.run(e), i = { [l]: b(r.data, [
+	_preprocess(e, t) {
+		let [n, r] = this._spec.inputSize, i = this._pipeline.run(e, t), a = { [u]: S(i.data, [
 			1,
 			3,
-			n,
-			t
+			r,
+			n
 		]) };
-		return this._spec.needsSourceImage && (i[f] = N(e), i[d] = b(new Float32Array([r.scale]), [1]), i[u] = b(new Float32Array([r.padLeft, r.padTop]), [2])), {
-			feeds: i,
-			scale: r.scale,
-			padLeft: r.padLeft,
-			padTop: r.padTop
+		return this._spec.needsSourceImage && (a[p] = I(e), a[f] = S(new Float32Array([i.scale]), [1]), a[d] = S(new Float32Array([i.padLeft, i.padTop]), [2])), {
+			feeds: a,
+			scale: i.scale,
+			padLeft: i.padLeft,
+			padTop: i.padTop,
+			reused: i.reused
 		};
 	}
 	_toOriginal(e, n, r) {
@@ -120,7 +127,7 @@ var D = class n extends T {
 		return new t(l((c(0) - a) / i), u((c(1) - o) / i), l((c(2) - a) / i), u((c(3) - o) / i));
 	}
 	_classify(e, t, n) {
-		let r = this._spec.applySoftmax ? C(e) : e, { indices: i, values: a } = w(r, n ?? null), o = [];
+		let r = this._spec.applySoftmax ? T(e) : e, { indices: i, values: a } = E(r, n ?? null), o = [];
 		for (let e = 0; e < i.length; e++) {
 			let t = i[e] ?? 0, n = a[e] ?? 0, r = this._classifierLabels[t] ?? `class_${t}`;
 			o.push({
@@ -152,50 +159,40 @@ var D = class n extends T {
 		};
 	}
 };
-function O(e) {
-	let t = e.outputNames.indexOf(h);
+function j(e) {
+	let t = e.outputNames.indexOf(_);
 	if (t < 0) return null;
 	let n = e.outputShapes[t];
 	return n === void 0 || n.length === 0 ? null : n[n.length - 1] ?? null;
 }
-function k(e) {
+function M(e) {
 	let t = {};
 	for (let n = 0; n < e.length; n++) t[n] = e[n];
 	return t;
 }
-function A(t, n) {
+function N(t, n) {
 	let r = t[n];
 	if (r === void 0) throw new e(`The fused pipeline is missing its '${n}' output.`);
 	return r;
 }
-function j(e, t) {
-	return A(e, t).data;
+function P(e, t) {
+	return s(N(e, t).data);
 }
-function M(e, t) {
-	let n = A(e, t).data, r = [];
+function F(e, t) {
+	let n = N(e, t).data, r = [];
 	for (let e = 0; e < n.length; e++) r.push(Number(n[e]));
 	return r;
 }
-function N(e) {
-	let t = v(y(e), e.width, e.height, 3);
-	return b(t, [
+function I(e) {
+	let t = b(x(e), e.width, e.height, 3);
+	return S(t, [
 		1,
 		3,
 		e.height,
 		e.width
 	]);
 }
-function P(e, t) {
-	let [r, i, a, o] = t.asIntXyxy(), s = Math.max(0, r), c = Math.max(0, i), l = Math.min(e.width, a), u = Math.min(e.height, o);
-	if (l <= s || u <= c) return new n(/* @__PURE__ */ new Uint8Array(), 0, 0);
-	let d = l - s, f = u - c, p = new Uint8Array(d * f * 3);
-	for (let t = 0; t < f; t++) {
-		let n = ((c + t) * e.width + s) * 3;
-		p.set(e.data.subarray(n, n + d * 3), t * d * 3);
-	}
-	return new n(p, d, f);
-}
-function F(e, t, n, r, i, a) {
+function L(e, t, n, r, i, a, o) {
 	return {
 		classId: e,
 		className: t,
@@ -206,10 +203,24 @@ function F(e, t, n, r, i, a) {
 		conf: n,
 		box: r,
 		croppedImage: i,
-		classification: a
+		classification: a,
+		mask: o
 	};
 }
-function I(e, t) {
+function R(e, t, r, i, a) {
+	let { width: o, height: s } = a;
+	if (o < 1 || s < 1) return null;
+	let c = r * i, l = e.subarray(t * c, (t + 1) * c), u = new Uint8Array(o * s);
+	for (let e = 0; e < s; e++) {
+		let t = Math.min(Math.floor(e * i / s), i - 1);
+		for (let n = 0; n < o; n++) {
+			let i = Math.min(Math.floor(n * r / o), r - 1);
+			u[e * o + n] = (l[t * r + i] ?? 0) > .5 ? 255 : 0;
+		}
+	}
+	return new n(u, o, s);
+}
+function z(e, t) {
 	let n = new Float32Array(e.length * 4), r = new Int32Array(e.length), a = new Float32Array(e.length);
 	return e.forEach((e, t) => {
 		let [i, o, s, c] = e.bbox.xyxy;
@@ -217,6 +228,6 @@ function I(e, t) {
 	}), new i(n, r, a, t);
 }
 //#endregion
-export { D as DetectClassify };
+export { A as DetectClassify, R as maskToBox };
 
 //# sourceMappingURL=detectClassify.js.map

package/dist/vision/tasks/detectClassify.js.map

@@ -1 +1 @@
-{"version":3,"file":"detectClassify.js","names":[],"sources":["../../../src/vision/tasks/detectClassify.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Run a fused detect→classify pipeline in the browser.\n *\n * The file this loads was built by the Python SDK's `ort_vision_sdk.compose`,\n * and it already contains both models plus the crop-and-resize bridge between\n * them. That matters far more in a tab than on a server: two models mean two\n * `.onnx` downloads, two WASM/WebGPU session initializations, and a per-crop\n * round trip through JavaScript to slice, resize and restack the regions before\n * the second model can see them. A fused pipeline has one download, one session\n * and no round trip — the crops are produced and consumed inside the graph.\n */\n\nimport type * as ort from \"onnxruntime-web\";\n\nimport { FusionError } from \"../core/exceptions\";\nimport { type ModelSource, type OrtSessionOptions, OrtSession } from \"../core/session\";\nimport { SpeedTimer } from \"../core/timing\";\nimport {\n    INPUT_IMAGE,\n    INPUT_PAD,\n    INPUT_SCALE,\n    INPUT_SOURCE,\n    OUTPUT_BOXES,\n    OUTPUT_CLASSES,\n    OUTPUT_NUM_DETECTIONS,\n    OUTPUT_PROBS,\n    OUTPUT_SCORES,\n    type FusionSpec,\n    readFusionSpec,\n} from \"../fusion\";\nimport { type ImageInput, loadImage } from \"../io/image\";\nimport { type LabelSpec, resolveLabels } from \"../labels\";\nimport { softmax, topK } from \"../postprocess/classification\";\nimport { toCHW, toFloat32, toFloat32Tensor } from \"../preprocess/image\";\nimport { LetterboxPipeline, zeroTensorData } from \"../preprocess/pipeline\";\nimport { Boxes, DetectClassifyResults } from \"../results\";\nimport {\n    BoundingBox,\n    RGBImage,\n    type ClassProbability,\n    type ClassificationResult,\n    type DetectionResult,\n} from \"../types\";\nimport { VisionTask, requireDetections } from \"./base\";\n\nexport interface DetectClassifyOptions extends OrtSessionOptions {\n    /**\n     * Class label spec for the **detection** stage — see {@link resolveLabels}.\n     * Defaults to the names recorded at fusion time, falling back to the COCO\n     * 80-class preset when the fusion recorded none.\n     */\n    readonly labels?: LabelSpec;\n    /**\n     * Class label spec for the **classification** stage. Defaults to the recorded\n     * names, falling back to generated `class_<id>` names.\n     */\n    readonly classifierLabels?: LabelSpec;\n    /**\n     * If `true`, a run that finds nothing throws {@link NoDetectionsError}\n     * instead of returning an empty envelope. Default `false`, because looking\n     * and finding nothing is a successful inference. Turn it on when an empty\n     * result means the surrounding pipeline should stop rather than carry on with\n     * zero rows. Can be overridden per `predict` call.\n     */\n    readonly raiseOnEmpty?: boolean;\n}\n\nexport interface DetectClassifyPredictOptions {\n    /**\n     * Drop detections scoring below this. The graph's own NMS threshold was fixed\n     * at fusion time and cannot be lowered here — this only filters further.\n     */\n    readonly confThreshold?: number;\n    /** If set, keep only detections whose detector `classId` is in this list. */\n    readonly classes?: readonly number[];\n    /** Truncate each detection's `classification.probabilities` to its top-k entries. */\n    readonly topK?: number;\n    /** Override the constructor's `raiseOnEmpty` setting for this call. */\n    readonly raiseOnEmpty?: boolean;\n}\n\n/**\n * Detector and classifier running as a single ONNX model.\n *\n * Everything the pipeline needs to know about itself — the resolution to\n * letterbox to, whether it wants the full-resolution image as well, whether its\n * classifier output still needs a softmax, the class names of both stages — was\n * written into the file at fusion time and is read back here. Nothing is\n * restated on the JavaScript side, so nothing can drift out of step with the\n * Python side that built it.\n *\n * @example\n * ```typescript\n * const pipeline = await DetectClassify.create(\"/models/pipeline.onnx\");\n * const result = (await pipeline.predict(\"/images/flock.jpg\"))[0];\n * for (const detection of result) {\n *   console.log(detection.name, detection.conf, detection.classification?.name);\n * }\n * ```\n */\nexport class DetectClassify extends VisionTask {\n    private constructor(\n        session: OrtSession,\n        private readonly _spec: FusionSpec,\n        private readonly _labels: readonly string[],\n        private readonly _names: Readonly<Record<number, string>>,\n        private readonly _classifierLabels: readonly string[],\n        private readonly _classifierNames: Readonly<Record<number, string>>,\n        private readonly _raiseOnEmpty: boolean,\n    ) {\n        super(session);\n    }\n\n    private _pipelineCache: LetterboxPipeline | null = null;\n\n    /**\n     * Run the model once on zero-filled inputs, paying one-time costs up front.\n     *\n     * Worth more here than on a single-stage task: a fused pipeline is two models\n     * plus the bridge in one graph, so the first inference compiles shaders for\n     * all of it. Calling this while a loading spinner is still up moves that cost\n     * somewhere the user is already waiting.