tempest-react-sdk 0.57.0 → 0.59.0

package/dist/vision/core/session.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"session.cjs","names":[],"sources":["../../../src/vision/core/session.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Thin wrapper around `onnxruntime-web` `InferenceSession` with typed metadata.\n */\n\nimport type * as ort from \"onnxruntime-web\";\nimport * as ortRuntime from \"onnxruntime-web\";\n\nimport { InferenceError, ModelLoadError } from \"./exceptions\";\nimport { type DeclaredShape, declaredShapesFrom } from \"./graph\";\nimport { readModelMetadata } from \"./metadata\";\nimport { resolveProviders } from \"./providers\";\n\n/** Anything `InferenceSession.create` accepts. */\nexport type ModelSource = string | ArrayBufferLike | Uint8Array;\n\n/**\n * Fetch a model URL as bytes so its metadata can be read.\n *\n * Falls back to the URL itself when the fetch fails, letting ORT try its own\n * load path: losing the metadata map is a downgrade, but failing to load a model\n * that ORT could have fetched would be a regression.\n *\n * @param url Where the `.onnx` lives.\n * @returns The model bytes, or the original URL when they could not be fetched.\n */\nasync function fetchModel(url: string): Promise<Uint8Array | string> {\n    try {\n        const response = await fetch(url);\n        if (!response.ok) {\n            warnMetadataUnavailable(url, `HTTP ${response.status} ${response.statusText}`);\n            return url;\n        }\n        return new Uint8Array(await response.arrayBuffer());\n    } catch (err) {\n        warnMetadataUnavailable(url, (err as Error).message);\n        return url;\n    }\n}\n\n/**\n * Warn that a model's metadata could not be read, and say what that costs.\n *\n * The fallback itself is right — losing the metadata beats failing a load that\n * ORT could have completed on its own — but it used to be silent, and the\n * symptom it produces is remote from the cause: class names come back as\n * `class_0`, `class_1`, ... with nothing anywhere explaining why. Whoever hits\n * this needs to be told that passing `labels` is the way out.\n *\n * @param url The model URL that could not be fetched here.\n * @param reason What went wrong, as reported by `fetch`.\n */\nfunction warnMetadataUnavailable(url: string, reason: string): void {\n    console.warn(\n        `[@ort-vision-sdk/web] Could not fetch ${url} to read its metadata (${reason}). ` +\n            \"Letting ONNX Runtime load it instead: the model will work, but its baked-in \" +\n            \"class names are unavailable, so labels fall back to class_0, class_1, ... \" +\n            \"Pass `labels` explicitly to name them.\",\n    );\n}\n\nexport interface OrtSessionOptions {\n    /** Execution providers in preference order. `undefined` uses {@link DEFAULT_PROVIDERS}. */\n    readonly providers?: readonly string[];\n    /** Optional ORT session options forwarded to `InferenceSession.create`. */\n    readonly sessionOptions?: ort.InferenceSession.SessionOptions;\n    /**\n     * Whether to read the model's custom metadata map (`names`, `task`, `imgsz`).\n     * Defaults to `true`.\n     *\n     * The runtime does not expose that map, so it is read from the file itself —\n     * which means a URL model is fetched here and handed to ORT as bytes instead\n     * of letting ORT fetch it. That is the same single download either way, and\n     * it is what lets a task resolve its labels off the model. Set to `false` to\n     * keep the URL path untouched and leave {@link OrtSession.metadata} empty.\n     *\n     * `false` is also the escape hatch when a device cannot afford the bytes: the\n     * fetched buffer is dropped before ORT builds the graph (see\n     * {@link OrtSession.create}), but ORT's own load path still keeps the model out\n     * of reach of anything the SDK holds. A session built this way resolves its\n     * input size from the graph as usual — only the class names are lost, so a\n     * caller taking this route has to pass `labels` itself.\n     */\n    readonly readMetadata?: boolean;\n}\n\n/**\n * Wrap an ONNX Runtime Web `InferenceSession` with convenient metadata access.\n *\n * The wrapper exposes input/output names and the shapes the graph declares,\n * manages execution-provider selection, provides a typed {@link OrtSession.run}\n * method, and releases the native session through {@link OrtSession.release}.\n */\nexport class OrtSession {\n    private constructor(\n        private readonly _session: ort.InferenceSession,\n        public readonly providers: readonly string[],\n        private readonly _metadata: Readonly<Record<string, string>>,\n    ) {}\n\n    /**\n     * Load an ONNX model into an ORT inference session.\n     *\n     * The metadata map is read **before** the session is built, and that order is\n     * load-bearing on memory-constrained devices. ORT copies the model into its\n     * WASM heap and then allocates the graph and the weights on top of that copy;\n     * a `readModelMetadata` call placed after `InferenceSession.create` keeps the\n     * JavaScript-side buffer reachable across the whole build, so a 5 MB model\n     * costs 5 MB of JS heap plus 5 MB of WASM heap plus the weights at the same\n     * instant. Reading first makes the buffer collectable as soon as ORT has copied\n     * it — on a phone that was the difference between a session and\n     * `Can't create a session. failed to allocate a buffer of size N`.\n     *\n     * @param model Either a URL string, or a `Uint8Array`/`ArrayBuffer` containing the model bytes.\n     * @param options Provider list, pass-through `SessionOptions`, and whether to\n     *   read the model's metadata map (see {@link OrtSessionOptions.readMetadata}).\n     * @throws {@link ModelLoadError} if the model cannot be loaded.\n     */\n    static async create(model: ModelSource, options: OrtSessionOptions = {}): Promise<OrtSession> {\n        const providers = resolveProviders(options.providers);\n        const sessionOptions: ort.InferenceSession.SessionOptions = {\n            ...(options.sessionOptions ?? {}),\n            executionProviders:\n                providers as ort.InferenceSession.SessionOptions[\"executionProviders\"],\n        };\n        const wantsMetadata = options.readMetadata !== false;\n        const source = typeof model === \"string\" && wantsMetadata ? await fetchModel(model) : model;\n        const metadata =\n            wantsMetadata && typeof source !== \"string\" ? readModelMetadata(source) : {};\n\n        let session: ort.InferenceSession;\n        try {\n            if (typeof source === \"string\") {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else if (source instanceof Uint8Array) {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else {\n                session = await ortRuntime.InferenceSession.create(\n                    source as ArrayBuffer,\n                    sessionOptions,\n                );\n            }\n        } catch (err) {\n            throw new ModelLoadError(`Failed to load ONNX model: ${(err as Error).message}`, {\n                cause: err,\n            });\n        }\n\n        return new OrtSession(session, providers, metadata);\n    }\n\n    /** Names of the model's inputs, in declaration order. */\n    get inputNames(): readonly string[] {\n        return this._session.inputNames;\n    }\n\n    /** Name of the first (and usually only) input. */\n    get inputName(): string {\n        const name = this._session.inputNames[0];\n        if (name === undefined) {\n            throw new InferenceError(\"Model has no inputs.\");\n        }\n        return name;\n    }\n\n    /** Names of the model's outputs, in declaration order. */\n    get outputNames(): readonly string[] {\n        return this._session.outputNames;\n    }\n\n    /**\n     * Shapes the graph declares for its inputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Empty shapes mean the runtime\n     * reported no metadata — either a non-tensor input, or an `onnxruntime-web`\n     * older than 1.21, which predates input metadata.\n     */\n    get inputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.inputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first input, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get inputShape(): DeclaredShape {\n        return this.inputShapes[0] ?? [];\n    }\n\n    /**\n     * Shapes the graph declares for its outputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Reading them is how a task can\n     * tell how many classes a head emits without being told.\n     */\n    get outputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.outputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first output, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get outputShape(): DeclaredShape {\n        return this.outputShapes[0] ?? [];\n    }\n\n    /**\n     * The model's custom metadata map — `names`, `task`, `imgsz`, ... for an\n     * Ultralytics export.\n     *\n     * Read from the model's bytes at load time, since the runtime does not expose\n     * it. Empty when the session was created with `readMetadata: false`, from a\n     * URL that could not be fetched here, or from a model carrying no metadata.\n     */\n    get metadata(): Readonly<Record<string, string>> {\n        return this._metadata;\n    }\n\n    /**\n     * Release the native session and free its memory.\n     *\n     * Call it when a session is discarded while the page lives on — rebuilding a\n     * task at a different input size, swapping in a newer model. A failure from\n     * the runtime is ignored: a session being torn down has nothing left to fail\n     * at, and the caller is already moving on.\n     */\n    async release(): Promise<void> {\n        await this._session.release().catch(() => undefined);\n    }\n\n    /** The underlying `onnxruntime-web` session, for advanced use cases. */\n    get raw(): ort.InferenceSession {\n        return this._session;\n    }\n\n    /**\n     * Run inference and return all outputs.\n     *\n     * @param feeds Map of input name to `ort.Tensor`. Keys must match {@link inputNames}.\n     * @throws {@link InferenceError} if ORT raises any error during execution.\n     */\n    async run(feeds: Record<string, ort.Tensor>): Promise<Record<string, ort.Tensor>> {\n        try {\n            const result = await this._session.run(feeds);\n            return result as Record<string, ort.Tensor>;\n        } catch (err) {\n            throw new InferenceError(`Inference failed: ${(err as Error).message}`, { cause: err });\n        }\n    }\n}\n"],"mappings":"0NA0BA,eAAe,EAAW,EAA2C,CACjE,GAAI,CACA,IAAM,EAAW,MAAM,MAAM,CAAG,EAKhC,OAJK,EAAS,GAIP,IAAI,WAAW,MAAM,EAAS,YAAY,CAAC,GAH9C,EAAwB,EAAK,QAAQ,EAAS,OAAO,GAAG,EAAS,YAAY,EACtE,EAGf,OAAS,EAAK,CAEV,OADA,EAAwB,EAAM,EAAc,OAAO,EAC5C,CACX,CACJ,CAcA,SAAS,EAAwB,EAAa,EAAsB,CAChE,QAAQ,KACJ,yCAAyC,EAAI,yBAAyB,EAAO,kMAIjF,CACJ,CAkCA,IAAa,EAAb,MAAa,CAAW,CAEC,SACD,UACC,UAHrB,YACI,EACA,EACA,EACF,CAHmB,KAAA,SAAA,EACD,KAAA,UAAA,EACC,KAAA,UAAA,CAClB,CAoBH,aAAa,OAAO,EAAoB,EAA6B,CAAC,EAAwB,CAC1F,IAAM,EAAY,EAAA,iBAAiB,EAAQ,SAAS,EAC9C,EAAsD,CACxD,GAAI,EAAQ,gBAAkB,CAAC,EAC/B,mBACI,CACR,EACM,EAAgB,EAAQ,eAAiB,GACzC,EAAS,OAAO,GAAU,UAAY,EAAgB,MAAM,EAAW,CAAK,EAAI,EAChF,EACF,GAAiB,OAAO,GAAW,SAAW,EAAA,kBAAkB,CAAM,EAAI,CAAC,EAE3E,EACJ,GAAI,CACA,AAKI,GALA,OAAO,GAAW,UAEX,aAAkB,WADf,MAAM,EAAW,iBAAiB,OAAO,EAAQ,CAAc,EASjF,OAAS,EAAK,CACV,MAAM,IAAI,EAAA,eAAe,8BAA+B,EAAc,UAAW,CAC7E,MAAO,CACX,CAAC,CACL,CAEA,OAAO,IAAI,EAAW,EAAS,EAAW,CAAQ,CACtD,CAGA,IAAI,YAAgC,CAChC,OAAO,KAAK,SAAS,UACzB,CAGA,IAAI,WAAoB,CACpB,IAAM,EAAO,KAAK,SAAS,WAAW,GACtC,GAAI,IAAS,IAAA,GACT,MAAM,IAAI,EAAA,eAAe,sBAAsB,EAEnD,OAAO,CACX,CAGA,IAAI,aAAiC,CACjC,OAAO,KAAK,SAAS,WACzB,CASA,IAAI,aAAwC,CACxC,OAAO,EAAA,mBACH,KAAK,SAAS,aAElB,CACJ,CAOA,IAAI,YAA4B,CAC5B,OAAO,KAAK,YAAY,IAAM,CAAC,CACnC,CAQA,IAAI,cAAyC,CACzC,OAAO,EAAA,mBACH,KAAK,SAAS,cAElB,CACJ,CAOA,IAAI,aAA6B,CAC7B,OAAO,KAAK,aAAa,IAAM,CAAC,CACpC,CAUA,IAAI,UAA6C,CAC7C,OAAO,KAAK,SAChB,CAUA,MAAM,SAAyB,CAC3B,MAAM,KAAK,SAAS,QAAQ,CAAC,CAAC,UAAY,IAAA,EAAS,CACvD,CAGA,IAAI,KAA4B,CAC5B,OAAO,KAAK,QAChB,CAQA,MAAM,IAAI,EAAwE,CAC9E,GAAI,CAEA,OAAO,MADc,KAAK,SAAS,IAAI,CAAK,CAEhD,OAAS,EAAK,CACV,MAAM,IAAI,EAAA,eAAe,qBAAsB,EAAc,UAAW,CAAE,MAAO,CAAI,CAAC,CAC1F,CACJ,CACJ"}
+{"version":3,"file":"session.cjs","names":[],"sources":["../../../src/vision/core/session.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Thin wrapper around `onnxruntime-web` `InferenceSession` with typed metadata.\n */\n\nimport type * as ort from \"onnxruntime-web\";\nimport * as ortRuntime from \"onnxruntime-web\";\n\nimport { InferenceError, ModelLoadError } from \"./exceptions\";\nimport { type DeclaredShape, declaredShapesFrom } from \"./graph\";\nimport { readModelMetadata } from \"./metadata\";\nimport { FALLBACK_PROVIDER, detectProviders, resolveProviders } from \"./providers\";\n\n/** Anything `InferenceSession.create` accepts. */\nexport type ModelSource = string | ArrayBufferLike | Uint8Array;\n\n/**\n * Fetch a model URL as bytes so its metadata can be read.\n *\n * Falls back to the URL itself when the fetch fails, letting ORT try its own\n * load path: losing the metadata map is a downgrade, but failing to load a model\n * that ORT could have fetched would be a regression.\n *\n * @param url Where the `.onnx` lives.\n * @returns The model bytes, or the original URL when they could not be fetched.\n */\nasync function fetchModel(url: string): Promise<Uint8Array | string> {\n    try {\n        const response = await fetch(url);\n        if (!response.ok) {\n            warnMetadataUnavailable(url, `HTTP ${response.status} ${response.statusText}`);\n            return url;\n        }\n        return new Uint8Array(await response.arrayBuffer());\n    } catch (err) {\n        warnMetadataUnavailable(url, (err as Error).message);\n        return url;\n    }\n}\n\n/**\n * Warn that a model's metadata could not be read, and say what that costs.\n *\n * The fallback itself is right — losing the metadata beats failing a load that\n * ORT could have completed on its own — but it used to be silent, and the\n * symptom it produces is remote from the cause: class names come back as\n * `class_0`, `class_1`, ... with nothing anywhere explaining why. Whoever hits\n * this needs to be told that passing `labels` is the way out.\n *\n * @param url The model URL that could not be fetched here.\n * @param reason What went wrong, as reported by `fetch`.\n */\nfunction warnMetadataUnavailable(url: string, reason: string): void {\n    console.warn(\n        `[@ort-vision-sdk/web] Could not fetch ${url} to read its metadata (${reason}). ` +\n            \"Letting ONNX Runtime load it instead: the model will work, but its baked-in \" +\n            \"class names are unavailable, so labels fall back to class_0, class_1, ... \" +\n            \"Pass `labels` explicitly to name them.\",\n    );\n}\n\nexport interface OrtSessionOptions {\n    /**\n     * Execution providers in preference order. `undefined` uses {@link DEFAULT_PROVIDERS}.\n     *\n     * Naming one explicitly also opts into a `console.warn` when this browser\n     * cannot offer it, instead of falling back in silence.\n     */\n    readonly providers?: readonly string[];\n    /** Optional ORT session options forwarded to `InferenceSession.create`. */\n    readonly sessionOptions?: ort.InferenceSession.SessionOptions;\n    /**\n     * Whether to read the model's custom metadata map (`names`, `task`, `imgsz`).\n     * Defaults to `true`.\n     *\n     * The runtime does not expose that map, so it is read from the file itself —\n     * which means a URL model is fetched here and handed to ORT as bytes instead\n     * of letting ORT fetch it. That is the same single download either way, and\n     * it is what lets a task resolve its labels off the model. Set to `false` to\n     * keep the URL path untouched and leave {@link OrtSession.metadata} empty.\n     *\n     * `false` is also the escape hatch when a device cannot afford the bytes: the\n     * fetched buffer is dropped before ORT builds the graph (see\n     * {@link OrtSession.create}), but ORT's own load path still keeps the model out\n     * of reach of anything the SDK holds. A session built this way resolves its\n     * input size from the graph as usual — only the class names are lost, so a\n     * caller taking this route has to pass `labels` itself.\n     */\n    readonly readMetadata?: boolean;\n}\n\n/**\n * Warn when a provider named explicitly by the caller is not going to run.\n *\n * Only explicit requests are worth a warning. The default list exists precisely\n * so that falling from `webgpu` to `wasm` is the expected outcome — but a caller\n * who wrote `providers: [\"webgpu\"]` and lands on WASM has a page several times\n * slower than intended and nothing in the console to explain it.\n *\n * @param requested Providers that were asked for.\n * @param effective Providers that survived capability detection.\n */\nfunction warnOnDroppedProviders(requested: readonly string[], effective: readonly string[]): void {\n    const dropped = requested.filter((provider) => !effective.includes(provider));\n    if (dropped.length === 0) {\n        return;\n    }\n    console.warn(\n        `This browser cannot offer the requested execution provider(s) ${JSON.stringify(dropped)}; ` +\n            `the session will run on ${JSON.stringify(effective)}. Inference still produces correct ` +\n            \"results, on the fallback provider.\",\n    );\n}\n\n/**\n * Wrap an ONNX Runtime Web `InferenceSession` with convenient metadata access.\n *\n * The wrapper exposes input/output names and the shapes the graph declares,\n * manages execution-provider selection, provides a typed {@link OrtSession.run}\n * method, and releases the native session through {@link OrtSession.release}.\n */\nexport class OrtSession {\n    private constructor(\n        private readonly _session: ort.InferenceSession,\n        /**\n         * Execution providers this session is expected to run on.\n         *\n         * The requested list narrowed to what this browser can actually offer — a\n         * `webgpu` entry survives only where an adapter exists. Best-effort: ORT-Web\n         * exposes no way to ask which provider a session ended up on, so an entry\n         * here means \"not ruled out\", not \"confirmed\". See\n         * {@link requestedProviders} for what was asked for.\n         *\n         * When nothing survives — a caller asking for `webgpu` alone on a device\n         * without it — this falls back to {@link FALLBACK_PROVIDER}, which ORT-Web\n         * can always run. Handing ORT the unsatisfiable list instead makes\n         * `InferenceSession.create` reject with \"no available backend found\", so the\n         * page gets no inference at all rather than the slow-but-working fallback\n         * the `console.warn` describes. Measured in a real Chromium, where\n         * `navigator.gpu` exists but yields no adapter.\n         */\n        public readonly providers: readonly string[],\n        private readonly _metadata: Readonly<Record<string, string>>,\n        /**\n         * Execution providers that were asked for, after defaults were applied.\n         *\n         * Kept separate because ORT-Web falls back silently: a page that asks for\n         * `webgpu` on a device without it runs on WASM and is told nothing.\n         */\n        public readonly requestedProviders: readonly string[],\n    ) {}\n\n    /**\n     * Load an ONNX model into an ORT inference session.\n     *\n     * The metadata map is read **before** the session is built, and that order is\n     * load-bearing on memory-constrained devices. ORT copies the model into its\n     * WASM heap and then allocates the graph and the weights on top of that copy;\n     * a `readModelMetadata` call placed after `InferenceSession.create` keeps the\n     * JavaScript-side buffer reachable across the whole build, so a 5 MB model\n     * costs 5 MB of JS heap plus 5 MB of WASM heap plus the weights at the same\n     * instant. Reading first makes the buffer collectable as soon as ORT has copied\n     * it — on a phone that was the difference between a session and\n     * `Can't create a session. failed to allocate a buffer of size N`.\n     *\n     * @param model Either a URL string, or a `Uint8Array`/`ArrayBuffer` containing the model bytes.\n     * @param options Provider list, pass-through `SessionOptions`, and whether to\n     *   read the model's metadata map (see {@link OrtSessionOptions.readMetadata}).\n     * @throws {@link ModelLoadError} if the model cannot be loaded.\n     */\n    static async create(model: ModelSource, options: OrtSessionOptions = {}): Promise<OrtSession> {\n        const requested = resolveProviders(options.providers);\n        const detected = await detectProviders(requested);\n        const providers = detected.length > 0 ? detected : [FALLBACK_PROVIDER];\n        if (options.providers !== undefined && options.providers.length > 0) {\n            warnOnDroppedProviders(requested, providers);\n        }\n        const sessionOptions: ort.InferenceSession.SessionOptions = {\n            ...(options.sessionOptions ?? {}),\n            executionProviders:\n                providers as ort.InferenceSession.SessionOptions[\"executionProviders\"],\n        };\n        const wantsMetadata = options.readMetadata !== false;\n        const source = typeof model === \"string\" && wantsMetadata ? await fetchModel(model) : model;\n        const metadata =\n            wantsMetadata && typeof source !== \"string\" ? readModelMetadata(source) : {};\n\n        let session: ort.InferenceSession;\n        try {\n            if (typeof source === \"string\") {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else if (source instanceof Uint8Array) {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else {\n                session = await ortRuntime.InferenceSession.create(\n                    source as ArrayBuffer,\n                    sessionOptions,\n                );\n            }\n        } catch (err) {\n            throw new ModelLoadError(`Failed to load ONNX model: ${(err as Error).message}`, {\n                cause: err,\n            });\n        }\n\n        return new OrtSession(session, providers, metadata, requested);\n    }\n\n    /** Names of the model's inputs, in declaration order. */\n    get inputNames(): readonly string[] {\n        return this._session.inputNames;\n    }\n\n    /** Name of the first (and usually only) input. */\n    get inputName(): string {\n        const name = this._session.inputNames[0];\n        if (name === undefined) {\n            throw new InferenceError(\"Model has no inputs.\");\n        }\n        return name;\n    }\n\n    /** Names of the model's outputs, in declaration order. */\n    get outputNames(): readonly string[] {\n        return this._session.outputNames;\n    }\n\n    /**\n     * Shapes the graph declares for its inputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Empty shapes mean the runtime\n     * reported no metadata — either a non-tensor input, or an `onnxruntime-web`\n     * older than 1.21, which predates input metadata.\n     */\n    get inputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.inputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first input, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get inputShape(): DeclaredShape {\n        return this.inputShapes[0] ?? [];\n    }\n\n    /**\n     * Shapes the graph declares for its outputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Reading them is how a task can\n     * tell how many classes a head emits without being told.\n     */\n    get outputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.outputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first output, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get outputShape(): DeclaredShape {\n        return this.outputShapes[0] ?? [];\n    }\n\n    /**\n     * The model's custom metadata map — `names`, `task`, `imgsz`, ... for an\n     * Ultralytics export.\n     *\n     * Read from the model's bytes at load time, since the runtime does not expose\n     * it. Empty when the session was created with `readMetadata: false`, from a\n     * URL that could not be fetched here, or from a model carrying no metadata.\n     */\n    get metadata(): Readonly<Record<string, string>> {\n        return this._metadata;\n    }\n\n    /**\n     * Release the native session and free its memory.\n     *\n     * Call it when a session is discarded while the page lives on — rebuilding a\n     * task at a different input size, swapping in a newer model. A failure from\n     * the runtime is ignored: a session being torn down has nothing left to fail\n     * at, and the caller is already moving on.\n     */\n    async release(): Promise<void> {\n        await this._session.release().catch(() => undefined);\n    }\n\n    /** The underlying `onnxruntime-web` session, for advanced use cases. */\n    get raw(): ort.InferenceSession {\n        return this._session;\n    }\n\n    /**\n     * Run inference and return all outputs.\n     *\n     * @param feeds Map of input name to `ort.Tensor`. Keys must match {@link inputNames}.\n     * @throws {@link InferenceError} if ORT raises any error during execution.\n     */\n    async run(feeds: Record<string, ort.Tensor>): Promise<Record<string, ort.Tensor>> {\n        try {\n            const result = await this._session.run(feeds);\n            return result as Record<string, ort.Tensor>;\n        } catch (err) {\n            throw new InferenceError(`Inference failed: ${(err as Error).message}`, { cause: err });\n        }\n    }\n}\n"],"mappings":"0NA0BA,eAAe,EAAW,EAA2C,CACjE,GAAI,CACA,IAAM,EAAW,MAAM,MAAM,CAAG,EAKhC,OAJK,EAAS,GAIP,IAAI,WAAW,MAAM,EAAS,YAAY,CAAC,GAH9C,EAAwB,EAAK,QAAQ,EAAS,OAAO,GAAG,EAAS,YAAY,EACtE,EAGf,OAAS,EAAK,CAEV,OADA,EAAwB,EAAM,EAAc,OAAO,EAC5C,CACX,CACJ,CAcA,SAAS,EAAwB,EAAa,EAAsB,CAChE,QAAQ,KACJ,yCAAyC,EAAI,yBAAyB,EAAO,kMAIjF,CACJ,CA2CA,SAAS,EAAuB,EAA8B,EAAoC,CAC9F,IAAM,EAAU,EAAU,OAAQ,GAAa,CAAC,EAAU,SAAS,CAAQ,CAAC,EACxE,EAAQ,SAAW,GAGvB,QAAQ,KACJ,iEAAiE,KAAK,UAAU,CAAO,EAAE,4BAC1D,KAAK,UAAU,CAAS,EAAE,sEAE7D,CACJ,CASA,IAAa,EAAb,MAAa,CAAW,CAEC,SAkBD,UACC,UAOD,mBA3BpB,YACI,EAkBA,EACA,EAOA,EACF,CA3BmB,KAAA,SAAA,EAkBD,KAAA,UAAA,EACC,KAAA,UAAA,EAOD,KAAA,mBAAA,CACjB,CAoBH,aAAa,OAAO,EAAoB,EAA6B,CAAC,EAAwB,CAC1F,IAAM,EAAY,EAAA,iBAAiB,EAAQ,SAAS,EAC9C,EAAW,MAAM,EAAA,gBAAgB,CAAS,EAC1C,EAAY,EAAS,OAAS,EAAI,EAAW,CAAC,EAAA,iBAAiB,EACjE,EAAQ,YAAc,IAAA,IAAa,EAAQ,UAAU,OAAS,GAC9D,EAAuB,EAAW,CAAS,EAE/C,IAAM,EAAsD,CACxD,GAAI,EAAQ,gBAAkB,CAAC,EAC/B,mBACI,CACR,EACM,EAAgB,EAAQ,eAAiB,GACzC,EAAS,OAAO,GAAU,UAAY,EAAgB,MAAM,EAAW,CAAK,EAAI,EAChF,EACF,GAAiB,OAAO,GAAW,SAAW,EAAA,kBAAkB,CAAM,EAAI,CAAC,EAE3E,EACJ,GAAI,CACA,AAKI,GALA,OAAO,GAAW,UAEX,aAAkB,WADf,MAAM,EAAW,iBAAiB,OAAO,EAAQ,CAAc,EASjF,OAAS,EAAK,CACV,MAAM,IAAI,EAAA,eAAe,8BAA+B,EAAc,UAAW,CAC7E,MAAO,CACX,CAAC,CACL,CAEA,OAAO,IAAI,EAAW,EAAS,EAAW,EAAU,CAAS,CACjE,CAGA,IAAI,YAAgC,CAChC,OAAO,KAAK,SAAS,UACzB,CAGA,IAAI,WAAoB,CACpB,IAAM,EAAO,KAAK,SAAS,WAAW,GACtC,GAAI,IAAS,IAAA,GACT,MAAM,IAAI,EAAA,eAAe,sBAAsB,EAEnD,OAAO,CACX,CAGA,IAAI,aAAiC,CACjC,OAAO,KAAK,SAAS,WACzB,CASA,IAAI,aAAwC,CACxC,OAAO,EAAA,mBACH,KAAK,SAAS,aAElB,CACJ,CAOA,IAAI,YAA4B,CAC5B,OAAO,KAAK,YAAY,IAAM,CAAC,CACnC,CAQA,IAAI,cAAyC,CACzC,OAAO,EAAA,mBACH,KAAK,SAAS,cAElB,CACJ,CAOA,IAAI,aAA6B,CAC7B,OAAO,KAAK,aAAa,IAAM,CAAC,CACpC,CAUA,IAAI,UAA6C,CAC7C,OAAO,KAAK,SAChB,CAUA,MAAM,SAAyB,CAC3B,MAAM,KAAK,SAAS,QAAQ,CAAC,CAAC,UAAY,IAAA,EAAS,CACvD,CAGA,IAAI,KAA4B,CAC5B,OAAO,KAAK,QAChB,CAQA,MAAM,IAAI,EAAwE,CAC9E,GAAI,CAEA,OAAO,MADc,KAAK,SAAS,IAAI,CAAK,CAEhD,OAAS,EAAK,CACV,MAAM,IAAI,EAAA,eAAe,qBAAsB,EAAc,UAAW,CAAE,MAAO,CAAI,CAAC,CAC1F,CACJ,CACJ"}