\n     *\n     * @param runs How many warm-up inferences to run. One is enough for WASM;\n     *   WebGPU sometimes settles on the second.\n     */\n    async warmup(runs: number = 1): Promise<void> {\n        const [width, height] = this._spec.inputSize;\n        for (let i = 0; i < runs; i++) {\n            const feeds: Record<string, ort.Tensor> = {\n                [INPUT_IMAGE]: toFloat32Tensor(zeroTensorData(width, height), [\n                    1,\n                    3,\n                    height,\n                    width,\n                ]),\n            };\n            if (this._spec.needsSourceImage) {\n                feeds[INPUT_SOURCE] = toFloat32Tensor(zeroTensorData(width, height), [\n                    1,\n                    3,\n                    height,\n                    width,\n                ]);\n                feeds[INPUT_SCALE] = toFloat32Tensor(new Float32Array([1]), [1]);\n                feeds[INPUT_PAD] = toFloat32Tensor(new Float32Array([0, 0]), [2]);\n            }\n            await this._session.run(feeds);\n        }\n    }\n\n    /**\n     * The fused preprocessing pipeline, built on first use.\n     *\n     * Lazily, because constructing it allocates canvases: a pipeline built in an\n     * environment without a canvas implementation stays constructible, and only\n     * fails if it is actually asked to preprocess something.\n     */\n    private get _pipeline(): LetterboxPipeline {\n        if (this._pipelineCache === null) {\n            const [width, height] = this._spec.inputSize;\n            this._pipelineCache = new LetterboxPipeline(width, height);\n        }\n        return this._pipelineCache;\n    }\n\n    /**\n     * Load a fused pipeline and resolve both label spaces.\n     *\n     * @param model The fused `.onnx` — a URL, an `ArrayBuffer`, or bytes.\n     * @param options Label overrides plus the usual session options.\n     * @throws {@link FusionError} when the model carries no pipeline metadata,\n     *   i.e. it is a plain detector or classifier rather than something\n     *   `ort_vision_sdk.compose` produced.\n     */\n    static async create(\n        model: ModelSource,\n        options: DetectClassifyOptions = {},\n    ): Promise<DetectClassify> {\n        const session = await OrtSession.create(model, options);\n        const spec = readFusionSpec(session.metadata);\n        if (spec === null) {\n            throw new FusionError(\n                \"This model carries no fused-pipeline metadata, so DetectClassify cannot tell how to \" +\n                    \"drive it. Build one with ort_vision_sdk.compose.fuse_detect_classify, or load a \" +\n                    \"plain model with Detector/Classifier instead.\",\n            );\n        }\n\n        const labels = resolveLabels(options.labels ?? spec.detectorNames ?? \"coco\");\n        const classifierLabels = resolveLabels(options.classifierLabels ?? spec.classifierNames, {\n            numClasses: classifierClasses(session) ?? undefined,\n        });\n        return new DetectClassify(\n            session,\n            spec,\n            labels,\n            indexNames(labels),\n            classifierLabels,\n            indexNames(classifierLabels),\n            options.raiseOnEmpty ?? false,\n        );\n    }\n\n    /** The pipeline configuration recorded in the model at fusion time. */\n    get spec(): FusionSpec {\n        return this._spec;\n    }\n\n    /** The `[width, height]` the detection stage runs at. */\n    get inputSize(): readonly [number, number] {\n        return this._spec.inputSize;\n    }\n\n    /** Detector class labels indexed by class id. */\n    get labels(): readonly string[] {\n        return this._labels;\n    }\n\n    /** Detector class id → class name (matches Ultralytics' `model.names`). */\n    get names(): Readonly<Record<number, string>> {\n        return this._names;\n    }\n\n    /** Classifier class labels indexed by class id. */\n    get classifierLabels(): readonly string[] {\n        return this._classifierLabels;\n    }\n\n    /** Classifier class id → class name. */\n    get classifierNames(): Readonly<Record<number, string>> {\n        return this._classifierNames;\n    }\n\n    /**\n     * Alias for {@link predict} — call the pipeline like a torch `nn.Module`.\n     *\n     * Use as `pipeline.call(img)` since JavaScript class instances are not\n     * callable; for direct invocation, prefer `pipeline.predict(img)`.\n     */\n    async call(\n        image: ImageInput,\n        options: DetectClassifyPredictOptions = {},\n    ): Promise<DetectClassifyResults[]> {\n        return this.predict(image, options);\n    }\n\n    /**\n     * Run the pipeline on a single image.\n     *\n     * The returned envelope carries a {@link Speed} breakdown in `speed`. Its\n     * `inference` figure covers detection *and* classification, since the\n     * pipeline runs them as one graph and no boundary between them is observable\n     * from outside.\n     */\n    async predict(\n        image: ImageInput,\n        options: DetectClassifyPredictOptions = {},\n    ): Promise<DetectClassifyResults[]> {\n        const timer = new SpeedTimer();\n        const path = typeof image === \"string\" ? image : null;\n        const original = await loadImage(image);\n        timer.stage(\"load\");\n        const { feeds, scale, padLeft, padTop } = this._preprocess(original);\n        timer.stage(\"preprocess\");\n        const outputs = await this._session.run(feeds);\n        this._pipeline.release();\n        timer.stage(\"inference\");\n\n        const probsTensor = output(outputs, OUTPUT_PROBS);\n        const boxes = floats(outputs, OUTPUT_BOXES);\n        const scores = floats(outputs, OUTPUT_SCORES);\n        const classes = integers(outputs, OUTPUT_CLASSES);\n        const probs = probsTensor.data as Float32Array;\n        const reported = integers(outputs, OUTPUT_NUM_DETECTIONS)[0] ?? 0;\n        const rows = Math.min(reported, Math.floor(boxes.length / 4));\n        const classCount = probsTensor.dims[probsTensor.dims.length - 1] ?? 0;\n\n        const allowed = options.classes === undefined ? null : new Set(options.classes);\n        const floor = options.confThreshold ?? 0;\n        const detections: DetectionResult[] = [];\n        for (let row = 0; row < rows; row++) {\n            const classId = classes[row] ?? 0;\n            const confidence = scores[row] ?? 0;\n            if (confidence < floor || (allowed !== null && !allowed.has(classId))) continue;\n\n            const bbox = this._toOriginal(boxes, row, { scale, padLeft, padTop, original });\n            const cropped = crop(original, bbox);\n            detections.push(\n                detection(\n                    classId,\n                    this._names[classId] ?? `class_${classId}`,\n                    confidence,\n                    bbox,\n                    cropped,\n                    this._classify(\n                        probs.subarray(row * classCount, (row + 1) * classCount),\n                        cropped,\n                        options.topK,\n                    ),\n                ),\n            );\n        }\n\n        requireDetections(detections.length, {\n            raiseOnEmpty: options.raiseOnEmpty ?? this._raiseOnEmpty,\n            confThreshold: Math.max(floor, this._spec.confThreshold),\n            classes: options.classes,\n            path,\n        });\n\n        const origShape: readonly [number, number] = [original.height, original.width];\n        timer.stage(\"postprocess\");\n        return [\n            new DetectClassifyResults(\n                bulkBoxes(detections, origShape),\n                detections,\n                this._names,\n                this._classifierNames,\n                original,\n                origShape,\n                path,\n                timer.speed(),\n            ),\n        ];\n    }\n\n    /**\n     * Letterbox the image and build the graph's feeds.\n     *\n     * The detector input runs through {@link LetterboxPipeline}, which fuses the\n     * resize, the padding and the HWC-to-CHW float conversion into one\n     * `drawImage` plus one readback, and reuses its output buffer between frames.\n     * That buffer goes straight to ONNX Runtime, so `_pipeline.release()` must not\n     * be called until the run resolves.\n     *\n     * A pipeline fused with `cropSource: \"original\"` also takes the untouched\n     * image as a second input, plus the scale and padding of the letterbox — that\n     * is what lets the graph undo the letterbox transform internally and crop at\n     * native resolution instead of from the downscaled copy. That one is **not**\n     * letterboxed by definition, so it does not go through the fused path.\n     */\n    private _preprocess(image: RGBImage): {\n        feeds: Record<string, ort.Tensor>;\n        scale: number;\n        padLeft: number;\n        padTop: number;\n    } {\n        const [width, height] = this._spec.inputSize;\n        const boxed = this._pipeline.run(image);\n        const feeds: Record<string, ort.Tensor> = {\n            [INPUT_IMAGE]: toFloat32Tensor(boxed.data, [1, 3, height, width]),\n        };\n\n        if (this._spec.needsSourceImage) {\n            feeds[INPUT_SOURCE] = tensorOf(image);\n            feeds[INPUT_SCALE] = toFloat32Tensor(new Float32Array([boxed.scale]), [1]);\n            feeds[INPUT_PAD] = toFloat32Tensor(\n                new Float32Array([boxed.padLeft, boxed.padTop]),\n                [2],\n            );\n        }\n        return { feeds, scale: boxed.scale, padLeft: boxed.padLeft, padTop: boxed.padTop };\n    }\n\n    /**\n     * Map one letterboxed xyxy row back onto the original image.