package/dist/vision/core/session.js

@@ -1,38 +1,45 @@
 import { InferenceError as e, ModelLoadError as t } from "./exceptions.js";
 import { declaredShapesFrom as n } from "./graph.js";
 import { readModelMetadata as r } from "./metadata.js";
-import { resolveProviders as i } from "./providers.js";
-import * as a from "onnxruntime-web";
+import { FALLBACK_PROVIDER as i, detectProviders as a, resolveProviders as o } from "./providers.js";
+import * as s from "onnxruntime-web";
 //#region src/vision/core/session.ts
-async function o(e) {
+async function c(e) {
 	try {
 		let t = await fetch(e);
-		return t.ok ? new Uint8Array(await t.arrayBuffer()) : (s(e, `HTTP ${t.status} ${t.statusText}`), e);
+		return t.ok ? new Uint8Array(await t.arrayBuffer()) : (l(e, `HTTP ${t.status} ${t.statusText}`), e);
 	} catch (t) {
-		return s(e, t.message), e;
+		return l(e, t.message), e;
 	}
 }
-function s(e, t) {
+function l(e, t) {
 	console.warn(`[@ort-vision-sdk/web] Could not fetch ${e} to read its metadata (${t}). Letting ONNX Runtime load it instead: the model will work, but its baked-in class names are unavailable, so labels fall back to class_0, class_1, ... Pass \`labels\` explicitly to name them.`);
 }
-var c = class s {
+function u(e, t) {
+	let n = e.filter((e) => !t.includes(e));
+	n.length !== 0 && console.warn(`This browser cannot offer the requested execution provider(s) ${JSON.stringify(n)}; the session will run on ${JSON.stringify(t)}. Inference still produces correct results, on the fallback provider.`);
+}
+var d = class l {
 	_session;
 	providers;
 	_metadata;
-	constructor(e, t, n) {
-		this._session = e, this.providers = t, this._metadata = n;
+	requestedProviders;
+	constructor(e, t, n, r) {
+		this._session = e, this.providers = t, this._metadata = n, this.requestedProviders = r;
 	}
 	static async create(e, n = {}) {
-		let c = i(n.providers), l = {
+		let d = o(n.providers), f = await a(d), p = f.length > 0 ? f : [i];
+		n.providers !== void 0 && n.providers.length > 0 && u(d, p);
+		let m = {
 			...n.sessionOptions ?? {},
-			executionProviders: c
-		}, u = n.readMetadata !== !1, d = typeof e == "string" && u ? await o(e) : e, f = u && typeof d != "string" ? r(d) : {}, p;
+			executionProviders: p
+		}, h = n.readMetadata !== !1, g = typeof e == "string" && h ? await c(e) : e, _ = h && typeof g != "string" ? r(g) : {}, v;
 		try {
-			p = (typeof d == "string" || d instanceof Uint8Array, await a.InferenceSession.create(d, l));
+			v = (typeof g == "string" || g instanceof Uint8Array, await s.InferenceSession.create(g, m));
 		} catch (e) {
 			throw new t(`Failed to load ONNX model: ${e.message}`, { cause: e });
 		}
-		return new s(p, c, f);
+		return new l(v, p, _, d);
 	}
 	get inputNames() {
 		return this._session.inputNames;
@@ -75,6 +82,6 @@ var c = class s {
 	}
 };
 //#endregion
-export { c as OrtSession };
+export { d as OrtSession };
 
 //# sourceMappingURL=session.js.map

package/dist/vision/core/session.js.map

@@ -1 +1 @@
-{"version":3,"file":"session.js","names":[],"sources":["../../../src/vision/core/session.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Thin wrapper around `onnxruntime-web` `InferenceSession` with typed metadata.\n */\n\nimport type * as ort from \"onnxruntime-web\";\nimport * as ortRuntime from \"onnxruntime-web\";\n\nimport { InferenceError, ModelLoadError } from \"./exceptions\";\nimport { type DeclaredShape, declaredShapesFrom } from \"./graph\";\nimport { readModelMetadata } from \"./metadata\";\nimport { resolveProviders } from \"./providers\";\n\n/** Anything `InferenceSession.create` accepts. */\nexport type ModelSource = string | ArrayBufferLike | Uint8Array;\n\n/**\n * Fetch a model URL as bytes so its metadata can be read.\n *\n * Falls back to the URL itself when the fetch fails, letting ORT try its own\n * load path: losing the metadata map is a downgrade, but failing to load a model\n * that ORT could have fetched would be a regression.\n *\n * @param url Where the `.onnx` lives.\n * @returns The model bytes, or the original URL when they could not be fetched.\n */\nasync function fetchModel(url: string): Promise<Uint8Array | string> {\n    try {\n        const response = await fetch(url);\n        if (!response.ok) {\n            warnMetadataUnavailable(url, `HTTP ${response.status} ${response.statusText}`);\n            return url;\n        }\n        return new Uint8Array(await response.arrayBuffer());\n    } catch (err) {\n        warnMetadataUnavailable(url, (err as Error).message);\n        return url;\n    }\n}\n\n/**\n * Warn that a model's metadata could not be read, and say what that costs.\n *\n * The fallback itself is right — losing the metadata beats failing a load that\n * ORT could have completed on its own — but it used to be silent, and the\n * symptom it produces is remote from the cause: class names come back as\n * `class_0`, `class_1`, ... with nothing anywhere explaining why. Whoever hits\n * this needs to be told that passing `labels` is the way out.\n *\n * @param url The model URL that could not be fetched here.\n * @param reason What went wrong, as reported by `fetch`.\n */\nfunction warnMetadataUnavailable(url: string, reason: string): void {\n    console.warn(\n        `[@ort-vision-sdk/web] Could not fetch ${url} to read its metadata (${reason}). ` +\n            \"Letting ONNX Runtime load it instead: the model will work, but its baked-in \" +\n            \"class names are unavailable, so labels fall back to class_0, class_1, ... \" +\n            \"Pass `labels` explicitly to name them.\",\n    );\n}\n\nexport interface OrtSessionOptions {\n    /** Execution providers in preference order. `undefined` uses {@link DEFAULT_PROVIDERS}. */\n    readonly providers?: readonly string[];\n    /** Optional ORT session options forwarded to `InferenceSession.create`. */\n    readonly sessionOptions?: ort.InferenceSession.SessionOptions;\n    /**\n     * Whether to read the model's custom metadata map (`names`, `task`, `imgsz`).\n     * Defaults to `true`.\n     *\n     * The runtime does not expose that map, so it is read from the file itself —\n     * which means a URL model is fetched here and handed to ORT as bytes instead\n     * of letting ORT fetch it. That is the same single download either way, and\n     * it is what lets a task resolve its labels off the model. Set to `false` to\n     * keep the URL path untouched and leave {@link OrtSession.metadata} empty.\n     *\n     * `false` is also the escape hatch when a device cannot afford the bytes: the\n     * fetched buffer is dropped before ORT builds the graph (see\n     * {@link OrtSession.create}), but ORT's own load path still keeps the model out\n     * of reach of anything the SDK holds. A session built this way resolves its\n     * input size from the graph as usual — only the class names are lost, so a\n     * caller taking this route has to pass `labels` itself.\n     */\n    readonly readMetadata?: boolean;\n}\n\n/**\n * Wrap an ONNX Runtime Web `InferenceSession` with convenient metadata access.\n *\n * The wrapper exposes input/output names and the shapes the graph declares,\n * manages execution-provider selection, provides a typed {@link OrtSession.run}\n * method, and releases the native session through {@link OrtSession.release}.\n */\nexport class OrtSession {\n    private constructor(\n        private readonly _session: ort.InferenceSession,\n        public readonly providers: readonly string[],\n        private readonly _metadata: Readonly<Record<string, string>>,\n    ) {}\n\n    /**\n     * Load an ONNX model into an ORT inference session.\n     *\n     * The metadata map is read **before** the session is built, and that order is\n     * load-bearing on memory-constrained devices. ORT copies the model into its\n     * WASM heap and then allocates the graph and the weights on top of that copy;\n     * a `readModelMetadata` call placed after `InferenceSession.create` keeps the\n     * JavaScript-side buffer reachable across the whole build, so a 5 MB model\n     * costs 5 MB of JS heap plus 5 MB of WASM heap plus the weights at the same\n     * instant. Reading first makes the buffer collectable as soon as ORT has copied\n     * it — on a phone that was the difference between a session and\n     * `Can't create a session. failed to allocate a buffer of size N`.\n     *\n     * @param model Either a URL string, or a `Uint8Array`/`ArrayBuffer` containing the model bytes.\n     * @param options Provider list, pass-through `SessionOptions`, and whether to\n     *   read the model's metadata map (see {@link OrtSessionOptions.readMetadata}).\n     * @throws {@link ModelLoadError} if the model cannot be loaded.\n     */\n    static async create(model: ModelSource, options: OrtSessionOptions = {}): Promise<OrtSession> {\n        const providers = resolveProviders(options.providers);\n        const sessionOptions: ort.InferenceSession.SessionOptions = {\n            ...(options.sessionOptions ?? {}),\n            executionProviders:\n                providers as ort.InferenceSession.SessionOptions[\"executionProviders\"],\n        };\n        const wantsMetadata = options.readMetadata !== false;\n        const source = typeof model === \"string\" && wantsMetadata ? await fetchModel(model) : model;\n        const metadata =\n            wantsMetadata && typeof source !== \"string\" ? readModelMetadata(source) : {};\n\n        let session: ort.InferenceSession;\n        try {\n            if (typeof source === \"string\") {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else if (source instanceof Uint8Array) {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else {\n                session = await ortRuntime.InferenceSession.create(\n                    source as ArrayBuffer,\n                    sessionOptions,\n                );\n            }\n        } catch (err) {\n            throw new ModelLoadError(`Failed to load ONNX model: ${(err as Error).message}`, {\n                cause: err,\n            });\n        }\n\n        return new OrtSession(session, providers, metadata);\n    }\n\n    /** Names of the model's inputs, in declaration order. */\n    get inputNames(): readonly string[] {\n        return this._session.inputNames;\n    }\n\n    /** Name of the first (and usually only) input. */\n    get inputName(): string {\n        const name = this._session.inputNames[0];\n        if (name === undefined) {\n            throw new InferenceError(\"Model has no inputs.\");\n        }\n        return name;\n    }\n\n    /** Names of the model's outputs, in declaration order. */\n    get outputNames(): readonly string[] {\n        return this._session.outputNames;\n    }\n\n    /**\n     * Shapes the graph declares for its inputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Empty shapes mean the runtime\n     * reported no metadata — either a non-tensor input, or an `onnxruntime-web`\n     * older than 1.21, which predates input metadata.\n     */\n    get inputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.inputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first input, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get inputShape(): DeclaredShape {\n        return this.inputShapes[0] ?? [];\n    }\n\n    /**\n     * Shapes the graph declares for its outputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Reading them is how a task can\n     * tell how many classes a head emits without being told.\n     */\n    get outputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.outputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first output, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get outputShape(): DeclaredShape {\n        return this.outputShapes[0] ?? [];\n    }\n\n    /**\n     * The model's custom metadata map — `names`, `task`, `imgsz`, ... for an\n     * Ultralytics export.\n     *\n     * Read from the model's bytes at load time, since the runtime does not expose\n     * it. Empty when the session was created with `readMetadata: false`, from a\n     * URL that could not be fetched here, or from a model carrying no metadata.\n     */\n    get metadata(): Readonly<Record<string, string>> {\n        return this._metadata;\n    }\n\n    /**\n     * Release the native session and free its memory.\n     *\n     * Call it when a session is discarded while the page lives on — rebuilding a\n     * task at a different input size, swapping in a newer model. A failure from\n     * the runtime is ignored: a session being torn down has nothing left to fail\n     * at, and the caller is already moving on.\n     */\n    async release(): Promise<void> {\n        await this._session.release().catch(() => undefined);\n    }\n\n    /** The underlying `onnxruntime-web` session, for advanced use cases. */\n    get raw(): ort.InferenceSession {\n        return this._session;\n    }\n\n    /**\n     * Run inference and return all outputs.\n     *\n     * @param feeds Map of input name to `ort.Tensor`. Keys must match {@link inputNames}.\n     * @throws {@link InferenceError} if ORT raises any error during execution.\n     */\n    async run(feeds: Record<string, ort.Tensor>): Promise<Record<string, ort.Tensor>> {\n        try {\n            const result = await this._session.run(feeds);\n            return result as Record<string, ort.Tensor>;\n        } catch (err) {\n            throw new InferenceError(`Inference failed: ${(err as Error).message}`, { cause: err });\n        }\n    }\n}\n"],"mappings":";;;;;;AA0BA,eAAe,EAAW,GAA2C;CACjE,IAAI;EACA,IAAM,IAAW,MAAM,MAAM,CAAG;EAKhC,OAJK,EAAS,KAIP,IAAI,WAAW,MAAM,EAAS,YAAY,CAAC,KAH9C,EAAwB,GAAK,QAAQ,EAAS,OAAO,GAAG,EAAS,YAAY,GACtE;CAGf,SAAS,GAAK;EAEV,OADA,EAAwB,GAAM,EAAc,OAAO,GAC5C;CACX;AACJ;AAcA,SAAS,EAAwB,GAAa,GAAsB;CAChE,QAAQ,KACJ,yCAAyC,EAAI,yBAAyB,EAAO,kMAIjF;AACJ;AAkCA,IAAa,IAAb,MAAa,EAAW;CAEC;CACD;CACC;CAHrB,YACI,GACA,GACA,GACF;EADmB,AAFA,KAAA,WAAA,GACD,KAAA,YAAA,GACC,KAAA,YAAA;CAClB;CAoBH,aAAa,OAAO,GAAoB,IAA6B,CAAC,GAAwB;EAC1F,IAAM,IAAY,EAAiB,EAAQ,SAAS,GAC9C,IAAsD;GACxD,GAAI,EAAQ,kBAAkB,CAAC;GAC/B,oBACI;EACR,GACM,IAAgB,EAAQ,iBAAiB,IACzC,IAAS,OAAO,KAAU,YAAY,IAAgB,MAAM,EAAW,CAAK,IAAI,GAChF,IACF,KAAiB,OAAO,KAAW,WAAW,EAAkB,CAAM,IAAI,CAAC,GAE3E;EACJ,IAAI;GACA,AAKI,KALA,OAAO,KAAW,YAEX,aAAkB,YADf,MAAM,EAAW,iBAAiB,OAAO,GAAQ,CAAc;EASjF,SAAS,GAAK;GACV,MAAM,IAAI,EAAe,8BAA+B,EAAc,WAAW,EAC7E,OAAO,EACX,CAAC;EACL;EAEA,OAAO,IAAI,EAAW,GAAS,GAAW,CAAQ;CACtD;CAGA,IAAI,aAAgC;EAChC,OAAO,KAAK,SAAS;CACzB;CAGA,IAAI,YAAoB;EACpB,IAAM,IAAO,KAAK,SAAS,WAAW;EACtC,IAAI,MAAS,KAAA,GACT,MAAM,IAAI,EAAe,sBAAsB;EAEnD,OAAO;CACX;CAGA,IAAI,cAAiC;EACjC,OAAO,KAAK,SAAS;CACzB;CASA,IAAI,cAAwC;EACxC,OAAO,EACH,KAAK,SAAS,aAElB;CACJ;CAOA,IAAI,aAA4B;EAC5B,OAAO,KAAK,YAAY,MAAM,CAAC;CACnC;CAQA,IAAI,eAAyC;EACzC,OAAO,EACH,KAAK,SAAS,cAElB;CACJ;CAOA,IAAI,cAA6B;EAC7B,OAAO,KAAK,aAAa,MAAM,CAAC;CACpC;CAUA,IAAI,WAA6C;EAC7C,OAAO,KAAK;CAChB;CAUA,MAAM,UAAyB;EAC3B,MAAM,KAAK,SAAS,QAAQ,CAAC,CAAC,YAAY,KAAA,CAAS;CACvD;CAGA,IAAI,MAA4B;EAC5B,OAAO,KAAK;CAChB;CAQA,MAAM,IAAI,GAAwE;EAC9E,IAAI;GAEA,OAAO,MADc,KAAK,SAAS,IAAI,CAAK;EAEhD,SAAS,GAAK;GACV,MAAM,IAAI,EAAe,qBAAsB,EAAc,WAAW,EAAE,OAAO,EAAI,CAAC;EAC1F;CACJ;AACJ"}
+{"version":3,"file":"session.js","names":[],"sources":["../../../src/vision/core/session.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Thin wrapper around `onnxruntime-web` `InferenceSession` with typed metadata.\n */\n\nimport type * as ort from \"onnxruntime-web\";\nimport * as ortRuntime from \"onnxruntime-web\";\n\nimport { InferenceError, ModelLoadError } from \"./exceptions\";\nimport { type DeclaredShape, declaredShapesFrom } from \"./graph\";\nimport { readModelMetadata } from \"./metadata\";\nimport { FALLBACK_PROVIDER, detectProviders, resolveProviders } from \"./providers\";\n\n/** Anything `InferenceSession.create` accepts. */\nexport type ModelSource = string | ArrayBufferLike | Uint8Array;\n\n/**\n * Fetch a model URL as bytes so its metadata can be read.\n *\n * Falls back to the URL itself when the fetch fails, letting ORT try its own\n * load path: losing the metadata map is a downgrade, but failing to load a model\n * that ORT could have fetched would be a regression.\n *\n * @param url Where the `.onnx` lives.\n * @returns The model bytes, or the original URL when they could not be fetched.\n */\nasync function fetchModel(url: string): Promise<Uint8Array | string> {\n    try {\n        const response = await fetch(url);\n        if (!response.ok) {\n            warnMetadataUnavailable(url, `HTTP ${response.status} ${response.statusText}`);\n            return url;\n        }\n        return new Uint8Array(await response.arrayBuffer());\n    } catch (err) {\n        warnMetadataUnavailable(url, (err as Error).message);\n        return url;\n    }\n}\n\n/**\n * Warn that a model's metadata could not be read, and say what that costs.\n *\n * The fallback itself is right — losing the metadata beats failing a load that\n * ORT could have completed on its own — but it used to be silent, and the\n * symptom it produces is remote from the cause: class names come back as\n * `class_0`, `class_1`, ... with nothing anywhere explaining why. Whoever hits\n * this needs to be told that passing `labels` is the way out.\n *\n * @param url The model URL that could not be fetched here.\n * @param reason What went wrong, as reported by `fetch`.\n */\nfunction warnMetadataUnavailable(url: string, reason: string): void {\n    console.warn(\n        `[@ort-vision-sdk/web] Could not fetch ${url} to read its metadata (${reason}). ` +\n            \"Letting ONNX Runtime load it instead: the model will work, but its baked-in \" +\n            \"class names are unavailable, so labels fall back to class_0, class_1, ... \" +\n            \"Pass `labels` explicitly to name them.\",\n    );\n}\n\nexport interface OrtSessionOptions {\n    /**\n     * Execution providers in preference order. `undefined` uses {@link DEFAULT_PROVIDERS}.\n     *\n     * Naming one explicitly also opts into a `console.warn` when this browser\n     * cannot offer it, instead of falling back in silence.\n     */\n    readonly providers?: readonly string[];\n    /** Optional ORT session options forwarded to `InferenceSession.create`. */\n    readonly sessionOptions?: ort.InferenceSession.SessionOptions;\n    /**\n     * Whether to read the model's custom metadata map (`names`, `task`, `imgsz`).\n     * Defaults to `true`.\n     *\n     * The runtime does not expose that map, so it is read from the file itself —\n     * which means a URL model is fetched here and handed to ORT as bytes instead\n     * of letting ORT fetch it. That is the same single download either way, and\n     * it is what lets a task resolve its labels off the model. Set to `false` to\n     * keep the URL path untouched and leave {@link OrtSession.metadata} empty.\n     *\n     * `false` is also the escape hatch when a device cannot afford the bytes: the\n     * fetched buffer is dropped before ORT builds the graph (see\n     * {@link OrtSession.create}), but ORT's own load path still keeps the model out\n     * of reach of anything the SDK holds. A session built this way resolves its\n     * input size from the graph as usual — only the class names are lost, so a\n     * caller taking this route has to pass `labels` itself.\n     */\n    readonly readMetadata?: boolean;\n}\n\n/**\n * Warn when a provider named explicitly by the caller is not going to run.\n *\n * Only explicit requests are worth a warning. The default list exists precisely\n * so that falling from `webgpu` to `wasm` is the expected outcome — but a caller\n * who wrote `providers: [\"webgpu\"]` and lands on WASM has a page several times\n * slower than intended and nothing in the console to explain it.\n *\n * @param requested Providers that were asked for.\n * @param effective Providers that survived capability detection.\n */\nfunction warnOnDroppedProviders(requested: readonly string[], effective: readonly string[]): void {\n    const dropped = requested.filter((provider) => !effective.includes(provider));\n    if (dropped.length === 0) {\n        return;\n    }\n    console.warn(\n        `This browser cannot offer the requested execution provider(s) ${JSON.stringify(dropped)}; ` +\n            `the session will run on ${JSON.stringify(effective)}. Inference still produces correct ` +\n            \"results, on the fallback provider.\",\n    );\n}\n\n/**\n * Wrap an ONNX Runtime Web `InferenceSession` with convenient metadata access.\n *\n * The wrapper exposes input/output names and the shapes the graph declares,\n * manages execution-provider selection, provides a typed {@link OrtSession.run}\n * method, and releases the native session through {@link OrtSession.release}.\n */\nexport class OrtSession {\n    private constructor(\n        private readonly _session: ort.InferenceSession,\n        /**\n         * Execution providers this session is expected to run on.\n         *\n         * The requested list narrowed to what this browser can actually offer — a\n         * `webgpu` entry survives only where an adapter exists. Best-effort: ORT-Web\n         * exposes no way to ask which provider a session ended up on, so an entry\n         * here means \"not ruled out\", not \"confirmed\". See\n         * {@link requestedProviders} for what was asked for.\n         *\n         * When nothing survives — a caller asking for `webgpu` alone on a device\n         * without it — this falls back to {@link FALLBACK_PROVIDER}, which ORT-Web\n         * can always run. Handing ORT the unsatisfiable list instead makes\n         * `InferenceSession.create` reject with \"no available backend found\", so the\n         * page gets no inference at all rather than the slow-but-working fallback\n         * the `console.warn` describes. Measured in a real Chromium, where\n         * `navigator.gpu` exists but yields no adapter.\n         */\n        public readonly providers: readonly string[],\n        private readonly _metadata: Readonly<Record<string, string>>,\n        /**\n         * Execution providers that were asked for, after defaults were applied.\n         *\n         * Kept separate because ORT-Web falls back silently: a page that asks for\n         * `webgpu` on a device without it runs on WASM and is told nothing.\n         */\n        public readonly requestedProviders: readonly string[],\n    ) {}\n\n    /**\n     * Load an ONNX model into an ORT inference session.\n     *\n     * The metadata map is read **before** the session is built, and that order is\n     * load-bearing on memory-constrained devices. ORT copies the model into its\n     * WASM heap and then allocates the graph and the weights on top of that copy;\n     * a `readModelMetadata` call placed after `InferenceSession.create` keeps the\n     * JavaScript-side buffer reachable across the whole build, so a 5 MB model\n     * costs 5 MB of JS heap plus 5 MB of WASM heap plus the weights at the same\n     * instant. Reading first makes the buffer collectable as soon as ORT has copied\n     * it — on a phone that was the difference between a session and\n     * `Can't create a session. failed to allocate a buffer of size N`.\n     *\n     * @param model Either a URL string, or a `Uint8Array`/`ArrayBuffer` containing the model bytes.\n     * @param options Provider list, pass-through `SessionOptions`, and whether to\n     *   read the model's metadata map (see {@link OrtSessionOptions.readMetadata}).\n     * @throws {@link ModelLoadError} if the model cannot be loaded.\n     */\n    static async create(model: ModelSource, options: OrtSessionOptions = {}): Promise<OrtSession> {\n        const requested = resolveProviders(options.providers);\n        const detected = await detectProviders(requested);\n        const providers = detected.length > 0 ? detected : [FALLBACK_PROVIDER];\n        if (options.providers !== undefined && options.providers.length > 0) {\n            warnOnDroppedProviders(requested, providers);\n        }\n        const sessionOptions: ort.InferenceSession.SessionOptions = {\n            ...(options.sessionOptions ?? {}),\n            executionProviders:\n                providers as ort.InferenceSession.SessionOptions[\"executionProviders\"],\n        };\n        const wantsMetadata = options.readMetadata !== false;\n        const source = typeof model === \"string\" && wantsMetadata ? await fetchModel(model) : model;\n        const metadata =\n            wantsMetadata && typeof source !== \"string\" ? readModelMetadata(source) : {};\n\n        let session: ort.InferenceSession;\n        try {\n            if (typeof source === \"string\") {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else if (source instanceof Uint8Array) {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else {\n                session = await ortRuntime.InferenceSession.create(\n                    source as ArrayBuffer,\n                    sessionOptions,\n                );\n            }\n        } catch (err) {\n            throw new ModelLoadError(`Failed to load ONNX model: ${(err as Error).message}`, {\n                cause: err,\n            });\n        }\n\n        return new OrtSession(session, providers, metadata, requested);\n    }\n\n    /** Names of the model's inputs, in declaration order. */\n    get inputNames(): readonly string[] {\n        return this._session.inputNames;\n    }\n\n    /** Name of the first (and usually only) input. */\n    get inputName(): string {\n        const name = this._session.inputNames[0];\n        if (name === undefined) {\n            throw new InferenceError(\"Model has no inputs.\");\n        }\n        return name;\n    }\n\n    /** Names of the model's outputs, in declaration order. */\n    get outputNames(): readonly string[] {\n        return this._session.outputNames;\n    }\n\n    /**\n     * Shapes the graph declares for its inputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Empty shapes mean the runtime\n     * reported no metadata — either a non-tensor input, or an `onnxruntime-web`\n     * older than 1.21, which predates input metadata.\n     */\n    get inputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.inputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first input, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get inputShape(): DeclaredShape {\n        return this.inputShapes[0] ?? [];\n    }\n\n    /**\n     * Shapes the graph declares for its outputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Reading them is how a task can\n     * tell how many classes a head emits without being told.\n     */\n    get outputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.outputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first output, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get outputShape(): DeclaredShape {\n        return this.outputShapes[0] ?? [];\n    }\n\n    /**\n     * The model's custom metadata map — `names`, `task`, `imgsz`, ... for an\n     * Ultralytics export.\n     *\n     * Read from the model's bytes at load time, since the runtime does not expose\n     * it. Empty when the session was created with `readMetadata: false`, from a\n     * URL that could not be fetched here, or from a model carrying no metadata.\n     */\n    get metadata(): Readonly<Record<string, string>> {\n        return this._metadata;\n    }\n\n    /**\n     * Release the native session and free its memory.\n     *\n     * Call it when a session is discarded while the page lives on — rebuilding a\n     * task at a different input size, swapping in a newer model. A failure from\n     * the runtime is ignored: a session being torn down has nothing left to fail\n     * at, and the caller is already moving on.\n     */\n    async release(): Promise<void> {\n        await this._session.release().catch(() => undefined);\n    }\n\n    /** The underlying `onnxruntime-web` session, for advanced use cases. */\n    get raw(): ort.InferenceSession {\n        return this._session;\n    }\n\n    /**\n     * Run inference and return all outputs.\n     *\n     * @param feeds Map of input name to `ort.Tensor`. Keys must match {@link inputNames}.\n     * @throws {@link InferenceError} if ORT raises any error during execution.\n     */\n    async run(feeds: Record<string, ort.Tensor>): Promise<Record<string, ort.Tensor>> {\n        try {\n            const result = await this._session.run(feeds);\n            return result as Record<string, ort.Tensor>;\n        } catch (err) {\n            throw new InferenceError(`Inference failed: ${(err as Error).message}`, { cause: err });\n        }\n    }\n}\n"],"mappings":";;;;;;AA0BA,eAAe,EAAW,GAA2C;CACjE,IAAI;EACA,IAAM,IAAW,MAAM,MAAM,CAAG;EAKhC,OAJK,EAAS,KAIP,IAAI,WAAW,MAAM,EAAS,YAAY,CAAC,KAH9C,EAAwB,GAAK,QAAQ,EAAS,OAAO,GAAG,EAAS,YAAY,GACtE;CAGf,SAAS,GAAK;EAEV,OADA,EAAwB,GAAM,EAAc,OAAO,GAC5C;CACX;AACJ;AAcA,SAAS,EAAwB,GAAa,GAAsB;CAChE,QAAQ,KACJ,yCAAyC,EAAI,yBAAyB,EAAO,kMAIjF;AACJ;AA2CA,SAAS,EAAuB,GAA8B,GAAoC;CAC9F,IAAM,IAAU,EAAU,QAAQ,MAAa,CAAC,EAAU,SAAS,CAAQ,CAAC;CACxE,EAAQ,WAAW,KAGvB,QAAQ,KACJ,iEAAiE,KAAK,UAAU,CAAO,EAAE,4BAC1D,KAAK,UAAU,CAAS,EAAE,sEAE7D;AACJ;AASA,IAAa,IAAb,MAAa,EAAW;CAEC;CAkBD;CACC;CAOD;CA3BpB,YACI,GAkBA,GACA,GAOA,GACF;EADkB,AA1BC,KAAA,WAAA,GAkBD,KAAA,YAAA,GACC,KAAA,YAAA,GAOD,KAAA,qBAAA;CACjB;CAoBH,aAAa,OAAO,GAAoB,IAA6B,CAAC,GAAwB;EAC1F,IAAM,IAAY,EAAiB,EAAQ,SAAS,GAC9C,IAAW,MAAM,EAAgB,CAAS,GAC1C,IAAY,EAAS,SAAS,IAAI,IAAW,CAAC,CAAiB;EACrE,AAAI,EAAQ,cAAc,KAAA,KAAa,EAAQ,UAAU,SAAS,KAC9D,EAAuB,GAAW,CAAS;EAE/C,IAAM,IAAsD;GACxD,GAAI,EAAQ,kBAAkB,CAAC;GAC/B,oBACI;EACR,GACM,IAAgB,EAAQ,iBAAiB,IACzC,IAAS,OAAO,KAAU,YAAY,IAAgB,MAAM,EAAW,CAAK,IAAI,GAChF,IACF,KAAiB,OAAO,KAAW,WAAW,EAAkB,CAAM,IAAI,CAAC,GAE3E;EACJ,IAAI;GACA,AAKI,KALA,OAAO,KAAW,YAEX,aAAkB,YADf,MAAM,EAAW,iBAAiB,OAAO,GAAQ,CAAc;EASjF,SAAS,GAAK;GACV,MAAM,IAAI,EAAe,8BAA+B,EAAc,WAAW,EAC7E,OAAO,EACX,CAAC;EACL;EAEA,OAAO,IAAI,EAAW,GAAS,GAAW,GAAU,CAAS;CACjE;CAGA,IAAI,aAAgC;EAChC,OAAO,KAAK,SAAS;CACzB;CAGA,IAAI,YAAoB;EACpB,IAAM,IAAO,KAAK,SAAS,WAAW;EACtC,IAAI,MAAS,KAAA,GACT,MAAM,IAAI,EAAe,sBAAsB;EAEnD,OAAO;CACX;CAGA,IAAI,cAAiC;EACjC,OAAO,KAAK,SAAS;CACzB;CASA,IAAI,cAAwC;EACxC,OAAO,EACH,KAAK,SAAS,aAElB;CACJ;CAOA,IAAI,aAA4B;EAC5B,OAAO,KAAK,YAAY,MAAM,CAAC;CACnC;CAQA,IAAI,eAAyC;EACzC,OAAO,EACH,KAAK,SAAS,cAElB;CACJ;CAOA,IAAI,cAA6B;EAC7B,OAAO,KAAK,aAAa,MAAM,CAAC;CACpC;CAUA,IAAI,WAA6C;EAC7C,OAAO,KAAK;CAChB;CAUA,MAAM,UAAyB;EAC3B,MAAM,KAAK,SAAS,QAAQ,CAAC,CAAC,YAAY,KAAA,CAAS;CACvD;CAGA,IAAI,MAA4B;EAC5B,OAAO,KAAK;CAChB;CAQA,MAAM,IAAI,GAAwE;EAC9E,IAAI;GAEA,OAAO,MADc,KAAK,SAAS,IAAI,CAAK;EAEhD,SAAS,GAAK;GACV,MAAM,IAAI,EAAe,qBAAsB,EAAc,WAAW,EAAE,OAAO,EAAI,CAAC;EAC1F;CACJ;AACJ"}