\n     *\n     * The graph always reports boxes in the detector's letterboxed pixel space,\n     * whichever crop source it was fused with, so both sources agree here.\n     */\n    private _toOriginal(\n        boxes: Float32Array,\n        row: number,\n        context: {\n            scale: number;\n            padLeft: number;\n            padTop: number;\n            original: RGBImage;\n        },\n    ): BoundingBox {\n        const { scale, padLeft, padTop, original } = context;\n        const at = (offset: number): number => boxes[row * 4 + offset] ?? 0;\n        const clampX = (value: number): number => Math.min(Math.max(value, 0), original.width);\n        const clampY = (value: number): number => Math.min(Math.max(value, 0), original.height);\n        return new BoundingBox(\n            clampX((at(0) - padLeft) / scale),\n            clampY((at(1) - padTop) / scale),\n            clampX((at(2) - padLeft) / scale),\n            clampY((at(3) - padTop) / scale),\n        );\n    }\n\n    /**\n     * Turn one row of the classifier output into a result object.\n     *\n     * @param row The output row for this detection.\n     * @param image The crop the row describes, carried so callers can display\n     *   what was classified.\n     * @param k Optional truncation of the probability list.\n     */\n    private _classify(\n        row: Float32Array,\n        image: RGBImage,\n        k: number | undefined,\n    ): ClassificationResult {\n        const scores = this._spec.applySoftmax ? softmax(row) : row;\n        const { indices, values } = topK(scores, k ?? null);\n        const probabilities: ClassProbability[] = [];\n        for (let i = 0; i < indices.length; i++) {\n            const classId = indices[i] ?? 0;\n            const probability = values[i] ?? 0;\n            const className = this._classifierLabels[classId] ?? `class_${classId}`;\n            probabilities.push({\n                classId,\n                className,\n                probability,\n                cls: classId,\n                name: className,\n                conf: probability,\n            });\n        }\n        const top = probabilities[0] ?? {\n            classId: 0,\n            className: \"class_0\",\n            probability: 0,\n            cls: 0,\n            name: \"class_0\",\n            conf: 0,\n        };\n        return {\n            classId: top.classId,\n            className: top.className,\n            confidence: top.probability,\n            cls: top.classId,\n            name: top.className,\n            conf: top.probability,\n            image,\n            probabilities,\n        };\n    }\n}\n\n/**\n * Read the classifier stage's class count off the `probs` output shape.\n *\n * @param session The loaded pipeline session.\n * @returns The class count, or `null` when the graph leaves that axis dynamic\n *   or declares no `probs` output — in which case label resolution falls back\n *   to whatever the fusion recorded.\n */\nfunction classifierClasses(session: OrtSession): number | null {\n    const index = session.outputNames.indexOf(OUTPUT_PROBS);\n    if (index < 0) return null;\n    const shape = session.outputShapes[index];\n    if (shape === undefined || shape.length === 0) return null;\n    return shape[shape.length - 1] ?? null;\n}\n\n/**\n * Build a class id → name record from an ordered label list.\n *\n * @param labels Labels indexed by class id.\n * @returns The equivalent record.\n */\nfunction indexNames(labels: readonly string[]): Readonly<Record<number, string>> {\n    const names: Record<number, string> = {};\n    for (let i = 0; i < labels.length; i++) names[i] = labels[i] as string;\n    return names;\n}\n\n/**\n * Fetch an output tensor by name.\n *\n * @param outputs The run's results.\n * @param name The output's name in the pipeline contract.\n * @returns The tensor.\n * @throws {@link FusionError} when the graph does not carry that output, which\n *   means the file is not a pipeline this version can drive.\n */\nfunction output(outputs: Record<string, ort.Tensor>, name: string): ort.Tensor {\n    const tensor = outputs[name];\n    if (tensor === undefined) {\n        throw new FusionError(`The fused pipeline is missing its '${name}' output.`);\n    }\n    return tensor;\n}\n\n/**\n * Fetch a float output by name.\n *\n * @param outputs The run's results.\n * @param name The output's name in the pipeline contract.\n * @returns The output's data.\n * @throws {@link FusionError} when the graph does not carry that output.\n */\nfunction floats(outputs: Record<string, ort.Tensor>, name: string): Float32Array {\n    return output(outputs, name).data as Float32Array;\n}\n\n/**\n * Fetch an integer output by name, normalizing ORT's 64-bit representation.\n *\n * ONNX Runtime Web returns `int64` tensors as `BigInt64Array`, whose values do\n * not compare or index like numbers. Class ids and detection counts are always\n * small, so widening them to `number` here is lossless and keeps every caller\n * free of `BigInt` handling.\n *\n * @param outputs The run's results.\n * @param name The output's name in the pipeline contract.\n * @returns The output's values as plain numbers.\n * @throws {@link FusionError} when the graph does not carry that output.\n */\nfunction integers(outputs: Record<string, ort.Tensor>, name: string): number[] {\n    const data = output(outputs, name).data as BigInt64Array | Int32Array | Float32Array;\n    const values: number[] = [];\n    for (let i = 0; i < data.length; i++) values.push(Number(data[i]));\n    return values;\n}\n\n/**\n * Convert an image to the `[1, 3, H, W]` float32 tensor the graph expects.\n *\n * @param image The image to convert.\n * @returns The batched CHW tensor, scaled to `[0, 1]`.\n */\nfunction tensorOf(image: RGBImage): ort.Tensor {\n    const chw = toCHW(toFloat32(image), image.width, image.height, 3);\n    return toFloat32Tensor(chw, [1, 3, image.height, image.width]);\n}\n\n/**\n * Cut the box region out of the original image.\n *\n * @param image The source image.\n * @param bbox The box, in original-image pixel coordinates.\n * @returns The cropped region, or a zero-sized image for a box with no area.\n */\nfunction crop(image: RGBImage, bbox: BoundingBox): RGBImage {\n    const [rawX1, rawY1, rawX2, rawY2] = bbox.asIntXyxy();\n    const x1 = Math.max(0, rawX1);\n    const y1 = Math.max(0, rawY1);\n    const x2 = Math.min(image.width, rawX2);\n    const y2 = Math.min(image.height, rawY2);\n    if (x2 <= x1 || y2 <= y1) return new RGBImage(new Uint8Array(0), 0, 0);\n\n    const width = x2 - x1;\n    const height = y2 - y1;\n    const out = new Uint8Array(width * height * 3);\n    for (let row = 0; row < height; row++) {\n        const offset = ((y1 + row) * image.width + x1) * 3;\n        out.set(image.data.subarray(offset, offset + width * 3), row * width * 3);\n    }\n    return new RGBImage(out, width, height);\n}\n\n/**\n * Assemble one detection, filling the Ultralytics-style aliases.\n *\n * @param classId Detector class index.\n * @param className Detector class name.\n * @param confidence Detection score.\n * @param bbox Box in original-image coordinates.\n * @param croppedImage The region the box covers.\n * @param classification What the classification stage said about that region.\n * @returns The detection object.\n */\nfunction detection(\n    classId: number,\n    className: string,\n    confidence: number,\n    bbox: BoundingBox,\n    croppedImage: RGBImage,\n    classification: ClassificationResult,\n): DetectionResult {\n    return {\n        classId,\n        className,\n        confidence,\n        bbox,\n        cls: classId,\n        name: className,\n        conf: confidence,\n        box: bbox,\n        croppedImage,\n        classification,\n    };\n}\n\n/**\n * Assemble the bulk-array `Boxes` view from per-instance detections.\n *\n * @param detections The surviving detections.\n * @param origShape `[height, width]` of the original image.\n * @returns The bulk view, empty when nothing survived.