package/dist/vision/fusion.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"fusion.cjs","names":[],"sources":["../../src/vision/fusion.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * The contract a fused pipeline carries inside its own `.onnx` file.\n *\n * A pipeline built by the Python SDK's `ort_vision_sdk.compose` is a single\n * graph that already contains the detector, the crop-and-resize bridge and the\n * classifier. Everything the runtime needs to drive it — the letterbox\n * resolution, the crop resolution, where the crops come from, how many\n * detections the graph emits, the class names of both stages — was decided at\n * fusion time and written into the model's metadata.\n *\n * This module reads it back. It is the browser half of a contract whose other\n * half lives in `ort_vision_sdk/fusion.py`: the same keys, the same encodings,\n * the same fallbacks. A pipeline fused once therefore runs identically in\n * Python and in a browser tab, off the same file.\n *\n * Building a pipeline stays a Python-side build step — there is no ONNX\n * protobuf writer here, and there is no reason for one: fusing is something you\n * do once next to your export pipeline, not in a page load.\n */\n\nimport { parseNames } from \"./core/metadata\";\n\n/**\n * Which tensor the bridge crops the detected boxes out of.\n *\n * - `\"detector_input\"`: the letterboxed tensor already fed to the detector.\n *   The fused graph then has a **single** image input, but a small object is\n *   cropped out of its downscaled copy.\n * - `\"original\"`: a second, full-resolution image input. The bridge undoes the\n *   letterbox transform in-graph and crops at native resolution. Two tensors to\n *   feed, still one session and one model load.\n */\nexport type CropSource = \"detector_input\" | \"original\";\n\n/** Value of the `ovs.kind` metadata key for a detector→classifier pipeline. */\nexport const FUSION_KIND_DETECT_CLASSIFY = \"detect_classify\";\n\n/**\n * Namespace for every metadata key the fusion writes.\n *\n * Namespaced on purpose: the detector's own Ultralytics metadata (`names`,\n * `task`, `imgsz`) is carried over into the fused model, and an un-prefixed key\n * would either collide with it or be mistaken for it.\n */\nexport const METADATA_PREFIX = \"ovs.\";\n\n/** Name of the fused graph's letterboxed detector input, `[1, 3, H, W]` float32 in `[0, 1]`. */\nexport const INPUT_IMAGE = \"images\";\n\n/** Name of the full-resolution input. Present only when `cropSource === \"original\"`. */\nexport const INPUT_SOURCE = \"source_image\";\n\n/** Name of the `[1]` float32 letterbox scale factor. Only with `cropSource === \"original\"`. */\nexport const INPUT_SCALE = \"letterbox_scale\";\n\n/** Name of the `[2]` float32 `[padLeft, padTop]`. Only with `cropSource === \"original\"`. */\nexport const INPUT_PAD = \"letterbox_pad\";\n\n/** Name of the `[K, 4]` float32 xyxy output, in **letterboxed** input pixels. */\nexport const OUTPUT_BOXES = \"boxes\";\n\n/** Name of the `[K]` float32 detection-confidence output. */\nexport const OUTPUT_SCORES = \"scores\";\n\n/** Name of the `[K]` int64 detector-class output. */\nexport const OUTPUT_CLASSES = \"classes\";\n\n/** Name of the `[1]` int64 output holding how many of the `K` rows are real. */\nexport const OUTPUT_NUM_DETECTIONS = \"num_detections\";\n\n/** Name of the `[K, numClassifierClasses]` float32 classifier output, one row per box. */\nexport const OUTPUT_PROBS = \"probs\";\n\n/** Everything a fused pipeline declares about how it must be driven. */\nexport interface FusionSpec {\n    /** Pipeline family. Only `\"detect_classify\"` exists today. */\n    readonly kind: string;\n    /** `[width, height]` the detector stage expects — the resolution to letterbox to. */\n    readonly inputSize: readonly [number, number];\n    /** `[width, height]` every crop is resampled to inside the graph. */\n    readonly cropSize: readonly [number, number];\n    /** Which tensor the crops are taken from. */\n    readonly cropSource: CropSource;\n    /**\n     * Fixed number of rows `K` every output carries, surplus zero-padded and\n     * counted by {@link OUTPUT_NUM_DETECTIONS}. `null` means the graph emits\n     * exactly as many rows as survived NMS.\n     */\n    readonly maxDetections: number | null;\n    /** Score threshold baked into the graph's NMS node. */\n    readonly confThreshold: number;\n    /** IoU threshold baked into the graph's NMS node. */\n    readonly iouThreshold: number;\n    /** Whether the classifier stage emits logits that still need a softmax. */\n    readonly applySoftmax: boolean;\n    /** Detector class names in class-id order, or `null` when the fusion recorded none. */\n    readonly detectorNames: readonly string[] | null;\n    /** Classifier class names in class-id order, or `null`. */\n    readonly classifierNames: readonly string[] | null;\n    /** Version of `ort-vision-sdk` that produced the file. */\n    readonly sdkVersion: string;\n    /** Whether driving this pipeline requires feeding the full-resolution input. */\n    readonly needsSourceImage: boolean;\n}\n\nconst KEY_KIND = \"kind\";\nconst KEY_SDK_VERSION = \"sdk_version\";\nconst KEY_INPUT_SIZE = \"input_size\";\nconst KEY_CROP_SIZE = \"crop_size\";\nconst KEY_CROP_SOURCE = \"crop_source\";\nconst KEY_MAX_DETECTIONS = \"max_detections\";\nconst KEY_CONF_THRESHOLD = \"conf_threshold\";\nconst KEY_IOU_THRESHOLD = \"iou_threshold\";\nconst KEY_APPLY_SOFTMAX = \"apply_softmax\";\nconst KEY_DETECTOR_NAMES = \"detector_names\";\nconst KEY_CLASSIFIER_NAMES = \"classifier_names\";\n\nconst DYNAMIC = \"dynamic\";\n\n/**\n * Decode a `\"640,640\"` pair.\n *\n * @param raw The encoded pair.\n * @returns `[width, height]`, or `null` when the value is missing or malformed.\n */\nfunction decodeSize(raw: string | undefined): readonly [number, number] | null {\n    if (!raw) return null;\n    const parts = raw.split(\",\");\n    if (parts.length !== 2) return null;\n    const width = Number(parts[0]);\n    const height = Number(parts[1]);\n    if (!Number.isInteger(width) || !Number.isInteger(height)) return null;\n    if (width < 1 || height < 1) return null;\n    return [width, height];\n}\n\n/**\n * Decode a float, falling back when the value is missing or malformed.\n *\n * @param raw The encoded value.\n * @param fallback Value to use when `raw` cannot be read.\n * @returns The parsed number, or `fallback`.\n */\nfunction decodeFloat(raw: string | undefined, fallback: number): number {\n    if (!raw) return fallback;\n    const value = Number(raw);\n    return Number.isFinite(value) ? value : fallback;\n}\n\n/**\n * Read a pipeline spec out of a model's custom metadata.\n *\n * Individual malformed entries fall back to the value a fusion would have used\n * by default — a single bad float is not a reason to reject an otherwise\n * loadable pipeline. A malformed resolution is fatal, because there is no safe\n * default for one.\n *\n * @param metadata A model's custom metadata map, as read by\n *   {@link readModelMetadata}.\n * @returns The decoded spec, or `null` when the model is not a fused pipeline —\n *   it carries no `ovs.kind` entry, or one naming a pipeline kind this version\n *   does not know how to drive.\n */\nexport function readFusionSpec(\n    metadata: Readonly<Record<string, string>> | undefined,\n): FusionSpec | null {\n    if (!metadata) return null;\n\n    const read: Record<string, string> = {};\n    for (const [key, value] of Object.entries(metadata)) {\n        if (key.startsWith(METADATA_PREFIX)) read[key.slice(METADATA_PREFIX.length)] = value;\n    }\n    if (read[KEY_KIND] !== FUSION_KIND_DETECT_CLASSIFY) return null;\n\n    const inputSize = decodeSize(read[KEY_INPUT_SIZE]);\n    const cropSize = decodeSize(read[KEY_CROP_SIZE]);\n    if (inputSize === null || cropSize === null) return null;\n\n    const rawMax = read[KEY_MAX_DETECTIONS] ?? DYNAMIC;\n    const parsedMax = Number(rawMax);\n    const maxDetections =\n        rawMax === DYNAMIC || !Number.isInteger(parsedMax) || parsedMax < 1 ? null : parsedMax;\n\n    const cropSource: CropSource =\n        read[KEY_CROP_SOURCE] === \"original\" ? \"original\" : \"detector_input\";\n\n    return {\n        kind: FUSION_KIND_DETECT_CLASSIFY,\n        inputSize,\n        cropSize,\n        cropSource,\n        maxDetections,\n        confThreshold: decodeFloat(read[KEY_CONF_THRESHOLD], 0.25),\n        iouThreshold: decodeFloat(read[KEY_IOU_THRESHOLD], 0.45),\n        applySoftmax: (read[KEY_APPLY_SOFTMAX] ?? \"1\") !== \"0\",\n        detectorNames: parseNames(read[KEY_DETECTOR_NAMES]),\n        classifierNames: parseNames(read[KEY_CLASSIFIER_NAMES]),\n        sdkVersion: read[KEY_SDK_VERSION] ?? \"\",\n        needsSourceImage: cropSource === \"original\",\n    };\n}\n"],"mappings":"uCAoCA,IAAa,EAA8B,kBAS9B,EAAkB,OAGlB,EAAc,SAGd,EAAe,eAGf,EAAc,kBAGd,EAAY,gBAGZ,EAAe,QAGf,EAAgB,SAGhB,EAAiB,UAGjB,EAAwB,iBAGxB,EAAe,QAkCtB,EAAW,OACX,EAAkB,cAClB,EAAiB,aACjB,EAAgB,YAChB,EAAkB,cAClB,EAAqB,iBACrB,EAAqB,iBACrB,EAAoB,gBACpB,EAAoB,gBACpB,EAAqB,iBACrB,EAAuB,mBAEvB,EAAU,UAQhB,SAAS,EAAW,EAA2D,CAC3E,GAAI,CAAC,EAAK,OAAO,KACjB,IAAM,EAAQ,EAAI,MAAM,GAAG,EAC3B,GAAI,EAAM,SAAW,EAAG,OAAO,KAC/B,IAAM,EAAQ,OAAO,EAAM,EAAE,EACvB,EAAS,OAAO,EAAM,EAAE,EAG9B,MAFI,CAAC,OAAO,UAAU,CAAK,GAAK,CAAC,OAAO,UAAU,CAAM,GACpD,EAAQ,GAAK,EAAS,EAAU,KAC7B,CAAC,EAAO,CAAM,CACzB,CASA,SAAS,EAAY,EAAyB,EAA0B,CACpE,GAAI,CAAC,EAAK,OAAO,EACjB,IAAM,EAAQ,OAAO,CAAG,EACxB,OAAO,OAAO,SAAS,CAAK,EAAI,EAAQ,CAC5C,CAgBA,SAAgB,EACZ,EACiB,CACjB,GAAI,CAAC,EAAU,OAAO,KAEtB,IAAM,EAA+B,CAAC,EACtC,IAAK,GAAM,CAAC,EAAK,KAAU,OAAO,QAAQ,CAAQ,EAC1C,EAAI,WAAA,MAA0B,IAAG,EAAK,EAAI,MAAM,CAAsB,GAAK,GAEnF,GAAI,EAAK,KAAA,kBAA2C,OAAO,KAE3D,IAAM,EAAY,EAAW,EAAK,EAAe,EAC3C,EAAW,EAAW,EAAK,EAAc,EAC/C,GAAI,IAAc,MAAQ,IAAa,KAAM,OAAO,KAEpD,IAAM,EAAS,EAAK,IAAuB,EACrC,EAAY,OAAO,CAAM,EACzB,EACF,IAAW,GAAW,CAAC,OAAO,UAAU,CAAS,GAAK,EAAY,EAAI,KAAO,EAE3E,EACF,EAAK,KAAqB,WAAa,WAAa,iBAExD,MAAO,CACH,KAAM,EACN,YACA,WACA,aACA,gBACA,cAAe,EAAY,EAAK,GAAqB,GAAI,EACzD,aAAc,EAAY,EAAK,GAAoB,GAAI,EACvD,cAAe,EAAK,IAAsB,OAAS,IACnD,cAAe,EAAA,WAAW,EAAK,EAAmB,EAClD,gBAAiB,EAAA,WAAW,EAAK,EAAqB,EACtD,WAAY,EAAK,IAAoB,GACrC,iBAAkB,IAAe,UACrC,CACJ"}
+{"version":3,"file":"fusion.cjs","names":[],"sources":["../../src/vision/fusion.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * The contract a fused pipeline carries inside its own `.onnx` file.\n *\n * A pipeline built by the Python SDK's `ort_vision_sdk.compose` is a single\n * graph that already contains the detector, the crop-and-resize bridge and the\n * classifier. Everything the runtime needs to drive it — the letterbox\n * resolution, the crop resolution, where the crops come from, how many\n * detections the graph emits, the class names of both stages — was decided at\n * fusion time and written into the model's metadata.\n *\n * This module reads it back. It is the browser half of a contract whose other\n * half lives in `ort_vision_sdk/fusion.py`: the same keys, the same encodings,\n * the same fallbacks. A pipeline fused once therefore runs identically in\n * Python and in a browser tab, off the same file.\n *\n * Building a pipeline stays a Python-side build step — there is no ONNX\n * protobuf writer here, and there is no reason for one: fusing is something you\n * do once next to your export pipeline, not in a page load.\n */\n\nimport { parseNames } from \"./core/metadata\";\n\n/**\n * Which tensor the bridge crops the detected boxes out of.\n *\n * - `\"detector_input\"`: the letterboxed tensor already fed to the detector.\n *   The fused graph then has a **single** image input, but a small object is\n *   cropped out of its downscaled copy.\n * - `\"original\"`: a second, full-resolution image input. The bridge undoes the\n *   letterbox transform in-graph and crops at native resolution. Two tensors to\n *   feed, still one session and one model load.\n */\nexport type CropSource = \"detector_input\" | \"original\";\n\n/** Value of the `ovs.kind` metadata key for a detector→classifier pipeline. */\nexport const FUSION_KIND_DETECT_CLASSIFY = \"detect_classify\";\n\n/**\n * Namespace for every metadata key the fusion writes.\n *\n * Namespaced on purpose: the detector's own Ultralytics metadata (`names`,\n * `task`, `imgsz`) is carried over into the fused model, and an un-prefixed key\n * would either collide with it or be mistaken for it.\n */\nexport const METADATA_PREFIX = \"ovs.\";\n\n/** Name of the fused graph's letterboxed detector input, `[1, 3, H, W]` float32 in `[0, 1]`. */\nexport const INPUT_IMAGE = \"images\";\n\n/** Name of the full-resolution input. Present only when `cropSource === \"original\"`. */\nexport const INPUT_SOURCE = \"source_image\";\n\n/** Name of the `[1]` float32 letterbox scale factor. Only with `cropSource === \"original\"`. */\nexport const INPUT_SCALE = \"letterbox_scale\";\n\n/** Name of the `[2]` float32 `[padLeft, padTop]`. Only with `cropSource === \"original\"`. */\nexport const INPUT_PAD = \"letterbox_pad\";\n\n/**\n * Name of the `[K, 4]` float32 xyxy output, in **letterboxed** input pixels.\n *\n * A file fused by `ort-vision-sdk` 0.9.0 or later reports the box that was\n * actually classified: clamped to the image the crop came from, exactly as\n * RoiAlign received it. Older files report the raw box, so one that ran off the\n * frame draws a rectangle wider than the region the classifier saw.\n */\nexport const OUTPUT_BOXES = \"boxes\";\n\n/** Name of the `[K]` float32 detection-confidence output. */\nexport const OUTPUT_SCORES = \"scores\";\n\n/** Name of the `[K]` int64 detector-class output. */\nexport const OUTPUT_CLASSES = \"classes\";\n\n/** Name of the `[1]` int64 output holding how many of the `K` rows are real. */\nexport const OUTPUT_NUM_DETECTIONS = \"num_detections\";\n\n/** Name of the `[K, numClassifierClasses]` float32 classifier output, one row per box. */\nexport const OUTPUT_PROBS = \"probs\";\n\n/** Everything a fused pipeline declares about how it must be driven. */\nexport interface FusionSpec {\n    /** Pipeline family. Only `\"detect_classify\"` exists today. */\n    readonly kind: string;\n    /** `[width, height]` the detector stage expects — the resolution to letterbox to. */\n    readonly inputSize: readonly [number, number];\n    /** `[width, height]` every crop is resampled to inside the graph. */\n    readonly cropSize: readonly [number, number];\n    /** Which tensor the crops are taken from. */\n    readonly cropSource: CropSource;\n    /**\n     * Fixed number of rows `K` every output carries, surplus zero-padded and\n     * counted by {@link OUTPUT_NUM_DETECTIONS}. `null` means the graph emits\n     * exactly as many rows as survived NMS.\n     */\n    readonly maxDetections: number | null;\n    /** Score threshold baked into the graph's NMS node. */\n    readonly confThreshold: number;\n    /** IoU threshold baked into the graph's NMS node. */\n    readonly iouThreshold: number;\n    /** Whether the classifier stage emits logits that still need a softmax. */\n    readonly applySoftmax: boolean;\n    /** Detector class names in class-id order, or `null` when the fusion recorded none. */\n    readonly detectorNames: readonly string[] | null;\n    /** Classifier class names in class-id order, or `null`. */\n    readonly classifierNames: readonly string[] | null;\n    /** Version of `ort-vision-sdk` that produced the file. */\n    readonly sdkVersion: string;\n    /** Whether driving this pipeline requires feeding the full-resolution input. */\n    readonly needsSourceImage: boolean;\n}\n\nconst KEY_KIND = \"kind\";\nconst KEY_SDK_VERSION = \"sdk_version\";\nconst KEY_INPUT_SIZE = \"input_size\";\nconst KEY_CROP_SIZE = \"crop_size\";\nconst KEY_CROP_SOURCE = \"crop_source\";\nconst KEY_MAX_DETECTIONS = \"max_detections\";\nconst KEY_CONF_THRESHOLD = \"conf_threshold\";\nconst KEY_IOU_THRESHOLD = \"iou_threshold\";\nconst KEY_APPLY_SOFTMAX = \"apply_softmax\";\nconst KEY_DETECTOR_NAMES = \"detector_names\";\nconst KEY_CLASSIFIER_NAMES = \"classifier_names\";\n\nconst DYNAMIC = \"dynamic\";\n\n/**\n * Decode a `\"640,640\"` pair.\n *\n * @param raw The encoded pair.\n * @returns `[width, height]`, or `null` when the value is missing or malformed.\n */\nfunction decodeSize(raw: string | undefined): readonly [number, number] | null {\n    if (!raw) return null;\n    const parts = raw.split(\",\");\n    if (parts.length !== 2) return null;\n    const width = Number(parts[0]);\n    const height = Number(parts[1]);\n    if (!Number.isInteger(width) || !Number.isInteger(height)) return null;\n    if (width < 1 || height < 1) return null;\n    return [width, height];\n}\n\n/**\n * Decode a float, falling back when the value is missing or malformed.\n *\n * @param raw The encoded value.\n * @param fallback Value to use when `raw` cannot be read.\n * @returns The parsed number, or `fallback`.\n */\nfunction decodeFloat(raw: string | undefined, fallback: number): number {\n    if (!raw) return fallback;\n    const value = Number(raw);\n    return Number.isFinite(value) ? value : fallback;\n}\n\n/**\n * Read a pipeline spec out of a model's custom metadata.\n *\n * Individual malformed entries fall back to the value a fusion would have used\n * by default — a single bad float is not a reason to reject an otherwise\n * loadable pipeline. A malformed resolution is fatal, because there is no safe\n * default for one.\n *\n * @param metadata A model's custom metadata map, as read by\n *   {@link readModelMetadata}.\n * @returns The decoded spec, or `null` when the model is not a fused pipeline —\n *   it carries no `ovs.kind` entry, or one naming a pipeline kind this version\n *   does not know how to drive.\n */\nexport function readFusionSpec(\n    metadata: Readonly<Record<string, string>> | undefined,\n): FusionSpec | null {\n    if (!metadata) return null;\n\n    const read: Record<string, string> = {};\n    for (const [key, value] of Object.entries(metadata)) {\n        if (key.startsWith(METADATA_PREFIX)) read[key.slice(METADATA_PREFIX.length)] = value;\n    }\n    if (read[KEY_KIND] !== FUSION_KIND_DETECT_CLASSIFY) return null;\n\n    const inputSize = decodeSize(read[KEY_INPUT_SIZE]);\n    const cropSize = decodeSize(read[KEY_CROP_SIZE]);\n    if (inputSize === null || cropSize === null) return null;\n\n    const rawMax = read[KEY_MAX_DETECTIONS] ?? DYNAMIC;\n    const parsedMax = Number(rawMax);\n    const maxDetections =\n        rawMax === DYNAMIC || !Number.isInteger(parsedMax) || parsedMax < 1 ? null : parsedMax;\n\n    const cropSource: CropSource =\n        read[KEY_CROP_SOURCE] === \"original\" ? \"original\" : \"detector_input\";\n\n    return {\n        kind: FUSION_KIND_DETECT_CLASSIFY,\n        inputSize,\n        cropSize,\n        cropSource,\n        maxDetections,\n        confThreshold: decodeFloat(read[KEY_CONF_THRESHOLD], 0.25),\n        iouThreshold: decodeFloat(read[KEY_IOU_THRESHOLD], 0.45),\n        applySoftmax: (read[KEY_APPLY_SOFTMAX] ?? \"1\") !== \"0\",\n        detectorNames: parseNames(read[KEY_DETECTOR_NAMES]),\n        classifierNames: parseNames(read[KEY_CLASSIFIER_NAMES]),\n        sdkVersion: read[KEY_SDK_VERSION] ?? \"\",\n        needsSourceImage: cropSource === \"original\",\n    };\n}\n"],"mappings":"uCAoCA,IAAa,EAA8B,kBAS9B,EAAkB,OAGlB,EAAc,SAGd,EAAe,eAGf,EAAc,kBAGd,EAAY,gBAUZ,EAAe,QAGf,EAAgB,SAGhB,EAAiB,UAGjB,EAAwB,iBAGxB,EAAe,QAkCtB,EAAW,OACX,EAAkB,cAClB,EAAiB,aACjB,EAAgB,YAChB,EAAkB,cAClB,EAAqB,iBACrB,EAAqB,iBACrB,EAAoB,gBACpB,EAAoB,gBACpB,EAAqB,iBACrB,EAAuB,mBAEvB,EAAU,UAQhB,SAAS,EAAW,EAA2D,CAC3E,GAAI,CAAC,EAAK,OAAO,KACjB,IAAM,EAAQ,EAAI,MAAM,GAAG,EAC3B,GAAI,EAAM,SAAW,EAAG,OAAO,KAC/B,IAAM,EAAQ,OAAO,EAAM,EAAE,EACvB,EAAS,OAAO,EAAM,EAAE,EAG9B,MAFI,CAAC,OAAO,UAAU,CAAK,GAAK,CAAC,OAAO,UAAU,CAAM,GACpD,EAAQ,GAAK,EAAS,EAAU,KAC7B,CAAC,EAAO,CAAM,CACzB,CASA,SAAS,EAAY,EAAyB,EAA0B,CACpE,GAAI,CAAC,EAAK,OAAO,EACjB,IAAM,EAAQ,OAAO,CAAG,EACxB,OAAO,OAAO,SAAS,CAAK,EAAI,EAAQ,CAC5C,CAgBA,SAAgB,EACZ,EACiB,CACjB,GAAI,CAAC,EAAU,OAAO,KAEtB,IAAM,EAA+B,CAAC,EACtC,IAAK,GAAM,CAAC,EAAK,KAAU,OAAO,QAAQ,CAAQ,EAC1C,EAAI,WAAA,MAA0B,IAAG,EAAK,EAAI,MAAM,CAAsB,GAAK,GAEnF,GAAI,EAAK,KAAA,kBAA2C,OAAO,KAE3D,IAAM,EAAY,EAAW,EAAK,EAAe,EAC3C,EAAW,EAAW,EAAK,EAAc,EAC/C,GAAI,IAAc,MAAQ,IAAa,KAAM,OAAO,KAEpD,IAAM,EAAS,EAAK,IAAuB,EACrC,EAAY,OAAO,CAAM,EACzB,EACF,IAAW,GAAW,CAAC,OAAO,UAAU,CAAS,GAAK,EAAY,EAAI,KAAO,EAE3E,EACF,EAAK,KAAqB,WAAa,WAAa,iBAExD,MAAO,CACH,KAAM,EACN,YACA,WACA,aACA,gBACA,cAAe,EAAY,EAAK,GAAqB,GAAI,EACzD,aAAc,EAAY,EAAK,GAAoB,GAAI,EACvD,cAAe,EAAK,IAAsB,OAAS,IACnD,cAAe,EAAA,WAAW,EAAK,EAAmB,EAClD,gBAAiB,EAAA,WAAW,EAAK,EAAqB,EACtD,WAAY,EAAK,IAAoB,GACrC,iBAAkB,IAAe,UACrC,CACJ"}

package/dist/vision/fusion.js.map

@@ -1 +1 @@
-{"version":3,"file":"fusion.js","names":[],"sources":["../../src/vision/fusion.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * The contract a fused pipeline carries inside its own `.onnx` file.\n *\n * A pipeline built by the Python SDK's `ort_vision_sdk.compose` is a single\n * graph that already contains the detector, the crop-and-resize bridge and the\n * classifier. Everything the runtime needs to drive it — the letterbox\n * resolution, the crop resolution, where the crops come from, how many\n * detections the graph emits, the class names of both stages — was decided at\n * fusion time and written into the model's metadata.\n *\n * This module reads it back. It is the browser half of a contract whose other\n * half lives in `ort_vision_sdk/fusion.py`: the same keys, the same encodings,\n * the same fallbacks. A pipeline fused once therefore runs identically in\n * Python and in a browser tab, off the same file.\n *\n * Building a pipeline stays a Python-side build step — there is no ONNX\n * protobuf writer here, and there is no reason for one: fusing is something you\n * do once next to your export pipeline, not in a page load.\n */\n\nimport { parseNames } from \"./core/metadata\";\n\n/**\n * Which tensor the bridge crops the detected boxes out of.\n *\n * - `\"detector_input\"`: the letterboxed tensor already fed to the detector.\n *   The fused graph then has a **single** image input, but a small object is\n *   cropped out of its downscaled copy.\n * - `\"original\"`: a second, full-resolution image input. The bridge undoes the\n *   letterbox transform in-graph and crops at native resolution. Two tensors to\n *   feed, still one session and one model load.\n */\nexport type CropSource = \"detector_input\" | \"original\";\n\n/** Value of the `ovs.kind` metadata key for a detector→classifier pipeline. */\nexport const FUSION_KIND_DETECT_CLASSIFY = \"detect_classify\";\n\n/**\n * Namespace for every metadata key the fusion writes.\n *\n * Namespaced on purpose: the detector's own Ultralytics metadata (`names`,\n * `task`, `imgsz`) is carried over into the fused model, and an un-prefixed key\n * would either collide with it or be mistaken for it.\n */\nexport const METADATA_PREFIX = \"ovs.\";\n\n/** Name of the fused graph's letterboxed detector input, `[1, 3, H, W]` float32 in `[0, 1]`. */\nexport const INPUT_IMAGE = \"images\";\n\n/** Name of the full-resolution input. Present only when `cropSource === \"original\"`. */\nexport const INPUT_SOURCE = \"source_image\";\n\n/** Name of the `[1]` float32 letterbox scale factor. Only with `cropSource === \"original\"`. */\nexport const INPUT_SCALE = \"letterbox_scale\";\n\n/** Name of the `[2]` float32 `[padLeft, padTop]`. Only with `cropSource === \"original\"`. */\nexport const INPUT_PAD = \"letterbox_pad\";\n\n/** Name of the `[K, 4]` float32 xyxy output, in **letterboxed** input pixels. */\nexport const OUTPUT_BOXES = \"boxes\";\n\n/** Name of the `[K]` float32 detection-confidence output. */\nexport const OUTPUT_SCORES = \"scores\";\n\n/** Name of the `[K]` int64 detector-class output. */\nexport const OUTPUT_CLASSES = \"classes\";\n\n/** Name of the `[1]` int64 output holding how many of the `K` rows are real. */\nexport const OUTPUT_NUM_DETECTIONS = \"num_detections\";\n\n/** Name of the `[K, numClassifierClasses]` float32 classifier output, one row per box. */\nexport const OUTPUT_PROBS = \"probs\";\n\n/** Everything a fused pipeline declares about how it must be driven. */\nexport interface FusionSpec {\n    /** Pipeline family. Only `\"detect_classify\"` exists today. */\n    readonly kind: string;\n    /** `[width, height]` the detector stage expects — the resolution to letterbox to. */\n    readonly inputSize: readonly [number, number];\n    /** `[width, height]` every crop is resampled to inside the graph. */\n    readonly cropSize: readonly [number, number];\n    /** Which tensor the crops are taken from. */\n    readonly cropSource: CropSource;\n    /**\n     * Fixed number of rows `K` every output carries, surplus zero-padded and\n     * counted by {@link OUTPUT_NUM_DETECTIONS}. `null` means the graph emits\n     * exactly as many rows as survived NMS.\n     */\n    readonly maxDetections: number | null;\n    /** Score threshold baked into the graph's NMS node. */\n    readonly confThreshold: number;\n    /** IoU threshold baked into the graph's NMS node. */\n    readonly iouThreshold: number;\n    /** Whether the classifier stage emits logits that still need a softmax. */\n    readonly applySoftmax: boolean;\n    /** Detector class names in class-id order, or `null` when the fusion recorded none. */\n    readonly detectorNames: readonly string[] | null;\n    /** Classifier class names in class-id order, or `null`. */\n    readonly classifierNames: readonly string[] | null;\n    /** Version of `ort-vision-sdk` that produced the file. */\n    readonly sdkVersion: string;\n    /** Whether driving this pipeline requires feeding the full-resolution input. */\n    readonly needsSourceImage: boolean;\n}\n\nconst KEY_KIND = \"kind\";\nconst KEY_SDK_VERSION = \"sdk_version\";\nconst KEY_INPUT_SIZE = \"input_size\";\nconst KEY_CROP_SIZE = \"crop_size\";\nconst KEY_CROP_SOURCE = \"crop_source\";\nconst KEY_MAX_DETECTIONS = \"max_detections\";\nconst KEY_CONF_THRESHOLD = \"conf_threshold\";\nconst KEY_IOU_THRESHOLD = \"iou_threshold\";\nconst KEY_APPLY_SOFTMAX = \"apply_softmax\";\nconst KEY_DETECTOR_NAMES = \"detector_names\";\nconst KEY_CLASSIFIER_NAMES = \"classifier_names\";\n\nconst DYNAMIC = \"dynamic\";\n\n/**\n * Decode a `\"640,640\"` pair.\n *\n * @param raw The encoded pair.\n * @returns `[width, height]`, or `null` when the value is missing or malformed.\n */\nfunction decodeSize(raw: string | undefined): readonly [number, number] | null {\n    if (!raw) return null;\n    const parts = raw.split(\",\");\n    if (parts.length !== 2) return null;\n    const width = Number(parts[0]);\n    const height = Number(parts[1]);\n    if (!Number.isInteger(width) || !Number.isInteger(height)) return null;\n    if (width < 1 || height < 1) return null;\n    return [width, height];\n}\n\n/**\n * Decode a float, falling back when the value is missing or malformed.\n *\n * @param raw The encoded value.\n * @param fallback Value to use when `raw` cannot be read.\n * @returns The parsed number, or `fallback`.\n */\nfunction decodeFloat(raw: string | undefined, fallback: number): number {\n    if (!raw) return fallback;\n    const value = Number(raw);\n    return Number.isFinite(value) ? value : fallback;\n}\n\n/**\n * Read a pipeline spec out of a model's custom metadata.\n *\n * Individual malformed entries fall back to the value a fusion would have used\n * by default — a single bad float is not a reason to reject an otherwise\n * loadable pipeline. A malformed resolution is fatal, because there is no safe\n * default for one.\n *\n * @param metadata A model's custom metadata map, as read by\n *   {@link readModelMetadata}.\n * @returns The decoded spec, or `null` when the model is not a fused pipeline —\n *   it carries no `ovs.kind` entry, or one naming a pipeline kind this version\n *   does not know how to drive.\n */\nexport function readFusionSpec(\n    metadata: Readonly<Record<string, string>> | undefined,\n): FusionSpec | null {\n    if (!metadata) return null;\n\n    const read: Record<string, string> = {};\n    for (const [key, value] of Object.entries(metadata)) {\n        if (key.startsWith(METADATA_PREFIX)) read[key.slice(METADATA_PREFIX.length)] = value;\n    }\n    if (read[KEY_KIND] !== FUSION_KIND_DETECT_CLASSIFY) return null;\n\n    const inputSize = decodeSize(read[KEY_INPUT_SIZE]);\n    const cropSize = decodeSize(read[KEY_CROP_SIZE]);\n    if (inputSize === null || cropSize === null) return null;\n\n    const rawMax = read[KEY_MAX_DETECTIONS] ?? DYNAMIC;\n    const parsedMax = Number(rawMax);\n    const maxDetections =\n        rawMax === DYNAMIC || !Number.isInteger(parsedMax) || parsedMax < 1 ? null : parsedMax;\n\n    const cropSource: CropSource =\n        read[KEY_CROP_SOURCE] === \"original\" ? \"original\" : \"detector_input\";\n\n    return {\n        kind: FUSION_KIND_DETECT_CLASSIFY,\n        inputSize,\n        cropSize,\n        cropSource,\n        maxDetections,\n        confThreshold: decodeFloat(read[KEY_CONF_THRESHOLD], 0.25),\n        iouThreshold: decodeFloat(read[KEY_IOU_THRESHOLD], 0.45),\n        applySoftmax: (read[KEY_APPLY_SOFTMAX] ?? \"1\") !== \"0\",\n        detectorNames: parseNames(read[KEY_DETECTOR_NAMES]),\n        classifierNames: parseNames(read[KEY_CLASSIFIER_NAMES]),\n        sdkVersion: read[KEY_SDK_VERSION] ?? \"\",\n        needsSourceImage: cropSource === \"original\",\n    };\n}\n"],"mappings":";;AAoCA,IAAa,IAA8B,mBAS9B,IAAkB,QAGlB,IAAc,UAGd,IAAe,gBAGf,IAAc,mBAGd,IAAY,iBAGZ,IAAe,SAGf,IAAgB,UAGhB,IAAiB,WAGjB,IAAwB,kBAGxB,IAAe,SAkCtB,IAAW,QACX,IAAkB,eAClB,IAAiB,cACjB,IAAgB,aAChB,IAAkB,eAClB,IAAqB,kBACrB,IAAqB,kBACrB,IAAoB,iBACpB,IAAoB,iBACpB,IAAqB,kBACrB,IAAuB,oBAEvB,IAAU;AAQhB,SAAS,EAAW,GAA2D;CAC3E,IAAI,CAAC,GAAK,OAAO;CACjB,IAAM,IAAQ,EAAI,MAAM,GAAG;CAC3B,IAAI,EAAM,WAAW,GAAG,OAAO;CAC/B,IAAM,IAAQ,OAAO,EAAM,EAAE,GACvB,IAAS,OAAO,EAAM,EAAE;CAG9B,OAFI,CAAC,OAAO,UAAU,CAAK,KAAK,CAAC,OAAO,UAAU,CAAM,KACpD,IAAQ,KAAK,IAAS,IAAU,OAC7B,CAAC,GAAO,CAAM;AACzB;AASA,SAAS,EAAY,GAAyB,GAA0B;CACpE,IAAI,CAAC,GAAK,OAAO;CACjB,IAAM,IAAQ,OAAO,CAAG;CACxB,OAAO,OAAO,SAAS,CAAK,IAAI,IAAQ;AAC5C;AAgBA,SAAgB,EACZ,GACiB;CACjB,IAAI,CAAC,GAAU,OAAO;CAEtB,IAAM,IAA+B,CAAC;CACtC,KAAK,IAAM,CAAC,GAAK,MAAU,OAAO,QAAQ,CAAQ,GAC9C,AAAI,EAAI,WAAA,MAA0B,MAAG,EAAK,EAAI,MAAM,CAAsB,KAAK;CAEnF,IAAI,EAAK,OAAA,mBAA2C,OAAO;CAE3D,IAAM,IAAY,EAAW,EAAK,EAAe,GAC3C,IAAW,EAAW,EAAK,EAAc;CAC/C,IAAI,MAAc,QAAQ,MAAa,MAAM,OAAO;CAEpD,IAAM,IAAS,EAAK,MAAuB,GACrC,IAAY,OAAO,CAAM,GACzB,IACF,MAAW,KAAW,CAAC,OAAO,UAAU,CAAS,KAAK,IAAY,IAAI,OAAO,GAE3E,IACF,EAAK,OAAqB,aAAa,aAAa;CAExD,OAAO;EACH,MAAM;EACN;EACA;EACA;EACA;EACA,eAAe,EAAY,EAAK,IAAqB,GAAI;EACzD,cAAc,EAAY,EAAK,IAAoB,GAAI;EACvD,eAAe,EAAK,MAAsB,SAAS;EACnD,eAAe,EAAW,EAAK,EAAmB;EAClD,iBAAiB,EAAW,EAAK,EAAqB;EACtD,YAAY,EAAK,MAAoB;EACrC,kBAAkB,MAAe;CACrC;AACJ"}
+{"version":3,"file":"fusion.js","names":[],"sources":["../../src/vision/fusion.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * The contract a fused pipeline carries inside its own `.onnx` file.\n *\n * A pipeline built by the Python SDK's `ort_vision_sdk.compose` is a single\n * graph that already contains the detector, the crop-and-resize bridge and the\n * classifier. Everything the runtime needs to drive it — the letterbox\n * resolution, the crop resolution, where the crops come from, how many\n * detections the graph emits, the class names of both stages — was decided at\n * fusion time and written into the model's metadata.\n *\n * This module reads it back. It is the browser half of a contract whose other\n * half lives in `ort_vision_sdk/fusion.py`: the same keys, the same encodings,\n * the same fallbacks. A pipeline fused once therefore runs identically in\n * Python and in a browser tab, off the same file.\n *\n * Building a pipeline stays a Python-side build step — there is no ONNX\n * protobuf writer here, and there is no reason for one: fusing is something you\n * do once next to your export pipeline, not in a page load.\n */\n\nimport { parseNames } from \"./core/metadata\";\n\n/**\n * Which tensor the bridge crops the detected boxes out of.\n *\n * - `\"detector_input\"`: the letterboxed tensor already fed to the detector.\n *   The fused graph then has a **single** image input, but a small object is\n *   cropped out of its downscaled copy.\n * - `\"original\"`: a second, full-resolution image input. The bridge undoes the\n *   letterbox transform in-graph and crops at native resolution. Two tensors to\n *   feed, still one session and one model load.\n */\nexport type CropSource = \"detector_input\" | \"original\";\n\n/** Value of the `ovs.kind` metadata key for a detector→classifier pipeline. */\nexport const FUSION_KIND_DETECT_CLASSIFY = \"detect_classify\";\n\n/**\n * Namespace for every metadata key the fusion writes.\n *\n * Namespaced on purpose: the detector's own Ultralytics metadata (`names`,\n * `task`, `imgsz`) is carried over into the fused model, and an un-prefixed key\n * would either collide with it or be mistaken for it.\n */\nexport const METADATA_PREFIX = \"ovs.\";\n\n/** Name of the fused graph's letterboxed detector input, `[1, 3, H, W]` float32 in `[0, 1]`. */\nexport const INPUT_IMAGE = \"images\";\n\n/** Name of the full-resolution input. Present only when `cropSource === \"original\"`. */\nexport const INPUT_SOURCE = \"source_image\";\n\n/** Name of the `[1]` float32 letterbox scale factor. Only with `cropSource === \"original\"`. */\nexport const INPUT_SCALE = \"letterbox_scale\";\n\n/** Name of the `[2]` float32 `[padLeft, padTop]`. Only with `cropSource === \"original\"`. */\nexport const INPUT_PAD = \"letterbox_pad\";\n\n/**\n * Name of the `[K, 4]` float32 xyxy output, in **letterboxed** input pixels.\n *\n * A file fused by `ort-vision-sdk` 0.9.0 or later reports the box that was\n * actually classified: clamped to the image the crop came from, exactly as\n * RoiAlign received it. Older files report the raw box, so one that ran off the\n * frame draws a rectangle wider than the region the classifier saw.\n */\nexport const OUTPUT_BOXES = \"boxes\";\n\n/** Name of the `[K]` float32 detection-confidence output. */\nexport const OUTPUT_SCORES = \"scores\";\n\n/** Name of the `[K]` int64 detector-class output. */\nexport const OUTPUT_CLASSES = \"classes\";\n\n/** Name of the `[1]` int64 output holding how many of the `K` rows are real. */\nexport const OUTPUT_NUM_DETECTIONS = \"num_detections\";\n\n/** Name of the `[K, numClassifierClasses]` float32 classifier output, one row per box. */\nexport const OUTPUT_PROBS = \"probs\";\n\n/** Everything a fused pipeline declares about how it must be driven. */\nexport interface FusionSpec {\n    /** Pipeline family. Only `\"detect_classify\"` exists today. */\n    readonly kind: string;\n    /** `[width, height]` the detector stage expects — the resolution to letterbox to. */\n    readonly inputSize: readonly [number, number];\n    /** `[width, height]` every crop is resampled to inside the graph. */\n    readonly cropSize: readonly [number, number];\n    /** Which tensor the crops are taken from. */\n    readonly cropSource: CropSource;\n    /**\n     * Fixed number of rows `K` every output carries, surplus zero-padded and\n     * counted by {@link OUTPUT_NUM_DETECTIONS}. `null` means the graph emits\n     * exactly as many rows as survived NMS.\n     */\n    readonly maxDetections: number | null;\n    /** Score threshold baked into the graph's NMS node. */\n    readonly confThreshold: number;\n    /** IoU threshold baked into the graph's NMS node. */\n    readonly iouThreshold: number;\n    /** Whether the classifier stage emits logits that still need a softmax. */\n    readonly applySoftmax: boolean;\n    /** Detector class names in class-id order, or `null` when the fusion recorded none. */\n    readonly detectorNames: readonly string[] | null;\n    /** Classifier class names in class-id order, or `null`. */\n    readonly classifierNames: readonly string[] | null;\n    /** Version of `ort-vision-sdk` that produced the file. */\n    readonly sdkVersion: string;\n    /** Whether driving this pipeline requires feeding the full-resolution input. */\n    readonly needsSourceImage: boolean;\n}\n\nconst KEY_KIND = \"kind\";\nconst KEY_SDK_VERSION = \"sdk_version\";\nconst KEY_INPUT_SIZE = \"input_size\";\nconst KEY_CROP_SIZE = \"crop_size\";\nconst KEY_CROP_SOURCE = \"crop_source\";\nconst KEY_MAX_DETECTIONS = \"max_detections\";\nconst KEY_CONF_THRESHOLD = \"conf_threshold\";\nconst KEY_IOU_THRESHOLD = \"iou_threshold\";\nconst KEY_APPLY_SOFTMAX = \"apply_softmax\";\nconst KEY_DETECTOR_NAMES = \"detector_names\";\nconst KEY_CLASSIFIER_NAMES = \"classifier_names\";\n\nconst DYNAMIC = \"dynamic\";\n\n/**\n * Decode a `\"640,640\"` pair.\n *\n * @param raw The encoded pair.\n * @returns `[width, height]`, or `null` when the value is missing or malformed.\n */\nfunction decodeSize(raw: string | undefined): readonly [number, number] | null {\n    if (!raw) return null;\n    const parts = raw.split(\",\");\n    if (parts.length !== 2) return null;\n    const width = Number(parts[0]);\n    const height = Number(parts[1]);\n    if (!Number.isInteger(width) || !Number.isInteger(height)) return null;\n    if (width < 1 || height < 1) return null;\n    return [width, height];\n}\n\n/**\n * Decode a float, falling back when the value is missing or malformed.\n *\n * @param raw The encoded value.\n * @param fallback Value to use when `raw` cannot be read.\n * @returns The parsed number, or `fallback`.\n */\nfunction decodeFloat(raw: string | undefined, fallback: number): number {\n    if (!raw) return fallback;\n    const value = Number(raw);\n    return Number.isFinite(value) ? value : fallback;\n}\n\n/**\n * Read a pipeline spec out of a model's custom metadata.\n *\n * Individual malformed entries fall back to the value a fusion would have used\n * by default — a single bad float is not a reason to reject an otherwise\n * loadable pipeline. A malformed resolution is fatal, because there is no safe\n * default for one.\n *\n * @param metadata A model's custom metadata map, as read by\n *   {@link readModelMetadata}.\n * @returns The decoded spec, or `null` when the model is not a fused pipeline —\n *   it carries no `ovs.kind` entry, or one naming a pipeline kind this version\n *   does not know how to drive.\n */\nexport function readFusionSpec(\n    metadata: Readonly<Record<string, string>> | undefined,\n): FusionSpec | null {\n    if (!metadata) return null;\n\n    const read: Record<string, string> = {};\n    for (const [key, value] of Object.entries(metadata)) {\n        if (key.startsWith(METADATA_PREFIX)) read[key.slice(METADATA_PREFIX.length)] = value;\n    }\n    if (read[KEY_KIND] !== FUSION_KIND_DETECT_CLASSIFY) return null;\n\n    const inputSize = decodeSize(read[KEY_INPUT_SIZE]);\n    const cropSize = decodeSize(read[KEY_CROP_SIZE]);\n    if (inputSize === null || cropSize === null) return null;\n\n    const rawMax = read[KEY_MAX_DETECTIONS] ?? DYNAMIC;\n    const parsedMax = Number(rawMax);\n    const maxDetections =\n        rawMax === DYNAMIC || !Number.isInteger(parsedMax) || parsedMax < 1 ? null : parsedMax;\n\n    const cropSource: CropSource =\n        read[KEY_CROP_SOURCE] === \"original\" ? \"original\" : \"detector_input\";\n\n    return {\n        kind: FUSION_KIND_DETECT_CLASSIFY,\n        inputSize,\n        cropSize,\n        cropSource,\n        maxDetections,\n        confThreshold: decodeFloat(read[KEY_CONF_THRESHOLD], 0.25),\n        iouThreshold: decodeFloat(read[KEY_IOU_THRESHOLD], 0.45),\n        applySoftmax: (read[KEY_APPLY_SOFTMAX] ?? \"1\") !== \"0\",\n        detectorNames: parseNames(read[KEY_DETECTOR_NAMES]),\n        classifierNames: parseNames(read[KEY_CLASSIFIER_NAMES]),\n        sdkVersion: read[KEY_SDK_VERSION] ?? \"\",\n        needsSourceImage: cropSource === \"original\",\n    };\n}\n"],"mappings":";;AAoCA,IAAa,IAA8B,mBAS9B,IAAkB,QAGlB,IAAc,UAGd,IAAe,gBAGf,IAAc,mBAGd,IAAY,iBAUZ,IAAe,SAGf,IAAgB,UAGhB,IAAiB,WAGjB,IAAwB,kBAGxB,IAAe,SAkCtB,IAAW,QACX,IAAkB,eAClB,IAAiB,cACjB,IAAgB,aAChB,IAAkB,eAClB,IAAqB,kBACrB,IAAqB,kBACrB,IAAoB,iBACpB,IAAoB,iBACpB,IAAqB,kBACrB,IAAuB,oBAEvB,IAAU;AAQhB,SAAS,EAAW,GAA2D;CAC3E,IAAI,CAAC,GAAK,OAAO;CACjB,IAAM,IAAQ,EAAI,MAAM,GAAG;CAC3B,IAAI,EAAM,WAAW,GAAG,OAAO;CAC/B,IAAM,IAAQ,OAAO,EAAM,EAAE,GACvB,IAAS,OAAO,EAAM,EAAE;CAG9B,OAFI,CAAC,OAAO,UAAU,CAAK,KAAK,CAAC,OAAO,UAAU,CAAM,KACpD,IAAQ,KAAK,IAAS,IAAU,OAC7B,CAAC,GAAO,CAAM;AACzB;AASA,SAAS,EAAY,GAAyB,GAA0B;CACpE,IAAI,CAAC,GAAK,OAAO;CACjB,IAAM,IAAQ,OAAO,CAAG;CACxB,OAAO,OAAO,SAAS,CAAK,IAAI,IAAQ;AAC5C;AAgBA,SAAgB,EACZ,GACiB;CACjB,IAAI,CAAC,GAAU,OAAO;CAEtB,IAAM,IAA+B,CAAC;CACtC,KAAK,IAAM,CAAC,GAAK,MAAU,OAAO,QAAQ,CAAQ,GAC9C,AAAI,EAAI,WAAA,MAA0B,MAAG,EAAK,EAAI,MAAM,CAAsB,KAAK;CAEnF,IAAI,EAAK,OAAA,mBAA2C,OAAO;CAE3D,IAAM,IAAY,EAAW,EAAK,EAAe,GAC3C,IAAW,EAAW,EAAK,EAAc;CAC/C,IAAI,MAAc,QAAQ,MAAa,MAAM,OAAO;CAEpD,IAAM,IAAS,EAAK,MAAuB,GACrC,IAAY,OAAO,CAAM,GACzB,IACF,MAAW,KAAW,CAAC,OAAO,UAAU,CAAS,KAAK,IAAY,IAAI,OAAO,GAE3E,IACF,EAAK,OAAqB,aAAa,aAAa;CAExD,OAAO;EACH,MAAM;EACN;EACA;EACA;EACA;EACA,eAAe,EAAY,EAAK,IAAqB,GAAI;EACzD,cAAc,EAAY,EAAK,IAAoB,GAAI;EACvD,eAAe,EAAK,MAAsB,SAAS;EACnD,eAAe,EAAW,EAAK,EAAmB;EAClD,iBAAiB,EAAW,EAAK,EAAqB;EACtD,YAAY,EAAK,MAAoB;EACrC,kBAAkB,MAAe;CACrC;AACJ"}