\n */\nfunction bulkBoxes(\n    detections: readonly DetectionResult[],\n    origShape: readonly [number, number],\n): Boxes {\n    const xyxy = new Float32Array(detections.length * 4);\n    const cls = new Int32Array(detections.length);\n    const conf = new Float32Array(detections.length);\n    detections.forEach((entry, index) => {\n        const [x1, y1, x2, y2] = entry.bbox.xyxy;\n        xyxy[index * 4] = x1;\n        xyxy[index * 4 + 1] = y1;\n        xyxy[index * 4 + 2] = x2;\n        xyxy[index * 4 + 3] = y2;\n        cls[index] = entry.classId;\n        conf[index] = entry.confidence;\n    });\n    return new Boxes(xyxy, cls, conf, origShape);\n}\n"],"mappings":";;;;;;;;;;;;;AAqGA,IAAa,IAAb,MAAa,UAAuB,EAAW;CAGtB;CACA;CACA;CACA;CACA;CACA;CAPrB,YACI,GACA,GACA,GACA,GACA,GACA,GACA,GACF;EADmB,AAEjB,MAAM,CAAO,GAPI,KAAA,QAAA,GACA,KAAA,UAAA,GACA,KAAA,SAAA,GACA,KAAA,oBAAA,GACA,KAAA,mBAAA,GACA,KAAA,gBAAA;CAGrB;CAEA,iBAAmD;CAanD,MAAM,OAAO,IAAe,GAAkB;EAC1C,IAAM,CAAC,GAAO,KAAU,KAAK,MAAM;EACnC,KAAK,IAAI,IAAI,GAAG,IAAI,GAAM,KAAK;GAC3B,IAAM,IAAoC,GACrC,IAAc,EAAgB,EAAe,GAAO,CAAM,GAAG;IAC1D;IACA;IACA;IACA;GACJ,CAAC,EACL;GAWA,AAVI,KAAK,MAAM,qBACX,EAAM,KAAgB,EAAgB,EAAe,GAAO,CAAM,GAAG;IACjE;IACA;IACA;IACA;GACJ,CAAC,GACD,EAAM,KAAe,EAAgB,IAAI,aAAa,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAC/D,EAAM,KAAa,EAAgB,IAAI,aAAa,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,IAEpE,MAAM,KAAK,SAAS,IAAI,CAAK;EACjC;CACJ;CASA,IAAY,YAA+B;EACvC,IAAI,KAAK,mBAAmB,MAAM;GAC9B,IAAM,CAAC,GAAO,KAAU,KAAK,MAAM;GACnC,KAAK,iBAAiB,IAAI,EAAkB,GAAO,CAAM;EAC7D;EACA,OAAO,KAAK;CAChB;CAWA,aAAa,OACT,GACA,IAAiC,CAAC,GACX;EACvB,IAAM,IAAU,MAAM,EAAW,OAAO,GAAO,CAAO,GAChD,IAAO,EAAe,EAAQ,QAAQ;EAC5C,IAAI,MAAS,MACT,MAAM,IAAI,EACN,mNAGJ;EAGJ,IAAM,IAAS,EAAc,EAAQ,UAAU,EAAK,iBAAiB,MAAM,GACrE,IAAmB,EAAc,EAAQ,oBAAoB,EAAK,iBAAiB,EACrF,YAAY,EAAkB,CAAO,KAAK,KAAA,EAC9C,CAAC;EACD,OAAO,IAAI,EACP,GACA,GACA,GACA,EAAW,CAAM,GACjB,GACA,EAAW,CAAgB,GAC3B,EAAQ,gBAAgB,EAC5B;CACJ;CAGA,IAAI,OAAmB;EACnB,OAAO,KAAK;CAChB;CAGA,IAAI,YAAuC;EACvC,OAAO,KAAK,MAAM;CACtB;CAGA,IAAI,SAA4B;EAC5B,OAAO,KAAK;CAChB;CAGA,IAAI,QAA0C;EAC1C,OAAO,KAAK;CAChB;CAGA,IAAI,mBAAsC;EACtC,OAAO,KAAK;CAChB;CAGA,IAAI,kBAAoD;EACpD,OAAO,KAAK;CAChB;CAQA,MAAM,KACF,GACA,IAAwC,CAAC,GACT;EAChC,OAAO,KAAK,QAAQ,GAAO,CAAO;CACtC;CAUA,MAAM,QACF,GACA,IAAwC,CAAC,GACT;EAChC,IAAM,IAAQ,IAAI,EAAW,GACvB,IAAO,OAAO,KAAU,WAAW,IAAQ,MAC3C,IAAW,MAAM,EAAU,CAAK;EACtC,EAAM,MAAM,MAAM;EAClB,IAAM,EAAE,UAAO,UAAO,YAAS,cAAW,KAAK,YAAY,CAAQ;EACnE,EAAM,MAAM,YAAY;EACxB,IAAM,IAAU,MAAM,KAAK,SAAS,IAAI,CAAK;EAE7C,AADA,KAAK,UAAU,QAAQ,GACvB,EAAM,MAAM,WAAW;EAEvB,IAAM,IAAc,EAAO,GAAS,CAAY,GAC1C,IAAQ,EAAO,GAAS,CAAY,GACpC,IAAS,EAAO,GAAS,CAAa,GACtC,IAAU,EAAS,GAAS,CAAc,GAC1C,IAAQ,EAAY,MACpB,IAAW,EAAS,GAAA,gBAA8B,CAAC,CAAC,MAAM,GAC1D,IAAO,KAAK,IAAI,GAAU,KAAK,MAAM,EAAM,SAAS,CAAC,CAAC,GACtD,IAAa,EAAY,KAAK,EAAY,KAAK,SAAS,MAAM,GAE9D,IAAU,EAAQ,YAAY,KAAA,IAAY,OAAO,IAAI,IAAI,EAAQ,OAAO,GACxE,IAAQ,EAAQ,iBAAiB,GACjC,IAAgC,CAAC;EACvC,KAAK,IAAI,IAAM,GAAG,IAAM,GAAM,KAAO;GACjC,IAAM,IAAU,EAAQ,MAAQ,GAC1B,IAAa,EAAO,MAAQ;GAClC,IAAI,IAAa,KAAU,MAAY,QAAQ,CAAC,EAAQ,IAAI,CAAO,GAAI;GAEvE,IAAM,IAAO,KAAK,YAAY,GAAO,GAAK;IAAE;IAAO;IAAS;IAAQ;GAAS,CAAC,GACxE,IAAU,EAAK,GAAU,CAAI;GACnC,EAAW,KACP,EACI,GACA,KAAK,OAAO,MAAY,SAAS,KACjC,GACA,GACA,GACA,KAAK,UACD,EAAM,SAAS,IAAM,IAAa,IAAM,KAAK,CAAU,GACvD,GACA,EAAQ,IACZ,CACJ,CACJ;EACJ;EAEA,EAAkB,EAAW,QAAQ;GACjC,cAAc,EAAQ,gBAAgB,KAAK;GAC3C,eAAe,KAAK,IAAI,GAAO,KAAK,MAAM,aAAa;GACvD,SAAS,EAAQ;GACjB;EACJ,CAAC;EAED,IAAM,IAAuC,CAAC,EAAS,QAAQ,EAAS,KAAK;EAE7E,OADA,EAAM,MAAM,aAAa,GAClB,CACH,IAAI,EACA,EAAU,GAAY,CAAS,GAC/B,GACA,KAAK,QACL,KAAK,kBACL,GACA,GACA,GACA,EAAM,MAAM,CAChB,CACJ;CACJ;CAiBA,YAAoB,GAKlB;EACE,IAAM,CAAC,GAAO,KAAU,KAAK,MAAM,WAC7B,IAAQ,KAAK,UAAU,IAAI,CAAK,GAChC,IAAoC,GACrC,IAAc,EAAgB,EAAM,MAAM;GAAC;GAAG;GAAG;GAAQ;EAAK,CAAC,EACpE;EAUA,OARI,KAAK,MAAM,qBACX,EAAM,KAAgB,EAAS,CAAK,GACpC,EAAM,KAAe,EAAgB,IAAI,aAAa,CAAC,EAAM,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,GACzE,EAAM,KAAa,EACf,IAAI,aAAa,CAAC,EAAM,SAAS,EAAM,MAAM,CAAC,GAC9C,CAAC,CAAC,CACN,IAEG;GAAE;GAAO,OAAO,EAAM;GAAO,SAAS,EAAM;GAAS,QAAQ,EAAM;EAAO;CACrF;CAQA,YACI,GACA,GACA,GAMW;EACX,IAAM,EAAE,UAAO,YAAS,WAAQ,gBAAa,GACvC,KAAM,MAA2B,EAAM,IAAM,IAAI,MAAW,GAC5D,KAAU,MAA0B,KAAK,IAAI,KAAK,IAAI,GAAO,CAAC,GAAG,EAAS,KAAK,GAC/E,KAAU,MAA0B,KAAK,IAAI,KAAK,IAAI,GAAO,CAAC,GAAG,EAAS,MAAM;EACtF,OAAO,IAAI,EACP,GAAQ,EAAG,CAAC,IAAI,KAAW,CAAK,GAChC,GAAQ,EAAG,CAAC,IAAI,KAAU,CAAK,GAC/B,GAAQ,EAAG,CAAC,IAAI,KAAW,CAAK,GAChC,GAAQ,EAAG,CAAC,IAAI,KAAU,CAAK,CACnC;CACJ;CAUA,UACI,GACA,GACA,GACoB;EACpB,IAAM,IAAS,KAAK,MAAM,eAAe,EAAQ,CAAG,IAAI,GAClD,EAAE,YAAS,cAAW,EAAK,GAAQ,KAAK,IAAI,GAC5C,IAAoC,CAAC;EAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,EAAQ,QAAQ,KAAK;GACrC,IAAM,IAAU,EAAQ,MAAM,GACxB,IAAc,EAAO,MAAM,GAC3B,IAAY,KAAK,kBAAkB,MAAY,SAAS;GAC9D,EAAc,KAAK;IACf;IACA;IACA;IACA,KAAK;IACL,MAAM;IACN,MAAM;GACV,CAAC;EACL;EACA,IAAM,IAAM,EAAc,MAAM;GAC5B,SAAS;GACT,WAAW;GACX,aAAa;GACb,KAAK;GACL,MAAM;GACN,MAAM;EACV;EACA,OAAO;GACH,SAAS,EAAI;GACb,WAAW,EAAI;GACf,YAAY,EAAI;GAChB,KAAK,EAAI;GACT,MAAM,EAAI;GACV,MAAM,EAAI;GACV;GACA;EACJ;CACJ;AACJ;AAUA,SAAS,EAAkB,GAAoC;CAC3D,IAAM,IAAQ,EAAQ,YAAY,QAAQ,CAAY;CACtD,IAAI,IAAQ,GAAG,OAAO;CACtB,IAAM,IAAQ,EAAQ,aAAa;CAEnC,OADI,MAAU,KAAA,KAAa,EAAM,WAAW,IAAU,OAC/C,EAAM,EAAM,SAAS,MAAM;AACtC;AAQA,SAAS,EAAW,GAA6D;CAC7E,IAAM,IAAgC,CAAC;CACvC,KAAK,IAAI,IAAI,GAAG,IAAI,EAAO,QAAQ,KAAK,EAAM,KAAK,EAAO;CAC1D,OAAO;AACX;AAWA,SAAS,EAAO,GAAqC,GAA0B;CAC3E,IAAM,IAAS,EAAQ;CACvB,IAAI,MAAW,KAAA,GACX,MAAM,IAAI,EAAY,sCAAsC,EAAK,UAAU;CAE/E,OAAO;AACX;AAUA,SAAS,EAAO,GAAqC,GAA4B;CAC7E,OAAO,EAAO,GAAS,CAAI,CAAC,CAAC;AACjC;AAeA,SAAS,EAAS,GAAqC,GAAwB;CAC3E,IAAM,IAAO,EAAO,GAAS,CAAI,CAAC,CAAC,MAC7B,IAAmB,CAAC;CAC1B,KAAK,IAAI,IAAI,GAAG,IAAI,EAAK,QAAQ,KAAK,EAAO,KAAK,OAAO,EAAK,EAAE,CAAC;CACjE,OAAO;AACX;AAQA,SAAS,EAAS,GAA6B;CAC3C,IAAM,IAAM,EAAM,EAAU,CAAK,GAAG,EAAM,OAAO,EAAM,QAAQ,CAAC;CAChE,OAAO,EAAgB,GAAK;EAAC;EAAG;EAAG,EAAM;EAAQ,EAAM;CAAK,CAAC;AACjE;AASA,SAAS,EAAK,GAAiB,GAA6B;CACxD,IAAM,CAAC,GAAO,GAAO,GAAO,KAAS,EAAK,UAAU,GAC9C,IAAK,KAAK,IAAI,GAAG,CAAK,GACtB,IAAK,KAAK,IAAI,GAAG,CAAK,GACtB,IAAK,KAAK,IAAI,EAAM,OAAO,CAAK,GAChC,IAAK,KAAK,IAAI,EAAM,QAAQ,CAAK;CACvC,IAAI,KAAM,KAAM,KAAM,GAAI,OAAO,IAAI,kBAAS,IAAI,WAAY,GAAG,GAAG,CAAC;CAErE,IAAM,IAAQ,IAAK,GACb,IAAS,IAAK,GACd,IAAM,IAAI,WAAW,IAAQ,IAAS,CAAC;CAC7C,KAAK,IAAI,IAAM,GAAG,IAAM,GAAQ,KAAO;EACnC,IAAM,MAAW,IAAK,KAAO,EAAM,QAAQ,KAAM;EACjD,EAAI,IAAI,EAAM,KAAK,SAAS,GAAQ,IAAS,IAAQ,CAAC,GAAG,IAAM,IAAQ,CAAC;CAC5E;CACA,OAAO,IAAI,EAAS,GAAK,GAAO,CAAM;AAC1C;AAaA,SAAS,EACL,GACA,GACA,GACA,GACA,GACA,GACe;CACf,OAAO;EACH;EACA;EACA;EACA;EACA,KAAK;EACL,MAAM;EACN,MAAM;EACN,KAAK;EACL;EACA;CACJ;AACJ;AASA,SAAS,EACL,GACA,GACK;CACL,IAAM,IAAO,IAAI,aAAa,EAAW,SAAS,CAAC,GAC7C,IAAM,IAAI,WAAW,EAAW,MAAM,GACtC,IAAO,IAAI,aAAa,EAAW,MAAM;CAU/C,OATA,EAAW,SAAS,GAAO,MAAU;EACjC,IAAM,CAAC,GAAI,GAAI,GAAI,KAAM,EAAM,KAAK;EAMpC,AALA,EAAK,IAAQ,KAAK,GAClB,EAAK,IAAQ,IAAI,KAAK,GACtB,EAAK,IAAQ,IAAI,KAAK,GACtB,EAAK,IAAQ,IAAI,KAAK,GACtB,EAAI,KAAS,EAAM,SACnB,EAAK,KAAS,EAAM;CACxB,CAAC,GACM,IAAI,EAAM,GAAM,GAAK,GAAM,CAAS;AAC/C"}
+{"version":3,"file":"detectClassify.js","names":[],"sources":["../../../src/vision/tasks/detectClassify.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Run a fused detect→classify pipeline in the browser.\n *\n * The file this loads was built by the Python SDK's `ort_vision_sdk.compose`,\n * and it already contains both models plus the crop-and-resize bridge between\n * them. That matters far more in a tab than on a server: two models mean two\n * `.onnx` downloads, two WASM/WebGPU session initializations, and a per-crop\n * round trip through JavaScript to slice, resize and restack the regions before\n * the second model can see them. A fused pipeline has one download, one session\n * and no round trip — the crops are produced and consumed inside the graph.\n */\n\nimport type * as ort from \"onnxruntime-web\";\n\nimport { FusionError } from \"../core/exceptions\";\nimport { type ModelSource, type OrtSessionOptions, OrtSession } from \"../core/session\";\nimport { SpeedTimer } from \"../core/timing\";\nimport {\n    INPUT_IMAGE,\n    INPUT_PAD,\n    INPUT_SCALE,\n    INPUT_SOURCE,\n    OUTPUT_BOXES,\n    OUTPUT_CLASSES,\n    OUTPUT_NUM_DETECTIONS,\n    OUTPUT_MASKS,\n    OUTPUT_PROBS,\n    OUTPUT_SCORES,\n    type FusionSpec,\n    readFusionSpec,\n} from \"../fusion\";\nimport { type ImageInput, loadImageSource } from \"../io/image\";\nimport { type LabelSpec, resolveLabels } from \"../labels\";\nimport { softmax, topK } from \"../postprocess/classification\";\nimport { asFloat32Array } from \"../core/dtypes\";\nimport { toCHW, toFloat32, toFloat32Tensor } from \"../preprocess/image\";\nimport { LetterboxPipeline, zeroTensorData } from \"../preprocess/pipeline\";\nimport { Boxes, DetectClassifyResults } from \"../results\";\nimport type { RGBImage } from \"../types\";\nimport {\n    BoundingBox,\n    Mask,\n    type ClassProbability,\n    type ClassificationResult,\n    type DetectionResult,\n} from \"../types\";\nimport { VisionTask, cropToBox, requireDetections } from \"./base\";\n\nexport interface DetectClassifyOptions extends OrtSessionOptions {\n    /**\n     * Class label spec for the **detection** stage — see {@link resolveLabels}.\n     * Defaults to the names recorded at fusion time, falling back to the COCO\n     * 80-class preset when the fusion recorded none.\n     */\n    readonly labels?: LabelSpec;\n    /**\n     * Class label spec for the **classification** stage. Defaults to the recorded\n     * names, falling back to generated `class_<id>` names.\n     */\n    readonly classifierLabels?: LabelSpec;\n    /**\n     * If `true`, a run that finds nothing throws {@link NoDetectionsError}\n     * instead of returning an empty envelope. Default `false`, because looking\n     * and finding nothing is a successful inference. Turn it on when an empty\n     * result means the surrounding pipeline should stop rather than carry on with\n     * zero rows. Can be overridden per `predict` call.\n     */\n    readonly raiseOnEmpty?: boolean;\n}\n\nexport interface DetectClassifyPredictOptions {\n    /**\n     * Drop detections scoring below this. The graph's own NMS threshold was fixed\n     * at fusion time and cannot be lowered here — this only filters further.\n     */\n    readonly confThreshold?: number;\n    /** If set, keep only detections whose detector `classId` is in this list. */\n    readonly classes?: readonly number[];\n    /** Truncate each detection's `classification.probabilities` to its top-k entries. */\n    readonly topK?: number;\n    /** Override the constructor's `raiseOnEmpty` setting for this call. */\n    readonly raiseOnEmpty?: boolean;\n}\n\n/**\n * Detector and classifier running as a single ONNX model.\n *\n * Everything the pipeline needs to know about itself — the resolution to\n * letterbox to, whether it wants the full-resolution image as well, whether its\n * classifier output still needs a softmax, the class names of both stages — was\n * written into the file at fusion time and is read back here. Nothing is\n * restated on the JavaScript side, so nothing can drift out of step with the\n * Python side that built it.\n *\n * @example\n * ```typescript\n * const pipeline = await DetectClassify.create(\"/models/pipeline.onnx\");\n * const result = (await pipeline.predict(\"/images/flock.jpg\"))[0];\n * for (const detection of result) {\n *   console.log(detection.name, detection.conf, detection.classification?.name);\n * }\n * ```\n */\nexport class DetectClassify extends VisionTask {\n    private constructor(\n        session: OrtSession,\n        private readonly _spec: FusionSpec,\n        private readonly _labels: readonly string[],\n        private readonly _names: Readonly<Record<number, string>>,\n        private readonly _classifierLabels: readonly string[],\n        private readonly _classifierNames: Readonly<Record<number, string>>,\n        private readonly _raiseOnEmpty: boolean,\n    ) {\n        super(session);\n    }\n\n    private _pipelineCache: LetterboxPipeline | null = null;\n\n    /**\n     * Run the model once on zero-filled inputs, paying one-time costs up front.\n     *\n     * Worth more here than on a single-stage task: a fused pipeline is two models\n     * plus the bridge in one graph, so the first inference compiles shaders for\n     * all of it. Calling this while a loading spinner is still up moves that cost\n     * somewhere the user is already waiting.\n     *\n     * @param runs How many warm-up inferences to run. One is enough for WASM;\n     *   WebGPU sometimes settles on the second.\n     */\n    async warmup(runs: number = 1): Promise<void> {\n        const [width, height] = this._spec.inputSize;\n        for (let i = 0; i < runs; i++) {\n            const feeds: Record<string, ort.Tensor> = {\n                [INPUT_IMAGE]: toFloat32Tensor(zeroTensorData(width, height), [\n                    1,\n                    3,\n                    height,\n                    width,\n                ]),\n            };\n            if (this._spec.needsSourceImage) {\n                feeds[INPUT_SOURCE] = toFloat32Tensor(zeroTensorData(width, height), [\n                    1,\n                    3,\n                    height,\n                    width,\n                ]);\n                feeds[INPUT_SCALE] = toFloat32Tensor(new Float32Array([1]), [1]);\n                feeds[INPUT_PAD] = toFloat32Tensor(new Float32Array([0, 0]), [2]);\n            }\n            await this._session.run(feeds);\n        }\n    }\n\n    /**\n     * The fused preprocessing pipeline, built on first use.\n     *\n     * Lazily, because constructing it allocates canvases: a pipeline built in an\n     * environment without a canvas implementation stays constructible, and only\n     * fails if it is actually asked to preprocess something.\n     */\n    private get _pipeline(): LetterboxPipeline {\n        if (this._pipelineCache === null) {\n            const [width, height] = this._spec.inputSize;\n            this._pipelineCache = new LetterboxPipeline(width, height);\n        }\n        return this._pipelineCache;\n    }\n\n    /**\n     * Load a fused pipeline and resolve both label spaces.\n     *\n     * @param model The fused `.onnx` — a URL, an `ArrayBuffer`, or bytes.\n     * @param options Label overrides plus the usual session options.\n     * @throws {@link FusionError} when the model carries no pipeline metadata,\n     *   i.e. it is a plain detector or classifier rather than something\n     *   `ort_vision_sdk.compose` produced.\n     */\n    static async create(\n        model: ModelSource,\n        options: DetectClassifyOptions = {},\n    ): Promise<DetectClassify> {\n        const session = await OrtSession.create(model, options);\n        const spec = readFusionSpec(session.metadata);\n        if (spec === null) {\n            throw new FusionError(\n                \"This model carries no fused-pipeline metadata, so DetectClassify cannot tell how to \" +\n                    \"drive it. Build one with ort_vision_sdk.compose.fuse_detect_classify, or load a \" +\n                    \"plain model with Detector/Classifier instead.\",\n            );\n        }\n\n        const labels = resolveLabels(options.labels ?? spec.detectorNames ?? \"coco\");\n        const classifierLabels = resolveLabels(options.classifierLabels ?? spec.classifierNames, {\n            numClasses: classifierClasses(session) ?? undefined,\n        });\n        return new DetectClassify(\n            session,\n            spec,\n            labels,\n            indexNames(labels),\n            classifierLabels,\n            indexNames(classifierLabels),\n            options.raiseOnEmpty ?? false,\n        );\n    }\n\n    /** The pipeline configuration recorded in the model at fusion time. */\n    get spec(): FusionSpec {\n        return this._spec;\n    }\n\n    /** The `[width, height]` the detection stage runs at. */\n    get inputSize(): readonly [number, number] {\n        return this._spec.inputSize;\n    }\n\n    /** Detector class labels indexed by class id. */\n    get labels(): readonly string[] {\n        return this._labels;\n    }\n\n    /** Detector class id → class name (matches Ultralytics' `model.names`). */\n    get names(): Readonly<Record<number, string>> {\n        return this._names;\n    }\n\n    /** Classifier class labels indexed by class id. */\n    get classifierLabels(): readonly string[] {\n        return this._classifierLabels;\n    }\n\n    /** Classifier class id → class name. */\n    get classifierNames(): Readonly<Record<number, string>> {\n        return this._classifierNames;\n    }\n\n    /**\n     * Alias for {@link predict} — call the pipeline like a torch `nn.Module`.\n     *\n     * Use as `pipeline.call(img)` since JavaScript class instances are not\n     * callable; for direct invocation, prefer `pipeline.predict(img)`.\n     */\n    async call(\n        image: ImageInput,\n        options: DetectClassifyPredictOptions = {},\n    ): Promise<DetectClassifyResults[]> {\n        return this.predict(image, options);\n    }\n\n    /**\n     * Run the pipeline on a single image.\n     *\n     * The returned envelope carries a {@link Speed} breakdown in `speed`. Its\n     * `inference` figure covers detection *and* classification, since the\n     * pipeline runs them as one graph and no boundary between them is observable\n     * from outside.\n     */\n    async predict(\n        image: ImageInput,\n        options: DetectClassifyPredictOptions = {},\n    ): Promise<DetectClassifyResults[]> {\n        const timer = new SpeedTimer();\n        const path = typeof image === \"string\" ? image : null;\n        const { image: original, canvas } = await loadImageSource(image);\n        timer.stage(\"load\");\n        const { feeds, scale, padLeft, padTop, reused } = this._preprocess(original, canvas);\n        timer.stage(\"preprocess\");\n        let outputs: Record<string, ort.Tensor>;\n        try {\n            outputs = await this._session.run(feeds);\n        } finally {\n            this._pipeline.release({ reused });\n        }\n        timer.stage(\"inference\");\n\n        const probsTensor = output(outputs, OUTPUT_PROBS);\n        const boxes = floats(outputs, OUTPUT_BOXES);\n        const scores = floats(outputs, OUTPUT_SCORES);\n        const classes = integers(outputs, OUTPUT_CLASSES);\n        const probs = asFloat32Array(probsTensor.data);\n        const reported = integers(outputs, OUTPUT_NUM_DETECTIONS)[0] ?? 0;\n        const masks = this._spec.hasMasks ? floats(outputs, OUTPUT_MASKS) : null;\n        const [cropWidth, cropHeight] = this._spec.cropSize;\n        const rows = Math.min(reported, Math.floor(boxes.length / 4));\n        const classCount = probsTensor.dims[probsTensor.dims.length - 1] ?? 0;\n\n        const allowed = options.classes === undefined ? null : new Set(options.classes);\n        const floor = options.confThreshold ?? 0;\n        const detections: DetectionResult[] = [];\n        for (let row = 0; row < rows; row++) {\n            const classId = classes[row] ?? 0;\n            const confidence = scores[row] ?? 0;\n            if (confidence < floor || (allowed !