package/dist/vision/index.cjs

@@ -1,2 +1,2 @@
-const e=require("./core/exceptions.cjs"),t=require("./types.cjs"),n=require("./core/timing.cjs"),r=require("./results.cjs"),i=require("./labels.cjs"),a=require("./core/graph.cjs"),o=require("./core/metadata.cjs"),s=require("./core/providers.cjs"),c=require("./core/session.cjs"),l=require("./io/image.cjs"),u=require("./fusion.cjs"),d=require("./preprocess/image.cjs"),f=require("./preprocess/pipeline.cjs"),p=require("./postprocess/classification.cjs"),m=require("./postprocess/detection.cjs"),h=require("./postprocess/segmentation.cjs"),g=require("./tasks/base.cjs"),_=require("./tasks/classifier.cjs"),v=require("./tasks/detectClassify.cjs"),y=require("./tasks/detector.cjs"),b=require("./tasks/segmenter.cjs");var x=`0.7.1`;exports.BoundingBox=t.BoundingBox,exports.Boxes=r.Boxes,exports.COCO_CLASSES=i.COCO_CLASSES,exports.ClassificationResults=r.ClassificationResults,exports.Classifier=_.Classifier,exports.DEFAULT_PROVIDERS=s.DEFAULT_PROVIDERS,exports.DetectClassify=v.DetectClassify,exports.DetectClassifyResults=r.DetectClassifyResults,exports.DetectionResults=r.DetectionResults,exports.Detector=y.Detector,exports.FUSION_KIND_DETECT_CLASSIFY=u.FUSION_KIND_DETECT_CLASSIFY,exports.FusionError=e.FusionError,exports.INPUT_IMAGE=u.INPUT_IMAGE,exports.INPUT_PAD=u.INPUT_PAD,exports.INPUT_SCALE=u.INPUT_SCALE,exports.INPUT_SOURCE=u.INPUT_SOURCE,exports.ImageLoadError=e.ImageLoadError,exports.InferenceError=e.InferenceError,exports.LabelMapError=e.LabelMapError,exports.LetterboxPipeline=f.LetterboxPipeline,exports.METADATA_PREFIX=u.METADATA_PREFIX,exports.Mask=t.Mask,exports.Masks=r.Masks,exports.ModelLoadError=e.ModelLoadError,exports.NoDetectionsError=e.NoDetectionsError,exports.OUTPUT_BOXES=u.OUTPUT_BOXES,exports.OUTPUT_CLASSES=u.OUTPUT_CLASSES,exports.OUTPUT_NUM_DETECTIONS=u.OUTPUT_NUM_DETECTIONS,exports.OUTPUT_PROBS=u.OUTPUT_PROBS,exports.OUTPUT_SCORES=u.OUTPUT_SCORES,exports.OrtSession=c.OrtSession,exports.OrtVisionError=e.OrtVisionError,exports.Probs=r.Probs,exports.ProviderNotAvailableError=e.ProviderNotAvailableError,exports.RGBImage=t.RGBImage,exports.ResizePipeline=f.ResizePipeline,exports.SegmentationResults=r.SegmentationResults,exports.Segmenter=b.Segmenter,exports.SpeedTimer=n.SpeedTimer,exports.VERSION=x,exports.VisionTask=g.VisionTask,exports.batchedNms=m.batchedNms,exports.classificationNumClasses=a.classificationNumClasses,exports.declaredShapesFrom=a.declaredShapesFrom,exports.decodeYolo=m.decodeYolo,exports.decodeYoloAnchors=m.decodeYoloAnchors,exports.decodeYoloSeg=h.decodeYoloSeg,exports.defaultLabels=i.defaultLabels,exports.detectionNumClasses=a.detectionNumClasses,exports.fromCv2=d.fromCv2,exports.letterbox=d.letterbox,exports.letterboxToTensorData=f.letterboxToTensorData,exports.loadImage=l.loadImage,exports.modelNames=o.modelNames,exports.nms=m.nms,exports.normalize=d.normalize,exports.parseNames=o.parseNames,exports.readFusionSpec=u.readFusionSpec,exports.readModelMetadata=o.readModelMetadata,exports.requireDetections=g.requireDetections,exports.resize=d.resize,exports.resizeToTensorData=f.resizeToTensorData,exports.resolveInputSize=a.resolveInputSize,exports.resolveLabels=i.resolveLabels,exports.resolveProviders=s.resolveProviders,exports.softmax=p.softmax,exports.spatialInputSize=a.spatialInputSize,exports.toCHW=d.toCHW,exports.toCv2=d.toCv2,exports.toFloat32=d.toFloat32,exports.toFloat32Tensor=d.toFloat32Tensor,exports.toTensor=d.toTensor,exports.topK=p.topK,exports.writePlanarFloat32=f.writePlanarFloat32,exports.zeroTensorData=f.zeroTensorData;
+const e=require("./core/exceptions.cjs"),t=require("./types.cjs"),n=require("./core/timing.cjs"),r=require("./results.cjs"),i=require("./labels.cjs"),a=require("./core/graph.cjs"),o=require("./core/metadata.cjs"),s=require("./core/providers.cjs"),c=require("./core/session.cjs"),l=require("./normalization.cjs"),u=require("./io/image.cjs"),d=require("./fusion.cjs"),f=require("./preprocess/image.cjs"),p=require("./preprocess/pipeline.cjs"),m=require("./postprocess/classification.cjs"),h=require("./postprocess/detection.cjs"),g=require("./postprocess/segmentation.cjs"),_=require("./tasks/base.cjs"),v=require("./tasks/classifier.cjs"),y=require("./tasks/detectClassify.cjs"),b=require("./tasks/detector.cjs"),x=require("./tasks/segmenter.cjs");var S=`0.8.1`;exports.BoundingBox=t.BoundingBox,exports.Boxes=r.Boxes,exports.COCO_CLASSES=i.COCO_CLASSES,exports.CUSTOM_NORMALIZATION=l.CUSTOM_NORMALIZATION,exports.ClassificationResults=r.ClassificationResults,exports.Classifier=v.Classifier,exports.DEFAULT_PROVIDERS=s.DEFAULT_PROVIDERS,exports.DetectClassify=y.DetectClassify,exports.DetectClassifyResults=r.DetectClassifyResults,exports.DetectionResults=r.DetectionResults,exports.Detector=b.Detector,exports.FUSION_KIND_DETECT_CLASSIFY=d.FUSION_KIND_DETECT_CLASSIFY,exports.FusionError=e.FusionError,exports.IDENTITY_MEAN=l.IDENTITY_MEAN,exports.IDENTITY_STD=l.IDENTITY_STD,exports.IMAGENET_MEAN=l.IMAGENET_MEAN,exports.IMAGENET_STD=l.IMAGENET_STD,exports.INPUT_IMAGE=d.INPUT_IMAGE,exports.INPUT_PAD=d.INPUT_PAD,exports.INPUT_SCALE=d.INPUT_SCALE,exports.INPUT_SOURCE=d.INPUT_SOURCE,exports.ImageLoadError=e.ImageLoadError,exports.InferenceError=e.InferenceError,exports.LabelMapError=e.LabelMapError,exports.LetterboxPipeline=p.LetterboxPipeline,exports.METADATA_PREFIX=d.METADATA_PREFIX,exports.Mask=t.Mask,exports.Masks=r.Masks,exports.ModelLoadError=e.ModelLoadError,exports.NoDetectionsError=e.NoDetectionsError,exports.OUTPUT_BOXES=d.OUTPUT_BOXES,exports.OUTPUT_CLASSES=d.OUTPUT_CLASSES,exports.OUTPUT_NUM_DETECTIONS=d.OUTPUT_NUM_DETECTIONS,exports.OUTPUT_PROBS=d.OUTPUT_PROBS,exports.OUTPUT_SCORES=d.OUTPUT_SCORES,exports.OrtSession=c.OrtSession,exports.OrtVisionError=e.OrtVisionError,exports.Probs=r.Probs,exports.ProviderNotAvailableError=e.ProviderNotAvailableError,exports.RGBImage=t.RGBImage,exports.ResizePipeline=p.ResizePipeline,exports.SegmentationResults=r.SegmentationResults,exports.Segmenter=x.Segmenter,exports.SpeedTimer=n.SpeedTimer,exports.VERSION=S,exports.VisionTask=_.VisionTask,exports.batchedNms=h.batchedNms,exports.classificationNumClasses=a.classificationNumClasses,exports.declaredShapesFrom=a.declaredShapesFrom,exports.decodeYolo=h.decodeYolo,exports.decodeYoloAnchors=h.decodeYoloAnchors,exports.decodeYoloSeg=g.decodeYoloSeg,exports.defaultLabels=i.defaultLabels,exports.detectProviders=s.detectProviders,exports.detectionNumClasses=a.detectionNumClasses,exports.fromCv2=f.fromCv2,exports.isUltralyticsClassifier=l.isUltralyticsClassifier,exports.letterbox=f.letterbox,exports.letterboxToTensorData=p.letterboxToTensorData,exports.loadImage=u.loadImage,exports.modelNames=o.modelNames,exports.nms=h.nms,exports.normalize=f.normalize,exports.parseNames=o.parseNames,exports.readFusionSpec=d.readFusionSpec,exports.readModelMetadata=o.readModelMetadata,exports.requireDetections=_.requireDetections,exports.resize=f.resize,exports.resizeToTensorData=p.resizeToTensorData,exports.resolveInputSize=a.resolveInputSize,exports.resolveLabels=i.resolveLabels,exports.resolveNormalization=l.resolveNormalization,exports.resolveProviders=s.resolveProviders,exports.softmax=m.softmax,exports.spatialInputSize=a.spatialInputSize,exports.toCHW=f.toCHW,exports.toCv2=f.toCv2,exports.toFloat32=f.toFloat32,exports.toFloat32Tensor=f.toFloat32Tensor,exports.toTensor=f.toTensor,exports.topK=m.topK,exports.writePlanarFloat32=p.writePlanarFloat32,exports.zeroTensorData=p.zeroTensorData;
 //# sourceMappingURL=index.cjs.map

package/dist/vision/index.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.cjs","names":[],"sources":["../../src/vision/index.ts"],"sourcesContent":["/**\n * `tempest-react-sdk/vision` — browser computer-vision inference with ONNX\n * Runtime Web (classification, detection, segmentation).\n *\n * Vendored from `@mauriciobenjamin700/ort-vision-sdk-web@0.7.1` (MIT, same\n * author) so it ships inside this SDK without an extra package install.\n * `onnxruntime-web` stays an optional peer dependency — install it (and ship\n * the matching `.wasm` files) only when you use this subpath.\n *\n * Do not hand-edit — regenerate with `npm run vendor:vision`.\n */\n\nexport {\n    BoundingBox,\n    Mask,\n    RGBImage,\n    type ClassProbability,\n    type ClassificationResult,\n    type DetectionResult,\n    type SegmentationResult,\n} from \"./types\";\n\nexport {\n    Boxes,\n    ClassificationResults,\n    DetectClassifyResults,\n    DetectionResults,\n    Masks,\n    Probs,\n    SegmentationResults,\n} from \"./results\";\n\nexport {\n    COCO_CLASSES,\n    type LabelSpec,\n    type ResolveLabelsOptions,\n    defaultLabels,\n    resolveLabels,\n} from \"./labels\";\n\nexport {\n    FusionError,\n    ImageLoadError,\n    InferenceError,\n    LabelMapError,\n    ModelLoadError,\n    NoDetectionsError,\n    OrtVisionError,\n    ProviderNotAvailableError,\n} from \"./core/exceptions\";\n\nexport { type ModelSource, type OrtSessionOptions, OrtSession } from \"./core/session\";\nexport {\n    type DeclaredDim,\n    type DeclaredShape,\n    type ResolveInputSizeOptions,\n    classificationNumClasses,\n    declaredShapesFrom,\n    detectionNumClasses,\n    resolveInputSize,\n    spatialInputSize,\n} from \"./core/graph\";\nexport { modelNames, parseNames, readModelMetadata } from \"./core/metadata\";\nexport { DEFAULT_PROVIDERS, resolveProviders } from \"./core/providers\";\nexport { type Speed, SpeedTimer } from \"./core/timing\";\n\nexport { type ImageInput, loadImage } from \"./io/image\";\n\nexport {\n    FUSION_KIND_DETECT_CLASSIFY,\n    INPUT_IMAGE,\n    INPUT_PAD,\n    INPUT_SCALE,\n    INPUT_SOURCE,\n    METADATA_PREFIX,\n    OUTPUT_BOXES,\n    OUTPUT_CLASSES,\n    OUTPUT_NUM_DETECTIONS,\n    OUTPUT_PROBS,\n    OUTPUT_SCORES,\n    type CropSource,\n    type FusionSpec,\n    readFusionSpec,\n} from \"./fusion\";\n\nexport {\n    type LetterboxResult,\n    fromCv2,\n    letterbox,\n    normalize,\n    resize,\n    toCHW,\n    toCv2,\n    toFloat32,\n    toFloat32Tensor,\n    toTensor,\n} from \"./preprocess/image\";\n\nexport {\n    type FusedLetterboxResult,\n    type FusedResizeResult,\n    LetterboxPipeline,\n    ResizePipeline,\n    letterboxToTensorData,\n    resizeToTensorData,\n    writePlanarFloat32,\n    zeroTensorData,\n} from \"./preprocess/pipeline\";\n\nexport { type TopKResult, softmax, topK } from \"./postprocess/classification\";\n\nexport {\n    type DecodeYoloAnchorsOptions,\n    type DecodeYoloOptions,\n    type DecodedAnchors,\n    type DecodedDetection,\n    batchedNms,\n    decodeYolo,\n    decodeYoloAnchors,\n    nms,\n} from \"./postprocess/detection\";\n\nexport {\n    type DecodeYoloSegOptions,\n    type DecodedSegmentation,\n    decodeYoloSeg,\n} from \"./postprocess/segmentation\";\n\nexport { VisionTask, requireDetections } from \"./tasks/base\";\nexport {\n    type ClassifierOptions,\n    type ClassifierPredictOptions,\n    Classifier,\n} from \"./tasks/classifier\";\nexport {\n    type DetectClassifyOptions,\n    type DetectClassifyPredictOptions,\n    DetectClassify,\n} from \"./tasks/detectClassify\";\nexport {\n    type DetectorHead,\n    type DetectorOptions,\n    type DetectorPredictOptions,\n    Detector,\n} from \"./tasks/detector\";\nexport {\n    type SegmenterHead,\n    type SegmenterOptions,\n    type SegmenterPredictOptions,\n    Segmenter,\n} from \"./tasks/segmenter\";\n\nexport const VERSION: string = \"0.7.1\";\n"],"mappings":"0sBAwJA,IAAa,EAAkB"}
+{"version":3,"file":"index.cjs","names":[],"sources":["../../src/vision/index.ts"],"sourcesContent":["/**\n * `tempest-react-sdk/vision` — browser computer-vision inference with ONNX\n * Runtime Web (classification, detection, segmentation).\n *\n * Vendored from `@mauriciobenjamin700/ort-vision-sdk-web@0.8.1` (MIT, same\n * author) so it ships inside this SDK without an extra package install.\n * `onnxruntime-web` stays an optional peer dependency — install it (and ship\n * the matching `.wasm` files) only when you use this subpath.\n *\n * Do not hand-edit — regenerate with `npm run vendor:vision`.\n */\n\nexport {\n    BoundingBox,\n    Mask,\n    RGBImage,\n    type ClassProbability,\n    type ClassificationResult,\n    type DetectionResult,\n    type SegmentationResult,\n} from \"./types\";\n\nexport {\n    Boxes,\n    ClassificationResults,\n    DetectClassifyResults,\n    DetectionResults,\n    Masks,\n    Probs,\n    SegmentationResults,\n} from \"./results\";\n\nexport {\n    COCO_CLASSES,\n    type LabelSpec,\n    type ResolveLabelsOptions,\n    defaultLabels,\n    resolveLabels,\n} from \"./labels\";\n\nexport {\n    FusionError,\n    ImageLoadError,\n    InferenceError,\n    LabelMapError,\n    ModelLoadError,\n    NoDetectionsError,\n    OrtVisionError,\n    ProviderNotAvailableError,\n} from \"./core/exceptions\";\n\nexport { type ModelSource, type OrtSessionOptions, OrtSession } from \"./core/session\";\nexport {\n    type DeclaredDim,\n    type DeclaredShape,\n    type ResolveInputSizeOptions,\n    classificationNumClasses,\n    declaredShapesFrom,\n    detectionNumClasses,\n    resolveInputSize,\n    spatialInputSize,\n} from \"./core/graph\";\nexport { modelNames, parseNames, readModelMetadata } from \"./core/metadata\";\nexport { DEFAULT_PROVIDERS, detectProviders, resolveProviders } from \"./core/providers\";\nexport { type Speed, SpeedTimer } from \"./core/timing\";\nexport {\n    CUSTOM_NORMALIZATION,\n    IDENTITY_MEAN,\n    IDENTITY_STD,\n    IMAGENET_MEAN,\n    IMAGENET_STD,\n    type Normalization,\n    type ResolvedNormalization,\n    isUltralyticsClassifier,\n    resolveNormalization,\n} from \"./normalization\";\n\nexport { type ImageInput, loadImage } from \"./io/image\";\n\nexport {\n    FUSION_KIND_DETECT_CLASSIFY,\n    INPUT_IMAGE,\n    INPUT_PAD,\n    INPUT_SCALE,\n    INPUT_SOURCE,\n    METADATA_PREFIX,\n    OUTPUT_BOXES,\n    OUTPUT_CLASSES,\n    OUTPUT_NUM_DETECTIONS,\n    OUTPUT_PROBS,\n    OUTPUT_SCORES,\n    type CropSource,\n    type FusionSpec,\n    readFusionSpec,\n} from \"./fusion\";\n\nexport {\n    type LetterboxResult,\n    fromCv2,\n    letterbox,\n    normalize,\n    resize,\n    toCHW,\n    toCv2,\n    toFloat32,\n    toFloat32Tensor,\n    toTensor,\n} from \"./preprocess/image\";\n\nexport {\n    type FusedLetterboxResult,\n    type FusedResizeResult,\n    LetterboxPipeline,\n    ResizePipeline,\n    letterboxToTensorData,\n    resizeToTensorData,\n    writePlanarFloat32,\n    zeroTensorData,\n} from \"./preprocess/pipeline\";\n\nexport { type TopKResult, softmax, topK } from \"./postprocess/classification\";\n\nexport {\n    type DecodeYoloAnchorsOptions,\n    type DecodeYoloOptions,\n    type DecodedAnchors,\n    type DecodedDetection,\n    batchedNms,\n    decodeYolo,\n    decodeYoloAnchors,\n    nms,\n} from \"./postprocess/detection\";\n\nexport {\n    type DecodeYoloSegOptions,\n    type DecodedSegmentation,\n    decodeYoloSeg,\n} from \"./postprocess/segmentation\";\n\nexport { VisionTask, requireDetections } from \"./tasks/base\";\nexport {\n    type ClassifierOptions,\n    type ClassifierPredictOptions,\n    Classifier,\n} from \"./tasks/classifier\";\nexport {\n    type DetectClassifyOptions,\n    type DetectClassifyPredictOptions,\n    DetectClassify,\n} from \"./tasks/detectClassify\";\nexport {\n    type DetectorHead,\n    type DetectorOptions,\n    type DetectorPredictOptions,\n    Detector,\n} from \"./tasks/detector\";\nexport {\n    type SegmenterHead,\n    type SegmenterOptions,\n    type SegmenterPredictOptions,\n    Segmenter,\n} from \"./tasks/segmenter\";\n\nexport const VERSION: string = \"0.8.1\";\n"],"mappings":"2uBAmKA,IAAa,EAAkB"}