== null && !allowed.has(classId))) continue;\n\n            const bbox = this._toOriginal(boxes, row, { scale, padLeft, padTop, original });\n            const cropped = cropToBox(original, bbox);\n            detections.push(\n                detection(\n                    classId,\n                    this._names[classId] ?? `class_${classId}`,\n                    confidence,\n                    bbox,\n                    cropped,\n                    this._classify(\n                        probs.subarray(row * classCount, (row + 1) * classCount),\n                        cropped,\n                        options.topK,\n                    ),\n                    masks === null ? null : maskToBox(masks, row, cropWidth, cropHeight, cropped),\n                ),\n            );\n        }\n\n        requireDetections(detections.length, {\n            raiseOnEmpty: options.raiseOnEmpty ?? this._raiseOnEmpty,\n            confThreshold: Math.max(floor, this._spec.confThreshold),\n            classes: options.classes,\n            path,\n        });\n\n        const origShape: readonly [number, number] = [original.height, original.width];\n        timer.stage(\"postprocess\");\n        return [\n            new DetectClassifyResults(\n                bulkBoxes(detections, origShape),\n                detections,\n                this._names,\n                this._classifierNames,\n                original,\n                origShape,\n                path,\n                timer.speed(),\n            ),\n        ];\n    }\n\n    /**\n     * Letterbox the image and build the graph's feeds.\n     *\n     * The detector input runs through {@link LetterboxPipeline}, which fuses the\n     * resize, the padding and the HWC-to-CHW float conversion into one\n     * `drawImage` plus one readback, and reuses its output buffer between frames.\n     * That buffer goes straight to ONNX Runtime, so `_pipeline.release()` must not\n     * be called until the run resolves.\n     *\n     * A pipeline fused with `cropSource: \"original\"` also takes the untouched\n     * image as a second input, plus the scale and padding of the letterbox — that\n     * is what lets the graph undo the letterbox transform internally and crop at\n     * native resolution instead of from the downscaled copy. That one is **not**\n     * letterboxed by definition, so it does not go through the fused path.\n     *\n     * @param image The decoded input.\n     * @param canvas The opaque canvas it was decoded on, when there is one — see\n     *   {@link LetterboxPipeline.run}.\n     */\n    private _preprocess(\n        image: RGBImage,\n        canvas: CanvasImageSource | null,\n    ): {\n        feeds: Record<string, ort.Tensor>;\n        scale: number;\n        padLeft: number;\n        padTop: number;\n        reused: boolean;\n    } {\n        const [width, height] = this._spec.inputSize;\n        const boxed = this._pipeline.run(image, canvas);\n        const feeds: Record<string, ort.Tensor> = {\n            [INPUT_IMAGE]: toFloat32Tensor(boxed.data, [1, 3, height, width]),\n        };\n\n        if (this._spec.needsSourceImage) {\n            feeds[INPUT_SOURCE] = tensorOf(image);\n            feeds[INPUT_SCALE] = toFloat32Tensor(new Float32Array([boxed.scale]), [1]);\n            feeds[INPUT_PAD] = toFloat32Tensor(\n                new Float32Array([boxed.padLeft, boxed.padTop]),\n                [2],\n            );\n        }\n        return {\n            feeds,\n            scale: boxed.scale,\n            padLeft: boxed.padLeft,\n            padTop: boxed.padTop,\n            reused: boxed.reused,\n        };\n    }\n\n    /**\n     * Map one letterboxed xyxy row back onto the original image.\n     *\n     * The graph always reports boxes in the detector's letterboxed pixel space,\n     * whichever crop source it was fused with, so both sources agree here.\n     */\n    private _toOriginal(\n        boxes: Float32Array,\n        row: number,\n        context: {\n            scale: number;\n            padLeft: number;\n            padTop: number;\n            original: RGBImage;\n        },\n    ): BoundingBox {\n        const { scale, padLeft, padTop, original } = context;\n        const at = (offset: number): number => boxes[row * 4 + offset] ?? 0;\n        const clampX = (value: number): number => Math.min(Math.max(value, 0), original.width);\n        const clampY = (value: number): number => Math.min(Math.max(value, 0), original.height);\n        return new BoundingBox(\n            clampX((at(0) - padLeft) / scale),\n            clampY((at(1) - padTop) / scale),\n            clampX((at(2) - padLeft) / scale),\n            clampY((at(3) - padTop) / scale),\n        );\n    }\n\n    /**\n     * Turn one row of the classifier output into a result object.\n     *\n     * @param row The output row for this detection.\n     * @param image The crop the row describes, carried so callers can display\n     *   what was classified.\n     * @param k Optional truncation of the probability list.\n     */\n    private _classify(\n        row: Float32Array,\n        image: RGBImage,\n        k: number | undefined,\n    ): ClassificationResult {\n        const scores = this._spec.applySoftmax ? softmax(row) : row;\n        const { indices, values } = topK(scores, k ?? null);\n        const probabilities: ClassProbability[] = [];\n        for (let i = 0; i < indices.length; i++) {\n            const classId = indices[i] ?? 0;\n            const probability = values[i] ?? 0;\n            const className = this._classifierLabels[classId] ?? `class_${classId}`;\n            probabilities.push({\n                classId,\n                className,\n                probability,\n                cls: classId,\n                name: className,\n                conf: probability,\n            });\n        }\n        const top = probabilities[0] ?? {\n            classId: 0,\n            className: \"class_0\",\n            probability: 0,\n            cls: 0,\n            name: \"class_0\",\n            conf: 0,\n        };\n        return {\n            classId: top.classId,\n            className: top.className,\n            confidence: top.probability,\n            cls: top.classId,\n            name: top.className,\n            conf: top.probability,\n            image,\n            probabilities,\n        };\n    }\n}\n\n/**\n * Read the classifier stage's class count off the `probs` output shape.\n *\n * @param session The loaded pipeline session.\n * @returns The class count, or `null` when the graph leaves that axis dynamic\n *   or declares no `probs` output — in which case label resolution falls back\n *   to whatever the fusion recorded.\n */\nfunction classifierClasses(session: OrtSession): number | null {\n    const index = session.outputNames.indexOf(OUTPUT_PROBS);\n    if (index < 0) return null;\n    const shape = session.outputShapes[index];\n    if (shape === undefined || shape.length === 0) return null;\n    return shape[shape.length - 1] ?? null;\n}\n\n/**\n * Build a class id → name record from an ordered label list.\n *\n * @param labels Labels indexed by class id.\n * @returns The equivalent record.\n */\nfunction indexNames(labels: readonly string[]): Readonly<Record<number, string>> {\n    const names: Record<number, string> = {};\n    for (let i = 0; i < labels.length; i++) names[i] = labels[i] as string;\n    return names;\n}\n\n/**\n * Fetch an output tensor by name.\n *\n * @param outputs The run's results.\n * @param name The output's name in the pipeline contract.\n * @returns The tensor.\n * @throws {@link FusionError} when the graph does not carry that output, which\n *   means the file is not a pipeline this version can drive.\n */\nfunction output(outputs: Record<string, ort.Tensor>, name: string): ort.Tensor {\n    const tensor = outputs[name];\n    if (tensor === undefined) {\n        throw new FusionError(`The fused pipeline is missing its '${name}' output.`);\n    }\n    return tensor;\n}\n\n/**\n * Fetch a float output by name.\n *\n * @param outputs The run's results.\n * @param name The output's name in the pipeline contract.\n * @returns The output's data.\n * @throws {@link FusionError} when the graph does not carry that output.\n */\nfunction floats(outputs: Record<string, ort.Tensor>, name: string): Float32Array {\n    return asFloat32Array(output(outputs, name).data);\n}\n\n/**\n * Fetch an integer output by name, normalizing ORT's 64-bit representation.\n *\n * ONNX Runtime Web returns `int64` tensors as `BigInt64Array`, whose values do\n * not compare or index like numbers. Class ids and detection counts are always\n * small, so widening them to `number` here is lossless and keeps every caller\n * free of `BigInt` handling.\n *\n * @param outputs The run's results.\n * @param name The output's name in the pipeline contract.\n * @returns The output's values as plain numbers.\n * @throws {@link FusionError} when the graph does not carry that output.\n */\nfunction integers(outputs: Record<string, ort.Tensor>, name: string): number[] {\n    const data = output(outputs, name).data as BigInt64Array | Int32Array | Float32Array;\n    const values: number[] = [];\n    for (let i = 0; i < data.length; i++) values.push(Number(data[i]));\n    return values;\n}\n\n/**\n * Convert an image to the `[1, 3, H, W]` float32 tensor the graph expects.\n *\n * @param image The image to convert.\n * @returns The batched CHW tensor, scaled to `[0, 1]`.\n */\nfunction tensorOf(image: RGBImage): ort.Tensor {\n    const chw = toCHW(toFloat32(image), image.width, image.height, 3);\n    return toFloat32Tensor(chw, [1, 3, image.height, image.width]);\n}\n\n/**\n * Assemble one detection, filling the Ultralytics-style aliases.\n *\n * @param classId Detector class index.\n * @param className Detector class name.\n * @param confidence Detection score.\n * @param bbox Box in original-image coordinates.\n * @param croppedImage The region the box covers.\n * @param classification What the classification stage said about that region.\n * @returns The detection object.\n */\nfunction detection(\n    classId: number,\n    className: string,\n    confidence: number,\n    bbox: BoundingBox,\n    croppedImage: RGBImage,\n    classification: ClassificationResult,\n    mask: Mask | null,\n): DetectionResult {\n    return {\n        classId,\n        className,\n        confidence,\n        bbox,\n        cls: classId,\n        name: className,\n        conf: confidence,\n        box: bbox,\n        croppedImage,\n        classification,\n        mask,\n    };\n}\n\n/**\n * Resample a crop-space mask onto the cropped image's own pixel grid.\n *\n * Exported for the test suite rather than for callers: it is not re-exported\n * from the package root, and a mask reaches user code already resampled, on\n * {@link DetectionResult.mask}.\n *\n * The graph computes masks at the resolution the segmenter was exported at,\n * which is rarely the size of the box in the original image. Resampling here is\n * what lets {@link DetectionResult.mask} carry the contract `Segmenter` already\n * produces, so a caller that handles one handles the other.\n *\n * Nearest-neighbour on purpose: the values are already thresholded, and\n * interpolating between them would invent edge pixels that are neither.\n *\n * @param masks The graph's `masks` output.\n * @param row Which detection row to read.\n * @param cropWidth Width of one mask in the graph's output.\n * @param cropHeight Height of one mask in the graph's output.\n * @param target The cropped image the mask must line up with.\n * @returns The mask on the crop's grid, or `null` when the box has no area.\n */\nexport function maskToBox(\n    masks: Float32Array,\n    row: number,\n    cropWidth: number,\n    cropHeight: number,\n    target: RGBImage,\n): Mask | null {\n    const { width, height } = target;\n    if (width < 1 || height < 1) return null;\n    const stride = cropWidth * cropHeight;\n    const source = masks.subarray(row * stride, (row + 1) * stride);\n    const data = new Uint8Array(width * height);\n    for (let y = 0; y < height; y++) {\n        const sourceRow = Math.min(Math.floor((y * cropHeight) / height), cropHeight - 1);\n        for (let x = 0; x < width; x++) {\n            const sourceColumn = Math.min(Math.floor((x * cropWidth) / width), cropWidth - 1);\n            data[y * width + x] =\n                (source[sourceRow * cropWidth + sourceColumn] ?? 0) > 0.5 ? 255 : 0;\n        }\n    }\n    return new Mask(data, width, height);\n}\n\n/**\n * Assemble the bulk-array `Boxes` view from per-instance detections.\n *\n * @param detections The surviving detections.\n * @param origShape `[height, width]` of the original image.\n * @returns The bulk view, empty when nothing survived.\n */\nfunction bulkBoxes(\n    detections: readonly DetectionResult[],\n    origShape: readonly [number, number],\n): Boxes {\n    const xyxy = new Float32Array(detections.length * 4);\n    const cls = new Int32Array(detections.length);\n    const conf = new Float32Array(detections.length);\n    detections.forEach((entry, index) => {\n        const [x1, y1, x2, y2] = entry.bbox.xyxy;\n        xyxy[index * 4] = x1;\n        xyxy[index * 4 + 1] = y1;\n        xyxy[index * 4 + 2] = x2;\n        xyxy[index * 4 + 3] = y2;\n        cls[index] = entry.classId;\n        conf[index] = entry.confidence;\n    });\n    return new Boxes(xyxy, cls, conf, origShape);\n}\n"],"mappings":";;;;;;;;;;;;;;AAwGA,IAAa,IAAb,MAAa,UAAuB,EAAW;CAGtB;CACA;CACA;CACA;CACA;CACA;CAPrB,YACI,GACA,GACA,GACA,GACA,GACA,GACA,GACF;EADmB,AAEjB,MAAM,CAAO,GAPI,KAAA,QAAA,GACA,KAAA,UAAA,GACA,KAAA,SAAA,GACA,KAAA,oBAAA,GACA,KAAA,mBAAA,GACA,KAAA,gBAAA;CAGrB;CAEA,iBAAmD;CAanD,MAAM,OAAO,IAAe,GAAkB;EAC1C,IAAM,CAAC,GAAO,KAAU,KAAK,MAAM;EACnC,KAAK,IAAI,IAAI,GAAG,IAAI,GAAM,KAAK;GAC3B,IAAM,IAAoC,GACrC,IAAc,EAAgB,EAAe,GAAO,CAAM,GAAG;IAC1D;IACA;IACA;IACA;GACJ,CAAC,EACL;GAWA,AAVI,KAAK,MAAM,qBACX,EAAM,KAAgB,EAAgB,EAAe,GAAO,CAAM,GAAG;IACjE;IACA;IACA;IACA;GACJ,CAAC,GACD,EAAM,KAAe,EAAgB,IAAI,aAAa,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,GAC/D,EAAM,KAAa,EAAgB,IAAI,aAAa,CAAC,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,IAEpE,MAAM,KAAK,SAAS,IAAI,CAAK;EACjC;CACJ;CASA,IAAY,YAA+B;EACvC,IAAI,KAAK,mBAAmB,MAAM;GAC9B,IAAM,CAAC,GAAO,KAAU,KAAK,MAAM;GACnC,KAAK,iBAAiB,IAAI,EAAkB,GAAO,CAAM;EAC7D;EACA,OAAO,KAAK;CAChB;CAWA,aAAa,OACT,GACA,IAAiC,CAAC,GACX;EACvB,IAAM,IAAU,MAAM,EAAW,OAAO,GAAO,CAAO,GAChD,IAAO,EAAe,EAAQ,QAAQ;EAC5C,IAAI,MAAS,MACT,MAAM,IAAI,EACN,mNAGJ;EAGJ,IAAM,IAAS,EAAc,EAAQ,UAAU,EAAK,iBAAiB,MAAM,GACrE,IAAmB,EAAc,EAAQ,oBAAoB,EAAK,iBAAiB,EACrF,YAAY,EAAkB,CAAO,KAAK,KAAA,EAC9C,CAAC;EACD,OAAO,IAAI,EACP,GACA,GACA,GACA,EAAW,CAAM,GACjB,GACA,EAAW,CAAgB,GAC3B,EAAQ,gBAAgB,EAC5B;CACJ;CAGA,IAAI,OAAmB;EACnB,OAAO,KAAK;CAChB;CAGA,IAAI,YAAuC;EACvC,OAAO,KAAK,MAAM;CACtB;CAGA,IAAI,SAA4B;EAC5B,OAAO,KAAK;CAChB;CAGA,IAAI,QAA0C;EAC1C,OAAO,KAAK;CAChB;CAGA,IAAI,mBAAsC;EACtC,OAAO,KAAK;CAChB;CAGA,IAAI,kBAAoD;EACpD,OAAO,KAAK;CAChB;CAQA,MAAM,KACF,GACA,IAAwC,CAAC,GACT;EAChC,OAAO,KAAK,QAAQ,GAAO,CAAO;CACtC;CAUA,MAAM,QACF,GACA,IAAwC,CAAC,GACT;EAChC,IAAM,IAAQ,IAAI,EAAW,GACvB,IAAO,OAAO,KAAU,WAAW,IAAQ,MAC3C,EAAE,OAAO,GAAU,cAAW,MAAM,EAAgB,CAAK;EAC/D,EAAM,MAAM,MAAM;EAClB,IAAM,EAAE,UAAO,UAAO,YAAS,WAAQ,cAAW,KAAK,YAAY,GAAU,CAAM;EACnF,EAAM,MAAM,YAAY;EACxB,IAAI;EACJ,IAAI;GACA,IAAU,MAAM,KAAK,SAAS,IAAI,CAAK;EAC3C,UAAU;GACN,KAAK,UAAU,QAAQ,EAAE,UAAO,CAAC;EACrC;EACA,EAAM,MAAM,WAAW;EAEvB,IAAM,IAAc,EAAO,GAAS,CAAY,GAC1C,IAAQ,EAAO,GAAS,CAAY,GACpC,IAAS,EAAO,GAAS,CAAa,GACtC,IAAU,EAAS,GAAS,CAAc,GAC1C,IAAQ,EAAe,EAAY,IAAI,GACvC,IAAW,EAAS,GAAA,gBAA8B,CAAC,CAAC,MAAM,GAC1D,IAAQ,KAAK,MAAM,WAAW,EAAO,GAAS,CAAY,IAAI,MAC9D,CAAC,GAAW,KAAc,KAAK,MAAM,UACrC,IAAO,KAAK,IAAI,GAAU,KAAK,MAAM,EAAM,SAAS,CAAC,CAAC,GACtD,IAAa,EAAY,KAAK,EAAY,KAAK,SAAS,MAAM,GAE9D,IAAU,EAAQ,YAAY,KAAA,IAAY,OAAO,IAAI,IAAI,EAAQ,OAAO,GACxE,IAAQ,EAAQ,iBAAiB,GACjC,IAAgC,CAAC;EACvC,KAAK,IAAI,IAAM,GAAG,IAAM,GAAM,KAAO;GACjC,IAAM,IAAU,EAAQ,MAAQ,GAC1B,IAAa,EAAO,MAAQ;GAClC,IAAI,IAAa,KAAU,MAAY,QAAQ,CAAC,EAAQ,IAAI,CAAO,GAAI;GAEvE,IAAM,IAAO,KAAK,YAAY,GAAO,GAAK;IAAE;IAAO;IAAS;IAAQ;GAAS,CAAC,GACxE,IAAU,EAAU,GAAU,CAAI;GACxC,EAAW,KACP,EACI,GACA,KAAK,OAAO,MAAY,SAAS,KACjC,GACA,GACA,GACA,KAAK,UACD,EAAM,SAAS,IAAM,IAAa,IAAM,KAAK,CAAU,GACvD,GACA,EAAQ,IACZ,GACA,MAAU,OAAO,OAAO,EAAU,GAAO,GAAK,GAAW,GAAY,CAAO,CAChF,CACJ;EACJ;EAEA,EAAkB,EAAW,QAAQ;GACjC,cAAc,EAAQ,gBAAgB,KAAK;GAC3C,eAAe,KAAK,IAAI,GAAO,KAAK,MAAM,aAAa;GACvD,SAAS,EAAQ;GACjB;EACJ,CAAC;EAED,IAAM,IAAuC,CAAC,EAAS,QAAQ,EAAS,KAAK;EAE7E,OADA,EAAM,MAAM,aAAa,GAClB,CACH,IAAI,EACA,EAAU,GAAY,CAAS,GAC/B,GACA,KAAK,QACL,KAAK,kBACL,GACA,GACA,GACA,EAAM,MAAM,CAChB,CACJ;CACJ;CAqBA,YACI,GACA,GAOF;EACE,IAAM,CAAC,GAAO,KAAU,KAAK,MAAM,WAC7B,IAAQ,KAAK,UAAU,IAAI,GAAO,CAAM,GACxC,IAAoC,GACrC,IAAc,EAAgB,EAAM,MAAM;GAAC;GAAG;GAAG;GAAQ;EAAK,CAAC,EACpE;EAUA,OARI,KAAK,MAAM,qBACX,EAAM,KAAgB,EAAS,CAAK,GACpC,EAAM,KAAe,EAAgB,IAAI,aAAa,CAAC,EAAM,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,GACzE,EAAM,KAAa,EACf,IAAI,aAAa,CAAC,EAAM,SAAS,EAAM,MAAM,CAAC,GAC9C,CAAC,CAAC,CACN,IAEG;GACH;GACA,OAAO,EAAM;GACb,SAAS,EAAM;GACf,QAAQ,EAAM;GACd,QAAQ,EAAM;EAClB;CACJ;CAQA,YACI,GACA,GACA,GAMW;EACX,IAAM,EAAE,UAAO,YAAS,WAAQ,gBAAa,GACvC,KAAM,MAA2B,EAAM,IAAM,IAAI,MAAW,GAC5D,KAAU,MAA0B,KAAK,IAAI,KAAK,IAAI,GAAO,CAAC,GAAG,EAAS,KAAK,GAC/E,KAAU,MAA0B,KAAK,IAAI,KAAK,IAAI,GAAO,CAAC,GAAG,EAAS,MAAM;EACtF,OAAO,IAAI,EACP,GAAQ,EAAG,CAAC,IAAI,KAAW,CAAK,GAChC,GAAQ,EAAG,CAAC,IAAI,KAAU,CAAK,GAC/B,GAAQ,EAAG,CAAC,IAAI,KAAW,CAAK,GAChC,GAAQ,EAAG,CAAC,IAAI,KAAU,CAAK,CACnC;CACJ;CAUA,UACI,GACA,GACA,GACoB;EACpB,IAAM,IAAS,KAAK,MAAM,eAAe,EAAQ,CAAG,IAAI,GAClD,EAAE,YAAS,cAAW,EAAK,GAAQ,KAAK,IAAI,GAC5C,IAAoC,CAAC;EAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,EAAQ,QAAQ,KAAK;GACrC,IAAM,IAAU,EAAQ,MAAM,GACxB,IAAc,EAAO,MAAM,GAC3B,IAAY,KAAK,kBAAkB,MAAY,SAAS;GAC9D,EAAc,KAAK;IACf;IACA;IACA;IACA,KAAK;IACL,MAAM;IACN,MAAM;GACV,CAAC;EACL;EACA,IAAM,IAAM,EAAc,MAAM;GAC5B,SAAS;GACT,WAAW;GACX,aAAa;GACb,KAAK;GACL,MAAM;GACN,MAAM;EACV;EACA,OAAO;GACH,SAAS,EAAI;GACb,WAAW,EAAI;GACf,YAAY,EAAI;GAChB,KAAK,EAAI;GACT,MAAM,EAAI;GACV,MAAM,EAAI;GACV;GACA;EACJ;CACJ;AACJ;AAUA,SAAS,EAAkB,GAAoC;CAC3D,IAAM,IAAQ,EAAQ,YAAY,QAAQ,CAAY;CACtD,IAAI,IAAQ,GAAG,OAAO;CACtB,IAAM,IAAQ,EAAQ,aAAa;CAEnC,OADI,MAAU,KAAA,KAAa,EAAM,WAAW,IAAU,OAC/C,EAAM,EAAM,SAAS,MAAM;AACtC;AAQA,SAAS,EAAW,GAA6D;CAC7E,IAAM,IAAgC,CAAC;CACvC,KAAK,IAAI,IAAI,GAAG,IAAI,EAAO,QAAQ,KAAK,EAAM,KAAK,EAAO;CAC1D,OAAO;AACX;AAWA,SAAS,EAAO,GAAqC,GAA0B;CAC3E,IAAM,IAAS,EAAQ;CACvB,IAAI,MAAW,KAAA,GACX,MAAM,IAAI,EAAY,sCAAsC,EAAK,UAAU;CAE/E,OAAO;AACX;AAUA,SAAS,EAAO,GAAqC,GAA4B;CAC7E,OAAO,EAAe,EAAO,GAAS,CAAI,CAAC,CAAC,IAAI;AACpD;AAeA,SAAS,EAAS,GAAqC,GAAwB;CAC3E,IAAM,IAAO,EAAO,GAAS,CAAI,CAAC,CAAC,MAC7B,IAAmB,CAAC;CAC1B,KAAK,IAAI,IAAI,GAAG,IAAI,EAAK,QAAQ,KAAK,EAAO,KAAK,OAAO,EAAK,EAAE,CAAC;CACjE,OAAO;AACX;AAQA,SAAS,EAAS,GAA6B;CAC3C,IAAM,IAAM,EAAM,EAAU,CAAK,GAAG,EAAM,OAAO,EAAM,QAAQ,CAAC;CAChE,OAAO,EAAgB,GAAK;EAAC;EAAG;EAAG,EAAM;EAAQ,EAAM;CAAK,CAAC;AACjE;AAaA,SAAS,EACL,GACA,GACA,GACA,GACA,GACA,GACA,GACe;CACf,OAAO;EACH;EACA;EACA;EACA;EACA,KAAK;EACL,MAAM;EACN,MAAM;EACN,KAAK;EACL;EACA;EACA;CACJ;AACJ;AAwBA,SAAgB,EACZ,GACA,GACA,GACA,GACA,GACW;CACX,IAAM,EAAE,UAAO,cAAW;CAC1B,IAAI,IAAQ,KAAK,IAAS,GAAG,OAAO;CACpC,IAAM,IAAS,IAAY,GACrB,IAAS,EAAM,SAAS,IAAM,IAAS,IAAM,KAAK,CAAM,GACxD,IAAO,IAAI,WAAW,IAAQ,CAAM;CAC1C,KAAK,IAAI,IAAI,GAAG,IAAI,GAAQ,KAAK;EAC7B,IAAM,IAAY,KAAK,IAAI,KAAK,MAAO,IAAI,IAAc,CAAM,GAAG,IAAa,CAAC;EAChF,KAAK,IAAI,IAAI,GAAG,IAAI,GAAO,KAAK;GAC5B,IAAM,IAAe,KAAK,IAAI,KAAK,MAAO,IAAI,IAAa,CAAK,GAAG,IAAY,CAAC;GAChF,EAAK,IAAI,IAAQ,MACZ,EAAO,IAAY,IAAY,MAAiB,KAAK,KAAM,MAAM;EAC1E;CACJ;CACA,OAAO,IAAI,EAAK,GAAM,GAAO,CAAM;AACvC;AASA,SAAS,EACL,GACA,GACK;CACL,IAAM,IAAO,IAAI,aAAa,EAAW,SAAS,CAAC,GAC7C,IAAM,IAAI,WAAW,EAAW,MAAM,GACtC,IAAO,IAAI,aAAa,EAAW,MAAM;CAU/C,OATA,EAAW,SAAS,GAAO,MAAU;EACjC,IAAM,CAAC,GAAI,GAAI,GAAI,KAAM,EAAM,KAAK;EAMpC,AALA,EAAK,IAAQ,KAAK,GAClB,EAAK,IAAQ,IAAI,KAAK,GACtB,EAAK,IAAQ,IAAI,KAAK,GACtB,EAAK,IAAQ,IAAI,KAAK,GACtB,EAAI,KAAS,EAAM,SACnB,EAAK,KAAS,EAAM;CACxB,CAAC,GACM,IAAI,EAAM,GAAM,GAAK,GAAM,CAAS;AAC/C"}

package/dist/vision/tasks/detector.cjs

@@ -1,2 +1,2 @@
-const e=require("../types.cjs"),t=require("../core/timing.cjs"),n=require("../results.cjs"),r=require("../labels.cjs"),i=require("../core/graph.cjs"),a=require("../core/metadata.cjs"),o=require("../core/session.cjs"),s=require("../io/image.cjs"),c=require("../preprocess/image.cjs"),l=require("../preprocess/pipeline.cjs"),u=require("../postprocess/detection.cjs"),d=require("./base.cjs");var f=class f extends d.VisionTask{_head;_labels;_names;_inputSize;_confThreshold;_iouThreshold;_maxDetections;_raiseOnEmpty;constructor(e,t,n,r,i,a,o,s,c){super(e),this._head=t,this._labels=n,this._names=r,this._inputSize=i,this._confThreshold=a,this._iouThreshold=o,this._maxDetections=s,this._raiseOnEmpty=c}_pipelineCache=null;async warmup(e=1){let[t,n]=this._inputSize;for(let r=0;r<e;r++){let e=c.toFloat32Tensor(l.zeroTensorData(t,n),[1,3,n,t]);await this._session.run({[this._session.inputName]:e})}}get _pipeline(){return this._pipelineCache===null&&(this._pipelineCache=new l.LetterboxPipeline(this._inputSize[0],this._inputSize[1])),this._pipelineCache}static async create(e,t={}){let n=t.head??`yolo`;if(n!==`yolo`)throw Error(`Unsupported detector head '${n}'. Supported: 'yolo'.`);let s=await o.OrtSession.create(e,t),c=t.numClasses??i.detectionNumClasses(s.outputShape)??void 0,l=r.resolveLabels(t.labels??a.modelNames(s.metadata)??r.defaultLabels(c),{numClasses:c}),u={};for(let e=0;e<l.length;e++)u[e]=l[e];return new f(s,n,l,u,i.resolveInputSize({graphShape:s.inputShape,requested:t.inputSize,fallback:[640,640]}),t.confThreshold??.25,t.iouThreshold??.45,t.maxDetections??300,t.raiseOnEmpty??!1)}get head(){return this._head}get labels(){return this._labels}get names(){return this._names}get inputSize(){return this._inputSize}get numClasses(){return this._labels.length}async call(e,t={}){return this.predict(e,t)}async predict(e,r={}){let i=new t.SpeedTimer,a=typeof e==`string`?e:null,o=await s.loadImage(e);i.stage(`load`);let{tensor:c,scale:l,padLeft:f,padTop:p}=this._preprocess(o);i.stage(`preprocess`);let m=await this._session.run({[this._session.inputName]:c});this._pipeline.release(),i.stage(`inference`);let h=this._session.outputNames[0];if(h===void 0)throw Error(`Detector model has no outputs.`);let g=m[h];if(g===void 0)throw Error(`Detector model output ${h} missing from run() result.`);let _=r.confThreshold??this._confThreshold,v=u.decodeYolo(g.data,g.dims,{originalWidth:o.width,originalHeight:o.height,padLeft:f,padTop:p,scale:l,confThreshold:_,iouThreshold:r.iouThreshold??this._iouThreshold,maxDetections:this._maxDetections}),y=r.classes===void 0?v:(()=>{let e=new Set(r.classes);return v.filter(t=>e.has(t.classId))})();d.requireDetections(y.length,{raiseOnEmpty:r.raiseOnEmpty??this._raiseOnEmpty,confThreshold:_,classes:r.classes,path:a});let b=y.map(e=>this._buildResult(o,e.bbox,e.classId,e.confidence)),x=[o.height,o.width],S=this._buildBoxes(b,x);return i.stage(`postprocess`),[new n.DetectionResults(S,b,this._names,o,x,a,i.speed())]}_preprocess(e){let[t,n]=this._inputSize,r=this._pipeline.run(e);return{tensor:c.toFloat32Tensor(r.data,[1,3,n,t]),scale:r.scale,padLeft:r.padLeft,padTop:r.padTop}}_buildResult(t,n,r,i){let[a,o,s,c]=n.asIntXyxy(),l=Math.max(0,a),u=Math.max(0,o),d=Math.min(t.width,s),f=Math.min(t.height,c),p;if(d>l&&f>u){let n=d-l,r=f-u,i=new Uint8Array(n*r*3);for(let e=0;e<r;e++){let r=((u+e)*t.width+l)*3;i.set(t.data.subarray(r,r+n*3),e*n*3)}p=new e.RGBImage(i,n,r)}else p=new e.RGBImage(new Uint8Array,0,0);let m=this._names[r]??`class_${r}`;return{classId:r,className:m,confidence:i,bbox:n,cls:r,name:m,conf:i,box:n,croppedImage:p}}_buildBoxes(e,t){let r=e.length,i=new Float32Array(r*4),a=new Int32Array(r),o=new Float32Array(r);for(let t=0;t<r;t++){let n=e[t];i[t*4]=n.bbox.x1,i[t*4+1]=n.bbox.y1,i[t*4+2]=n.bbox.x2,i[t*4+3]=n.bbox.y2,a[t]=n.classId,o[t]=n.confidence}return new n.Boxes(i,a,o,t)}};exports.Detector=f;
+const e=require("../core/timing.cjs"),t=require("../results.cjs"),n=require("../labels.cjs"),r=require("../core/dtypes.cjs"),i=require("../core/graph.cjs"),a=require("../core/metadata.cjs"),o=require("../core/session.cjs"),s=require("../io/image.cjs"),c=require("../preprocess/image.cjs"),l=require("../preprocess/pipeline.cjs"),u=require("../postprocess/detection.cjs"),d=require("./base.cjs");var f=class f extends d.VisionTask{_head;_labels;_names;_inputSize;_confThreshold;_iouThreshold;_maxDetections;_raiseOnEmpty;constructor(e,t,n,r,i,a,o,s,c){super(e),this._head=t,this._labels=n,this._names=r,this._inputSize=i,this._confThreshold=a,this._iouThreshold=o,this._maxDetections=s,this._raiseOnEmpty=c}_pipelineCache=null;async warmup(e=1){let[t,n]=this._inputSize;for(let r=0;r<e;r++){let e=c.toFloat32Tensor(l.zeroTensorData(t,n),[1,3,n,t]);await this._session.run({[this._session.inputName]:e})}}get _pipeline(){return this._pipelineCache===null&&(this._pipelineCache=new l.LetterboxPipeline(this._inputSize[0],this._inputSize[1])),this._pipelineCache}static async create(e,t={}){let r=t.head??`yolo`;if(r!==`yolo`)throw Error(`Unsupported detector head '${r}'. Supported: 'yolo'.`);let s=await o.OrtSession.create(e,t),c=t.numClasses??i.detectionNumClasses(s.outputShape)??void 0,l=n.resolveLabels(t.labels??a.modelNames(s.metadata)??n.defaultLabels(c),{numClasses:c}),u={};for(let e=0;e<l.length;e++)u[e]=l[e];return new f(s,r,l,u,i.resolveInputSize({graphShape:s.inputShape,requested:t.inputSize,fallback:[640,640]}),t.confThreshold??.25,t.iouThreshold??.45,t.maxDetections??300,t.raiseOnEmpty??!1)}get head(){return this._head}get labels(){return this._labels}get names(){return this._names}get inputSize(){return this._inputSize}get numClasses(){return this._labels.length}async call(e,t={}){return this.predict(e,t)}async predict(n,i={}){let a=new e.SpeedTimer,o=typeof n==`string`?n:null,{image:c,canvas:l,owned:f}=await s.loadImageSource(n);a.stage(`load`);let{tensor:p,scale:m,padLeft:h,padTop:g,reused:_}=this._preprocess(c,l);a.stage(`preprocess`);let v;try{v=await this._session.run({[this._session.inputName]:p})}finally{this._pipeline.release({reused:_})}a.stage(`inference`);let y=this._session.outputNames[0];if(y===void 0)throw Error(`Detector model has no outputs.`);let b=v[y];if(b===void 0)throw Error(`Detector model output ${y} missing from run() result.`);let x=i.confThreshold??this._confThreshold,S=u.decodeYolo(r.asFloat32Array(b.data),b.dims,{originalWidth:c.width,originalHeight:c.height,padLeft:h,padTop:g,scale:m,confThreshold:x,iouThreshold:i.iouThreshold??this._iouThreshold,maxDetections:this._maxDetections}),C=i.classes===void 0?S:(()=>{let e=new Set(i.classes);return S.filter(t=>e.has(t.classId))})();d.requireDetections(C.length,{raiseOnEmpty:i.raiseOnEmpty??this._raiseOnEmpty,confThreshold:x,classes:i.classes,path:o});let w=C.map(e=>this._buildResult(c,e.bbox,e.classId,e.confidence,f)),T=[c.height,c.width],E=this._buildBoxes(w,T);return a.stage(`postprocess`),[new t.DetectionResults(E,w,this._names,c,T,o,a.speed())]}_preprocess(e,t){let[n,r]=this._inputSize,i=this._pipeline.run(e,t);return{tensor:c.toFloat32Tensor(i.data,[1,3,r,n]),scale:i.scale,padLeft:i.padLeft,padTop:i.padTop,reused:i.reused}}_buildResult(e,t,n,r,i){let a=this._names[n]??`class_${n}`,o={classId:n,className:a,confidence:r,bbox:t,cls:n,name:a,conf:r,box:t};if(!i)return{...o,croppedImage:d.cropToBox(e,t)};let s=d.memoize(()=>d.cropToBox(e,t));return{...o,get croppedImage(){return s()}}}_buildBoxes(e,n){let r=e.length,i=new Float32Array(r*4),a=new Int32Array(r),o=new Float32Array(r);for(let t=0;t<r;t++){let n=e[t];i[t*4]=n.bbox.x1,i[t*4+1]=n.bbox.y1,i[t*4+2]=n.bbox.x2,i[t*4+3]=n.bbox.y2,a[t]=n.classId,o[t]=n.confidence}return new t.Boxes(i,a,o,n)}};exports.Detector=f;
 //# sourceMappingURL=detector.cjs.map