package/dist/vision/index.js

@@ -5,23 +5,24 @@ import { Boxes as f, ClassificationResults as p, DetectClassifyResults as m, Det
 import { COCO_CLASSES as y, defaultLabels as b, resolveLabels as x } from "./labels.js";
 import { classificationNumClasses as S, declaredShapesFrom as C, detectionNumClasses as w, resolveInputSize as T, spatialInputSize as E } from "./core/graph.js";
 import { modelNames as D, parseNames as O, readModelMetadata as k } from "./core/metadata.js";
-import { DEFAULT_PROVIDERS as A, resolveProviders as j } from "./core/providers.js";
-import { OrtSession as M } from "./core/session.js";
-import { loadImage as N } from "./io/image.js";
-import { FUSION_KIND_DETECT_CLASSIFY as P, INPUT_IMAGE as F, INPUT_PAD as I, INPUT_SCALE as L, INPUT_SOURCE as R, METADATA_PREFIX as z, OUTPUT_BOXES as B, OUTPUT_CLASSES as V, OUTPUT_NUM_DETECTIONS as H, OUTPUT_PROBS as U, OUTPUT_SCORES as W, readFusionSpec as G } from "./fusion.js";
-import { fromCv2 as K, letterbox as q, normalize as J, resize as Y, toCHW as X, toCv2 as Z, toFloat32 as Q, toFloat32Tensor as $, toTensor as ee } from "./preprocess/image.js";
-import { LetterboxPipeline as te, ResizePipeline as ne, letterboxToTensorData as re, resizeToTensorData as ie, writePlanarFloat32 as ae, zeroTensorData as oe } from "./preprocess/pipeline.js";
-import { softmax as se, topK as ce } from "./postprocess/classification.js";
-import { batchedNms as le, decodeYolo as ue, decodeYoloAnchors as de, nms as fe } from "./postprocess/detection.js";
-import { decodeYoloSeg as pe } from "./postprocess/segmentation.js";
-import { VisionTask as me, requireDetections as he } from "./tasks/base.js";
-import { Classifier as ge } from "./tasks/classifier.js";
-import { DetectClassify as _e } from "./tasks/detectClassify.js";
-import { Detector as ve } from "./tasks/detector.js";
-import { Segmenter as ye } from "./tasks/segmenter.js";
+import { DEFAULT_PROVIDERS as A, detectProviders as j, resolveProviders as M } from "./core/providers.js";
+import { OrtSession as N } from "./core/session.js";
+import { CUSTOM_NORMALIZATION as P, IDENTITY_MEAN as F, IDENTITY_STD as I, IMAGENET_MEAN as L, IMAGENET_STD as R, isUltralyticsClassifier as z, resolveNormalization as B } from "./normalization.js";
+import { loadImage as V } from "./io/image.js";
+import { FUSION_KIND_DETECT_CLASSIFY as H, INPUT_IMAGE as U, INPUT_PAD as W, INPUT_SCALE as G, INPUT_SOURCE as K, METADATA_PREFIX as q, OUTPUT_BOXES as J, OUTPUT_CLASSES as Y, OUTPUT_NUM_DETECTIONS as X, OUTPUT_PROBS as Z, OUTPUT_SCORES as Q, readFusionSpec as $ } from "./fusion.js";
+import { fromCv2 as ee, letterbox as te, normalize as ne, resize as re, toCHW as ie, toCv2 as ae, toFloat32 as oe, toFloat32Tensor as se, toTensor as ce } from "./preprocess/image.js";
+import { LetterboxPipeline as le, ResizePipeline as ue, letterboxToTensorData as de, resizeToTensorData as fe, writePlanarFloat32 as pe, zeroTensorData as me } from "./preprocess/pipeline.js";
+import { softmax as he, topK as ge } from "./postprocess/classification.js";
+import { batchedNms as _e, decodeYolo as ve, decodeYoloAnchors as ye, nms as be } from "./postprocess/detection.js";
+import { decodeYoloSeg as xe } from "./postprocess/segmentation.js";
+import { VisionTask as Se, requireDetections as Ce } from "./tasks/base.js";
+import { Classifier as we } from "./tasks/classifier.js";
+import { DetectClassify as Te } from "./tasks/detectClassify.js";
+import { Detector as Ee } from "./tasks/detector.js";
+import { Segmenter as De } from "./tasks/segmenter.js";
 //#region src/vision/index.ts
-var be = "0.7.1";
+var Oe = "0.8.1";
 //#endregion
-export { c as BoundingBox, f as Boxes, y as COCO_CLASSES, p as ClassificationResults, ge as Classifier, A as DEFAULT_PROVIDERS, _e as DetectClassify, m as DetectClassifyResults, h as DetectionResults, ve as Detector, P as FUSION_KIND_DETECT_CLASSIFY, e as FusionError, F as INPUT_IMAGE, I as INPUT_PAD, L as INPUT_SCALE, R as INPUT_SOURCE, t as ImageLoadError, n as InferenceError, r as LabelMapError, te as LetterboxPipeline, z as METADATA_PREFIX, l as Mask, g as Masks, i as ModelLoadError, a as NoDetectionsError, B as OUTPUT_BOXES, V as OUTPUT_CLASSES, H as OUTPUT_NUM_DETECTIONS, U as OUTPUT_PROBS, W as OUTPUT_SCORES, M as OrtSession, o as OrtVisionError, _ as Probs, s as ProviderNotAvailableError, u as RGBImage, ne as ResizePipeline, v as SegmentationResults, ye as Segmenter, d as SpeedTimer, be as VERSION, me as VisionTask, le as batchedNms, S as classificationNumClasses, C as declaredShapesFrom, ue as decodeYolo, de as decodeYoloAnchors, pe as decodeYoloSeg, b as defaultLabels, w as detectionNumClasses, K as fromCv2, q as letterbox, re as letterboxToTensorData, N as loadImage, D as modelNames, fe as nms, J as normalize, O as parseNames, G as readFusionSpec, k as readModelMetadata, he as requireDetections, Y as resize, ie as resizeToTensorData, T as resolveInputSize, x as resolveLabels, j as resolveProviders, se as softmax, E as spatialInputSize, X as toCHW, Z as toCv2, Q as toFloat32, $ as toFloat32Tensor, ee as toTensor, ce as topK, ae as writePlanarFloat32, oe as zeroTensorData };
+export { c as BoundingBox, f as Boxes, y as COCO_CLASSES, P as CUSTOM_NORMALIZATION, p as ClassificationResults, we as Classifier, A as DEFAULT_PROVIDERS, Te as DetectClassify, m as DetectClassifyResults, h as DetectionResults, Ee as Detector, H as FUSION_KIND_DETECT_CLASSIFY, e as FusionError, F as IDENTITY_MEAN, I as IDENTITY_STD, L as IMAGENET_MEAN, R as IMAGENET_STD, U as INPUT_IMAGE, W as INPUT_PAD, G as INPUT_SCALE, K as INPUT_SOURCE, t as ImageLoadError, n as InferenceError, r as LabelMapError, le as LetterboxPipeline, q as METADATA_PREFIX, l as Mask, g as Masks, i as ModelLoadError, a as NoDetectionsError, J as OUTPUT_BOXES, Y as OUTPUT_CLASSES, X as OUTPUT_NUM_DETECTIONS, Z as OUTPUT_PROBS, Q as OUTPUT_SCORES, N as OrtSession, o as OrtVisionError, _ as Probs, s as ProviderNotAvailableError, u as RGBImage, ue as ResizePipeline, v as SegmentationResults, De as Segmenter, d as SpeedTimer, Oe as VERSION, Se as VisionTask, _e as batchedNms, S as classificationNumClasses, C as declaredShapesFrom, ve as decodeYolo, ye as decodeYoloAnchors, xe as decodeYoloSeg, b as defaultLabels, j as detectProviders, w as detectionNumClasses, ee as fromCv2, z as isUltralyticsClassifier, te as letterbox, de as letterboxToTensorData, V as loadImage, D as modelNames, be as nms, ne as normalize, O as parseNames, $ as readFusionSpec, k as readModelMetadata, Ce as requireDetections, re as resize, fe as resizeToTensorData, T as resolveInputSize, x as resolveLabels, B as resolveNormalization, M as resolveProviders, he as softmax, E as spatialInputSize, ie as toCHW, ae as toCv2, oe as toFloat32, se as toFloat32Tensor, ce as toTensor, ge as topK, pe as writePlanarFloat32, me as zeroTensorData };
 
 //# sourceMappingURL=index.js.map

package/dist/vision/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","names":[],"sources":["../../src/vision/index.ts"],"sourcesContent":["/**\n * `tempest-react-sdk/vision` — browser computer-vision inference with ONNX\n * Runtime Web (classification, detection, segmentation).\n *\n * Vendored from `@mauriciobenjamin700/ort-vision-sdk-web@0.7.1` (MIT, same\n * author) so it ships inside this SDK without an extra package install.\n * `onnxruntime-web` stays an optional peer dependency — install it (and ship\n * the matching `.wasm` files) only when you use this subpath.\n *\n * Do not hand-edit — regenerate with `npm run vendor:vision`.\n */\n\nexport {\n    BoundingBox,\n    Mask,\n    RGBImage,\n    type ClassProbability,\n    type ClassificationResult,\n    type DetectionResult,\n    type SegmentationResult,\n} from \"./types\";\n\nexport {\n    Boxes,\n    ClassificationResults,\n    DetectClassifyResults,\n    DetectionResults,\n    Masks,\n    Probs,\n    SegmentationResults,\n} from \"./results\";\n\nexport {\n    COCO_CLASSES,\n    type LabelSpec,\n    type ResolveLabelsOptions,\n    defaultLabels,\n    resolveLabels,\n} from \"./labels\";\n\nexport {\n    FusionError,\n    ImageLoadError,\n    InferenceError,\n    LabelMapError,\n    ModelLoadError,\n    NoDetectionsError,\n    OrtVisionError,\n    ProviderNotAvailableError,\n} from \"./core/exceptions\";\n\nexport { type ModelSource, type OrtSessionOptions, OrtSession } from \"./core/session\";\nexport {\n    type DeclaredDim,\n    type DeclaredShape,\n    type ResolveInputSizeOptions,\n    classificationNumClasses,\n    declaredShapesFrom,\n    detectionNumClasses,\n    resolveInputSize,\n    spatialInputSize,\n} from \"./core/graph\";\nexport { modelNames, parseNames, readModelMetadata } from \"./core/metadata\";\nexport { DEFAULT_PROVIDERS, resolveProviders } from \"./core/providers\";\nexport { type Speed, SpeedTimer } from \"./core/timing\";\n\nexport { type ImageInput, loadImage } from \"./io/image\";\n\nexport {\n    FUSION_KIND_DETECT_CLASSIFY,\n    INPUT_IMAGE,\n    INPUT_PAD,\n    INPUT_SCALE,\n    INPUT_SOURCE,\n    METADATA_PREFIX,\n    OUTPUT_BOXES,\n    OUTPUT_CLASSES,\n    OUTPUT_NUM_DETECTIONS,\n    OUTPUT_PROBS,\n    OUTPUT_SCORES,\n    type CropSource,\n    type FusionSpec,\n    readFusionSpec,\n} from \"./fusion\";\n\nexport {\n    type LetterboxResult,\n    fromCv2,\n    letterbox,\n    normalize,\n    resize,\n    toCHW,\n    toCv2,\n    toFloat32,\n    toFloat32Tensor,\n    toTensor,\n} from \"./preprocess/image\";\n\nexport {\n    type FusedLetterboxResult,\n    type FusedResizeResult,\n    LetterboxPipeline,\n    ResizePipeline,\n    letterboxToTensorData,\n    resizeToTensorData,\n    writePlanarFloat32,\n    zeroTensorData,\n} from \"./preprocess/pipeline\";\n\nexport { type TopKResult, softmax, topK } from \"./postprocess/classification\";\n\nexport {\n    type DecodeYoloAnchorsOptions,\n    type DecodeYoloOptions,\n    type DecodedAnchors,\n    type DecodedDetection,\n    batchedNms,\n    decodeYolo,\n    decodeYoloAnchors,\n    nms,\n} from \"./postprocess/detection\";\n\nexport {\n    type DecodeYoloSegOptions,\n    type DecodedSegmentation,\n    decodeYoloSeg,\n} from \"./postprocess/segmentation\";\n\nexport { VisionTask, requireDetections } from \"./tasks/base\";\nexport {\n    type ClassifierOptions,\n    type ClassifierPredictOptions,\n    Classifier,\n} from \"./tasks/classifier\";\nexport {\n    type DetectClassifyOptions,\n    type DetectClassifyPredictOptions,\n    DetectClassify,\n} from \"./tasks/detectClassify\";\nexport {\n    type DetectorHead,\n    type DetectorOptions,\n    type DetectorPredictOptions,\n    Detector,\n} from \"./tasks/detector\";\nexport {\n    type SegmenterHead,\n    type SegmenterOptions,\n    type SegmenterPredictOptions,\n    Segmenter,\n} from \"./tasks/segmenter\";\n\nexport const VERSION: string = \"0.7.1\";\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;AAwJA,IAAa,KAAkB"}
+{"version":3,"file":"index.js","names":[],"sources":["../../src/vision/index.ts"],"sourcesContent":["/**\n * `tempest-react-sdk/vision` — browser computer-vision inference with ONNX\n * Runtime Web (classification, detection, segmentation).\n *\n * Vendored from `@mauriciobenjamin700/ort-vision-sdk-web@0.8.1` (MIT, same\n * author) so it ships inside this SDK without an extra package install.\n * `onnxruntime-web` stays an optional peer dependency — install it (and ship\n * the matching `.wasm` files) only when you use this subpath.\n *\n * Do not hand-edit — regenerate with `npm run vendor:vision`.\n */\n\nexport {\n    BoundingBox,\n    Mask,\n    RGBImage,\n    type ClassProbability,\n    type ClassificationResult,\n    type DetectionResult,\n    type SegmentationResult,\n} from \"./types\";\n\nexport {\n    Boxes,\n    ClassificationResults,\n    DetectClassifyResults,\n    DetectionResults,\n    Masks,\n    Probs,\n    SegmentationResults,\n} from \"./results\";\n\nexport {\n    COCO_CLASSES,\n    type LabelSpec,\n    type ResolveLabelsOptions,\n    defaultLabels,\n    resolveLabels,\n} from \"./labels\";\n\nexport {\n    FusionError,\n    ImageLoadError,\n    InferenceError,\n    LabelMapError,\n    ModelLoadError,\n    NoDetectionsError,\n    OrtVisionError,\n    ProviderNotAvailableError,\n} from \"./core/exceptions\";\n\nexport { type ModelSource, type OrtSessionOptions, OrtSession } from \"./core/session\";\nexport {\n    type DeclaredDim,\n    type DeclaredShape,\n    type ResolveInputSizeOptions,\n    classificationNumClasses,\n    declaredShapesFrom,\n    detectionNumClasses,\n    resolveInputSize,\n    spatialInputSize,\n} from \"./core/graph\";\nexport { modelNames, parseNames, readModelMetadata } from \"./core/metadata\";\nexport { DEFAULT_PROVIDERS, detectProviders, resolveProviders } from \"./core/providers\";\nexport { type Speed, SpeedTimer } from \"./core/timing\";\nexport {\n    CUSTOM_NORMALIZATION,\n    IDENTITY_MEAN,\n    IDENTITY_STD,\n    IMAGENET_MEAN,\n    IMAGENET_STD,\n    type Normalization,\n    type ResolvedNormalization,\n    isUltralyticsClassifier,\n    resolveNormalization,\n} from \"./normalization\";\n\nexport { type ImageInput, loadImage } from \"./io/image\";\n\nexport {\n    FUSION_KIND_DETECT_CLASSIFY,\n    INPUT_IMAGE,\n    INPUT_PAD,\n    INPUT_SCALE,\n    INPUT_SOURCE,\n    METADATA_PREFIX,\n    OUTPUT_BOXES,\n    OUTPUT_CLASSES,\n    OUTPUT_NUM_DETECTIONS,\n    OUTPUT_PROBS,\n    OUTPUT_SCORES,\n    type CropSource,\n    type FusionSpec,\n    readFusionSpec,\n} from \"./fusion\";\n\nexport {\n    type LetterboxResult,\n    fromCv2,\n    letterbox,\n    normalize,\n    resize,\n    toCHW,\n    toCv2,\n    toFloat32,\n    toFloat32Tensor,\n    toTensor,\n} from \"./preprocess/image\";\n\nexport {\n    type FusedLetterboxResult,\n    type FusedResizeResult,\n    LetterboxPipeline,\n    ResizePipeline,\n    letterboxToTensorData,\n    resizeToTensorData,\n    writePlanarFloat32,\n    zeroTensorData,\n} from \"./preprocess/pipeline\";\n\nexport { type TopKResult, softmax, topK } from \"./postprocess/classification\";\n\nexport {\n    type DecodeYoloAnchorsOptions,\n    type DecodeYoloOptions,\n    type DecodedAnchors,\n    type DecodedDetection,\n    batchedNms,\n    decodeYolo,\n    decodeYoloAnchors,\n    nms,\n} from \"./postprocess/detection\";\n\nexport {\n    type DecodeYoloSegOptions,\n    type DecodedSegmentation,\n    decodeYoloSeg,\n} from \"./postprocess/segmentation\";\n\nexport { VisionTask, requireDetections } from \"./tasks/base\";\nexport {\n    type ClassifierOptions,\n    type ClassifierPredictOptions,\n    Classifier,\n} from \"./tasks/classifier\";\nexport {\n    type DetectClassifyOptions,\n    type DetectClassifyPredictOptions,\n    DetectClassify,\n} from \"./tasks/detectClassify\";\nexport {\n    type DetectorHead,\n    type DetectorOptions,\n    type DetectorPredictOptions,\n    Detector,\n} from \"./tasks/detector\";\nexport {\n    type SegmenterHead,\n    type SegmenterOptions,\n    type SegmenterPredictOptions,\n    Segmenter,\n} from \"./tasks/segmenter\";\n\nexport const VERSION: string = \"0.8.1\";\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;AAmKA,IAAa,KAAkB"}

package/dist/vision/normalization.cjs

@@ -0,0 +1,2 @@
+var e=[.485,.456,.406],t=[.229,.224,.225],n=[0,0,0],r=[1,1,1],i=`custom`,a={imagenet:[e,t],ultralytics:[n,r],none:[n,r]};function o(e){return(e.author??``).trim().toLowerCase()===`ultralytics`&&(e.task??``).trim().toLowerCase()===`classify`}function s(e,t={}){let s=t.normalization??`auto`,l=t.mean!==void 0||t.std!==void 0;if(s!==`auto`&&a[s]===void 0)throw RangeError(`normalization must be one of 'auto', 'imagenet', 'ultralytics', 'none'; got ${JSON.stringify(s)}.`);if(l&&s!==`auto`)throw RangeError(`Pass either normalization=${JSON.stringify(s)} or explicit mean/std, not both.`);let u=o(e),d=s===`auto`?u?`ultralytics`:`imagenet`:s,[f,p]=a[d];if(!l)return{name:d,mean:f,std:p};let m=t.mean??f,h=t.std??p;return u&&!(c(m,n)&&c(h,r))&&console.warn(`The classifier is an Ultralytics export, whose classification head is trained on raw [0, 1] images, but mean=${JSON.stringify(m)} / std=${JSON.stringify(h)} was requested. It will be fed images normalized in a way it never saw in training, which degrades accuracy without raising anything. Drop mean/std to let normalization='auto' pick the identity.`),{name:i,mean:m,std:h}}function c(e,t){return e.every((e,n)=>e===t[n])}exports.CUSTOM_NORMALIZATION=i,exports.IDENTITY_MEAN=n,exports.IDENTITY_STD=r,exports.IMAGENET_MEAN=e,exports.IMAGENET_STD=t,exports.isUltralyticsClassifier=o,exports.resolveNormalization=s;
+//# sourceMappingURL=normalization.cjs.map

package/dist/vision/normalization.cjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"normalization.cjs","names":[],"sources":["../../src/vision/normalization.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Which preprocessing a classifier expects its input to have had.\n *\n * A classifier is trained on a specific tensor, and feeding it a differently\n * prepared one degrades it silently — no exception, no warning, just worse\n * predictions. The two families this SDK sees most disagree completely:\n * torchvision-style models want the ImageNet mean and deviation subtracted and\n * divided out, while an Ultralytics classification head consumes raw `[0, 1]`.\n *\n * Guessing wrong is not detectable from the outside, but it is not a guess: an\n * Ultralytics export stamps `author` and `task` into its own metadata, and every\n * task in this SDK already reads that map for the class names.\n *\n * Mirrors `ort_vision_sdk.normalization` in the Python SDK; the two must agree,\n * because the same model file is driven by both.\n */\n\n/**\n * Which preprocessing the classifier expects its input to have had.\n *\n * - `\"auto\"` (the default wherever it is accepted) reads the model's own export\n *   metadata and picks. An Ultralytics classification head gets\n *   `\"ultralytics\"`; anything else gets `\"imagenet\"`.\n * - `\"imagenet\"` subtracts the ImageNet mean and divides by the ImageNet\n *   deviation — the torchvision convention.\n * - `\"ultralytics\"` leaves the image in `[0, 1]`. Ultralytics' own classifier\n *   applies no mean/std at all, so anything else feeds it images it never saw\n *   in training.\n * - `\"none\"` is the same arithmetic as `\"ultralytics\"` — identity — under a\n *   name that says \"this model wants raw `[0, 1]`\" rather than naming a vendor.\n */\nexport type Normalization = \"auto\" | \"imagenet\" | \"ultralytics\" | \"none\";\n\n/** Per-channel RGB mean of ImageNet, the torchvision preprocessing convention. */\nexport const IMAGENET_MEAN: readonly [number, number, number] = [0.485, 0.456, 0.406];\n\n/** Per-channel RGB standard deviation of ImageNet. */\nexport const IMAGENET_STD: readonly [number, number, number] = [0.229, 0.224, 0.225];\n\n/** Mean that leaves an image untouched — what an Ultralytics classifier expects. */\nexport const IDENTITY_MEAN: readonly [number, number, number] = [0, 0, 0];\n\n/** Deviation that leaves an image untouched — what an Ultralytics classifier expects. */\nexport const IDENTITY_STD: readonly [number, number, number] = [1, 1, 1];\n\n/** Name reported when the caller supplied `mean`/`std` directly. */\nexport const CUSTOM_NORMALIZATION = \"custom\";\n\nconst PRESETS: Readonly<\n    Record<string, readonly [readonly [number, number, number], readonly [number, number, number]]>\n> = {\n    imagenet: [IMAGENET_MEAN, IMAGENET_STD],\n    ultralytics: [IDENTITY_MEAN, IDENTITY_STD],\n    none: [IDENTITY_MEAN, IDENTITY_STD],\n};\n\n/** What {@link resolveNormalization} settled on. */\nexport interface ResolvedNormalization {\n    /** The preset name, or `\"custom\"` when the caller supplied the values. */\n    readonly name: string;\n    /** Per-channel mean to subtract. */\n    readonly mean: readonly [number, number, number];\n    /** Per-channel deviation to divide by. */\n    readonly std: readonly [number, number, number];\n}\n\n/**\n * Whether a metadata map came out of `YOLO(...).export(format=\"onnx\")`.\n *\n * Every Ultralytics export stamps `author` and `task` into its metadata, and the\n * pair is unambiguous: `\"Ultralytics\"` plus `\"classify\"` is a classification head\n * from that codebase and nothing else.\n *\n * @param metadata The model's custom metadata map.\n */\nexport function isUltralyticsClassifier(metadata: Readonly<Record<string, string>>): boolean {\n    return (\n        (metadata.author ?? \"\").trim().toLowerCase() === \"ultralytics\" &&\n        (metadata.task ?? \"\").trim().toLowerCase() === \"classify\"\n    );\n}\n\n/**\n * Settle which `mean`/`std` to apply, and what to call the choice.\n *\n * Explicit `mean`/`std` always win — they are the escape hatch for a model whose\n * preprocessing neither preset describes. Anything they leave open falls back to\n * the preset, so passing only a `mean` does not silently reset the deviation\n * to 1.\n *\n * Warns (via `console.warn`) when the model is an Ultralytics export and the\n * supplied values are not the identity it was trained with. Nothing fails in\n * that case: the prediction has the right shape and is simply less accurate,\n * which is exactly why it is worth saying out loud.\n *\n * @param metadata The model's custom metadata map, read to detect the family.\n * @param options The preset asked for, plus any explicit `mean`/`std`.\n * @throws {RangeError} If `normalization` is not a known preset, or names one\n *   while `mean`/`std` are also supplied — two answers to the same question.\n */\nexport function resolveNormalization(\n    metadata: Readonly<Record<string, string>>,\n    options: {\n        readonly normalization?: Normalization;\n        readonly mean?: readonly [number, number, number];\n        readonly std?: readonly [number, number, number];\n    } = {},\n): ResolvedNormalization {\n    const normalization = options.normalization ?? \"auto\";\n    const explicit = options.mean !== undefined || options.std !== undefined;\n\n    if (normalization !== \"auto\" && PRESETS[normalization] === undefined) {\n        throw new RangeError(\n            `normalization must be one of 'auto', 'imagenet', 'ultralytics', 'none'; ` +\n                `got ${JSON.stringify(normalization)}.`,\n        );\n    }\n    if (explicit && normalization !== \"auto\") {\n        throw new RangeError(\n            `Pass either normalization=${JSON.stringify(normalization)} or explicit mean/std, not both.`,\n        );\n    }\n\n    const ultralytics = isUltralyticsClassifier(metadata);\n    const preset =\n        normalization === \"auto\" ? (ultralytics ? \"ultralytics\" : \"imagenet\") : normalization;\n    const [presetMean, presetStd] = PRESETS[preset] as readonly [\n        readonly [number, number, number],\n        readonly [number, number, number],\n    ];\n\n    if (!explicit) {\n        return { name: preset, mean: presetMean, std: presetStd };\n    }\n\n    const mean = options.mean ?? presetMean;\n    const std = options.std ?? presetStd;\n    if (ultralytics && !(isIdentity(mean, IDENTITY_MEAN) && isIdentity(std, IDENTITY_STD))) {\n        console.warn(\n            \"The classifier is an Ultralytics export, whose classification head is trained on raw \" +\n                `[0, 1] images, but mean=${JSON.stringify(mean)} / std=${JSON.stringify(std)} was ` +\n                \"requested. It will be fed images normalized in a way it never saw in training, which \" +\n                \"degrades accuracy without raising anything. Drop mean/std to let normalization='auto' \" +\n                \"pick the identity.\",\n        );\n    }\n    return { name: CUSTOM_NORMALIZATION, mean, std };\n}\n\n/**\n * Whether two channel triples are equal.\n *\n * @param value The triple to check.\n * @param reference The triple to compare against.\n */\nfunction isIdentity(\n    value: readonly [number, number, number],\n    reference: readonly [number, number, number],\n): boolean {\n    return value.every((entry, index) => entry === reference[index]);\n}\n"],"mappings":"AAmCA,IAAa,EAAmD,CAAC,KAAO,KAAO,IAAK,EAGvE,EAAkD,CAAC,KAAO,KAAO,IAAK,EAGtE,EAAmD,CAAC,EAAG,EAAG,CAAC,EAG3D,EAAkD,CAAC,EAAG,EAAG,CAAC,EAG1D,EAAuB,SAE9B,EAEF,CACA,SAAU,CAAC,EAAe,CAAY,EACtC,YAAa,CAAC,EAAe,CAAY,EACzC,KAAM,CAAC,EAAe,CAAY,CACtC,EAqBA,SAAgB,EAAwB,EAAqD,CACzF,OACK,EAAS,QAAU,GAAA,CAAI,KAAK,CAAC,CAAC,YAAY,IAAM,gBAChD,EAAS,MAAQ,GAAA,CAAI,KAAK,CAAC,CAAC,YAAY,IAAM,UAEvD,CAoBA,SAAgB,EACZ,EACA,EAII,CAAC,EACgB,CACrB,IAAM,EAAgB,EAAQ,eAAiB,OACzC,EAAW,EAAQ,OAAS,IAAA,IAAa,EAAQ,MAAQ,IAAA,GAE/D,GAAI,IAAkB,QAAU,EAAQ,KAAmB,IAAA,GACvD,MAAU,WACN,+EACW,KAAK,UAAU,CAAa,EAAE,EAC7C,EAEJ,GAAI,GAAY,IAAkB,OAC9B,MAAU,WACN,6BAA6B,KAAK,UAAU,CAAa,EAAE,iCAC/D,EAGJ,IAAM,EAAc,EAAwB,CAAQ,EAC9C,EACF,IAAkB,OAAU,EAAc,cAAgB,WAAc,EACtE,CAAC,EAAY,GAAa,EAAQ,GAKxC,GAAI,CAAC,EACD,MAAO,CAAE,KAAM,EAAQ,KAAM,EAAY,IAAK,CAAU,EAG5D,IAAM,EAAO,EAAQ,MAAQ,EACvB,EAAM,EAAQ,KAAO,EAU3B,OATI,GAAe,EAAE,EAAW,EAAM,CAAa,GAAK,EAAW,EAAK,CAAY,IAChF,QAAQ,KACJ,gHAC+B,KAAK,UAAU,CAAI,EAAE,SAAS,KAAK,UAAU,CAAG,EAAE,mMAIrF,EAEG,CAAE,KAAM,EAAsB,OAAM,KAAI,CACnD,CAQA,SAAS,EACL,EACA,EACO,CACP,OAAO,EAAM,OAAO,EAAO,IAAU,IAAU,EAAU,EAAM,CACnE"}

package/dist/vision/normalization.js

@@ -0,0 +1,49 @@
+//#region src/vision/normalization.ts
+var e = [
+	.485,
+	.456,
+	.406
+], t = [
+	.229,
+	.224,
+	.225
+], n = [
+	0,
+	0,
+	0
+], r = [
+	1,
+	1,
+	1
+], i = "custom", a = {
+	imagenet: [e, t],
+	ultralytics: [n, r],
+	none: [n, r]
+};
+function o(e) {
+	return (e.author ?? "").trim().toLowerCase() === "ultralytics" && (e.task ?? "").trim().toLowerCase() === "classify";
+}
+function s(e, t = {}) {
+	let s = t.normalization ?? "auto", l = t.mean !== void 0 || t.std !== void 0;
+	if (s !== "auto" && a[s] === void 0) throw RangeError(`normalization must be one of 'auto', 'imagenet', 'ultralytics', 'none'; got ${JSON.stringify(s)}.`);
+	if (l && s !== "auto") throw RangeError(`Pass either normalization=${JSON.stringify(s)} or explicit mean/std, not both.`);
+	let u = o(e), d = s === "auto" ? u ? "ultralytics" : "imagenet" : s, [f, p] = a[d];
+	if (!l) return {
+		name: d,
+		mean: f,
+		std: p
+	};
+	let m = t.mean ?? f, h = t.std ?? p;
+	return u && !(c(m, n) && c(h, r)) && console.warn(`The classifier is an Ultralytics export, whose classification head is trained on raw [0, 1] images, but mean=${JSON.stringify(m)} / std=${JSON.stringify(h)} was requested. It will be fed images normalized in a way it never saw in training, which degrades accuracy without raising anything. Drop mean/std to let normalization='auto' pick the identity.`), {
+		name: i,
+		mean: m,
+		std: h
+	};
+}
+function c(e, t) {
+	return e.every((e, n) => e === t[n]);
+}
+//#endregion
+export { i as CUSTOM_NORMALIZATION, n as IDENTITY_MEAN, r as IDENTITY_STD, e as IMAGENET_MEAN, t as IMAGENET_STD, o as isUltralyticsClassifier, s as resolveNormalization };
+
+//# sourceMappingURL=normalization.js.map

package/dist/vision/normalization.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"normalization.js","names":[],"sources":["../../src/vision/normalization.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Which preprocessing a classifier expects its input to have had.\n *\n * A classifier is trained on a specific tensor, and feeding it a differently\n * prepared one degrades it silently — no exception, no warning, just worse\n * predictions. The two families this SDK sees most disagree completely:\n * torchvision-style models want the ImageNet mean and deviation subtracted and\n * divided out, while an Ultralytics classification head consumes raw `[0, 1]`.\n *\n * Guessing wrong is not detectable from the outside, but it is not a guess: an\n * Ultralytics export stamps `author` and `task` into its own metadata, and every\n * task in this SDK already reads that map for the class names.\n *\n * Mirrors `ort_vision_sdk.normalization` in the Python SDK; the two must agree,\n * because the same model file is driven by both.\n */\n\n/**\n * Which preprocessing the classifier expects its input to have had.\n *\n * - `\"auto\"` (the default wherever it is accepted) reads the model's own export\n *   metadata and picks. An Ultralytics classification head gets\n *   `\"ultralytics\"`; anything else gets `\"imagenet\"`.\n * - `\"imagenet\"` subtracts the ImageNet mean and divides by the ImageNet\n *   deviation — the torchvision convention.\n * - `\"ultralytics\"` leaves the image in `[0, 1]`. Ultralytics' own classifier\n *   applies no mean/std at all, so anything else feeds it images it never saw\n *   in training.\n * - `\"none\"` is the same arithmetic as `\"ultralytics\"` — identity — under a\n *   name that says \"this model wants raw `[0, 1]`\" rather than naming a vendor.\n */\nexport type Normalization = \"auto\" | \"imagenet\" | \"ultralytics\" | \"none\";\n\n/** Per-channel RGB mean of ImageNet, the torchvision preprocessing convention. */\nexport const IMAGENET_MEAN: readonly [number, number, number] = [0.485, 0.456, 0.406];\n\n/** Per-channel RGB standard deviation of ImageNet. */\nexport const IMAGENET_STD: readonly [number, number, number] = [0.229, 0.224, 0.225];\n\n/** Mean that leaves an image untouched — what an Ultralytics classifier expects. */\nexport const IDENTITY_MEAN: readonly [number, number, number] = [0, 0, 0];\n\n/** Deviation that leaves an image untouched — what an Ultralytics classifier expects. */\nexport const IDENTITY_STD: readonly [number, number, number] = [1, 1, 1];\n\n/** Name reported when the caller supplied `mean`/`std` directly. */\nexport const CUSTOM_NORMALIZATION = \"custom\";\n\nconst PRESETS: Readonly<\n    Record<string, readonly [readonly [number, number, number], readonly [number, number, number]]>\n> = {\n    imagenet: [IMAGENET_MEAN, IMAGENET_STD],\n    ultralytics: [IDENTITY_MEAN, IDENTITY_STD],\n    none: [IDENTITY_MEAN, IDENTITY_STD],\n};\n\n/** What {@link resolveNormalization} settled on. */\nexport interface ResolvedNormalization {\n    /** The preset name, or `\"custom\"` when the caller supplied the values. */\n    readonly name: string;\n    /** Per-channel mean to subtract. */\n    readonly mean: readonly [number, number, number];\n    /** Per-channel deviation to divide by. */\n    readonly std: readonly [number, number, number];\n}\n\n/**\n * Whether a metadata map came out of `YOLO(...).export(format=\"onnx\")`.\n *\n * Every Ultralytics export stamps `author` and `task` into its metadata, and the\n * pair is unambiguous: `\"Ultralytics\"` plus `\"classify\"` is a classification head\n * from that codebase and nothing else.\n *\n * @param metadata The model's custom metadata map.\n */\nexport function isUltralyticsClassifier(metadata: Readonly<Record<string, string>>): boolean {\n    return (\n        (metadata.author ?? \"\").trim().toLowerCase() === \"ultralytics\" &&\n        (metadata.task ?? \"\").trim().toLowerCase() === \"classify\"\n    );\n}\n\n/**\n * Settle which `mean`/`std` to apply, and what to call the choice.\n *\n * Explicit `mean`/`std` always win — they are the escape hatch for a model whose\n * preprocessing neither preset describes. Anything they leave open falls back to\n * the preset, so passing only a `mean` does not silently reset the deviation\n * to 1.\n *\n * Warns (via `console.warn`) when the model is an Ultralytics export and the\n * supplied values are not the identity it was trained with. Nothing fails in\n * that case: the prediction has the right shape and is simply less accurate,\n * which is exactly why it is worth saying out loud.\n *\n * @param metadata The model's custom metadata map, read to detect the family.\n * @param options The preset asked for, plus any explicit `mean`/`std`.\n * @throws {RangeError} If `normalization` is not a known preset, or names one\n *   while `mean`/`std` are also supplied — two answers to the same question.\n */\nexport function resolveNormalization(\n    metadata: Readonly<Record<string, string>>,\n    options: {\n        readonly normalization?: Normalization;\n        readonly mean?: readonly [number, number, number];\n        readonly std?: readonly [number, number, number];\n    } = {},\n): ResolvedNormalization {\n    const normalization = options.normalization ?? \"auto\";\n    const explicit = options.mean !== undefined || options.std !== undefined;\n\n    if (normalization !== \"auto\" && PRESETS[normalization] === undefined) {\n        throw new RangeError(\n            `normalization must be one of 'auto', 'imagenet', 'ultralytics', 'none'; ` +\n                `got ${JSON.stringify(normalization)}.`,\n        );\n    }\n    if (explicit && normalization !== \"auto\") {\n        throw new RangeError(\n            `Pass either normalization=${JSON.stringify(normalization)} or explicit mean/std, not both.`,\n        );\n    }\n\n    const ultralytics = isUltralyticsClassifier(metadata);\n    const preset =\n        normalization === \"auto\" ? (ultralytics ? \"ultralytics\" : \"imagenet\") : normalization;\n    const [presetMean, presetStd] = PRESETS[preset] as readonly [\n        readonly [number, number, number],\n        readonly [number, number, number],\n    ];\n\n    if (!explicit) {\n        return { name: preset, mean: presetMean, std: presetStd };\n    }\n\n    const mean = options.mean ?? presetMean;\n    const std = options.std ?? presetStd;\n    if (ultralytics && !(isIdentity(mean, IDENTITY_MEAN) && isIdentity(std, IDENTITY_STD))) {\n        console.warn(\n            \"The classifier is an Ultralytics export, whose classification head is trained on raw \" +\n                `[0, 1] images, but mean=${JSON.stringify(mean)} / std=${JSON.stringify(std)} was ` +\n                \"requested. It will be fed images normalized in a way it never saw in training, which \" +\n                \"degrades accuracy without raising anything. Drop mean/std to let normalization='auto' \" +\n                \"pick the identity.\",\n        );\n    }\n    return { name: CUSTOM_NORMALIZATION, mean, std };\n}\n\n/**\n * Whether two channel triples are equal.\n *\n * @param value The triple to check.\n * @param reference The triple to compare against.\n */\nfunction isIdentity(\n    value: readonly [number, number, number],\n    reference: readonly [number, number, number],\n): boolean {\n    return value.every((entry, index) => entry === reference[index]);\n}\n"],"mappings":";AAmCA,IAAa,IAAmD;CAAC;CAAO;CAAO;AAAK,GAGvE,IAAkD;CAAC;CAAO;CAAO;AAAK,GAGtE,IAAmD;CAAC;CAAG;CAAG;AAAC,GAG3D,IAAkD;CAAC;CAAG;CAAG;AAAC,GAG1D,IAAuB,UAE9B,IAEF;CACA,UAAU,CAAC,GAAe,CAAY;CACtC,aAAa,CAAC,GAAe,CAAY;CACzC,MAAM,CAAC,GAAe,CAAY;AACtC;AAqBA,SAAgB,EAAwB,GAAqD;CACzF,QACK,EAAS,UAAU,GAAA,CAAI,KAAK,CAAC,CAAC,YAAY,MAAM,kBAChD,EAAS,QAAQ,GAAA,CAAI,KAAK,CAAC,CAAC,YAAY,MAAM;AAEvD;AAoBA,SAAgB,EACZ,GACA,IAII,CAAC,GACgB;CACrB,IAAM,IAAgB,EAAQ,iBAAiB,QACzC,IAAW,EAAQ,SAAS,KAAA,KAAa,EAAQ,QAAQ,KAAA;CAE/D,IAAI,MAAkB,UAAU,EAAQ,OAAmB,KAAA,GACvD,MAAU,WACN,+EACW,KAAK,UAAU,CAAa,EAAE,EAC7C;CAEJ,IAAI,KAAY,MAAkB,QAC9B,MAAU,WACN,6BAA6B,KAAK,UAAU,CAAa,EAAE,iCAC/D;CAGJ,IAAM,IAAc,EAAwB,CAAQ,GAC9C,IACF,MAAkB,SAAU,IAAc,gBAAgB,aAAc,GACtE,CAAC,GAAY,KAAa,EAAQ;CAKxC,IAAI,CAAC,GACD,OAAO;EAAE,MAAM;EAAQ,MAAM;EAAY,KAAK;CAAU;CAG5D,IAAM,IAAO,EAAQ,QAAQ,GACvB,IAAM,EAAQ,OAAO;CAU3B,OATI,KAAe,EAAE,EAAW,GAAM,CAAa,KAAK,EAAW,GAAK,CAAY,MAChF,QAAQ,KACJ,gHAC+B,KAAK,UAAU,CAAI,EAAE,SAAS,KAAK,UAAU,CAAG,EAAE,mMAIrF,GAEG;EAAE,MAAM;EAAsB;EAAM;CAAI;AACnD;AAQA,SAAS,EACL,GACA,GACO;CACP,OAAO,EAAM,OAAO,GAAO,MAAU,MAAU,EAAU,EAAM;AACnE"}

package/dist/vision/results.cjs

@@ -1,2 +1,2 @@
-var e={load:0,preprocess:0,inference:0,postprocess:0},t=class{xyxy;cls;conf;origShape;constructor(e,t,n,r){this.xyxy=e,this.cls=t,this.conf=n,this.origShape=r}get length(){return this.cls.length}get shape(){return[this.length,4]}get xywh(){let e=new Float32Array(this.xyxy.length);for(let t=0;t<this.length;t++){let n=this.xyxy[t*4],r=this.xyxy[t*4+1],i=this.xyxy[t*4+2],a=this.xyxy[t*4+3];e[t*4]=(n+i)/2,e[t*4+1]=(r+a)/2,e[t*4+2]=i-n,e[t*4+3]=a-r}return e}get xyxyn(){let[e,t]=this.origShape,n=new Float32Array(this.xyxy.length);if(this.length===0||t<=0||e<=0)return n;for(let r=0;r<this.length;r++)n[r*4]=this.xyxy[r*4]/t,n[r*4+1]=this.xyxy[r*4+1]/e,n[r*4+2]=this.xyxy[r*4+2]/t,n[r*4+3]=this.xyxy[r*4+3]/e;return n}get xywhn(){let e=this.xywh,[t,n]=this.origShape;if(this.length===0||n<=0||t<=0)return e;for(let r=0;r<this.length;r++)e[r*4]=e[r*4]/n,e[r*4+1]=e[r*4+1]/t,e[r*4+2]=e[r*4+2]/n,e[r*4+3]=e[r*4+3]/t;return e}get data(){let e=new Float32Array(this.length*6);for(let t=0;t<this.length;t++)e[t*6]=this.xyxy[t*4],e[t*6+1]=this.xyxy[t*4+1],e[t*6+2]=this.xyxy[t*4+2],e[t*6+3]=this.xyxy[t*4+3],e[t*6+4]=this.conf[t],e[t*6+5]=this.cls[t];return e}},n=class{data;constructor(e){this.data=e}get length(){return this.data.length}get shape(){return[this.length]}get top1(){return this._top(1).indices[0]??0}get top1conf(){return this._top(1).values[0]??0}get top5(){return this._top(5).indices}get top5conf(){return this._top(5).values}_cache=new Map;_top(e){let t=this._cache.get(e);if(t)return t;let n=this._topK(e);return this._cache.set(e,n),n}_topK(e){let t=Math.min(e,this.data.length),n=new Int32Array(t),r=new Float32Array(t);if(t===0)return{indices:n,values:r};r.fill(-1/0);for(let e=0;e<this.data.length;e++){let i=this.data[e];if(i<=r[t-1])continue;let a=t-1;for(;a>0&&r[a-1]<i;)r[a]=r[a-1],n[a]=n[a-1],--a;r[a]=i,n[a]=e}return{indices:n,values:r}}},r=class{data;xyxy;origShape;constructor(e,t,n){this.data=e,this.xyxy=t,this.origShape=n}get length(){return this.data.length}get shape(){return[this.length]}[Symbol.iterator](){return this.data[Symbol.iterator]()}},i=class{boxes;detections;names;origImg;origShape;path;speed;constructor(t,n,r,i,a,o=null,s=e){this.boxes=t,this.detections=n,this.names=r,this.origImg=i,this.origShape=a,this.path=o,this.speed=s}get length(){return this.detections.length}get(e){return this.detections[e]}[Symbol.iterator](){return this.detections[Symbol.iterator]()}},a=class{boxes;detections;names;classifierNames;origImg;origShape;path;speed;constructor(t,n,r,i,a,o,s=null,c=e){this.boxes=t,this.detections=n,this.names=r,this.classifierNames=i,this.origImg=a,this.origShape=o,this.path=s,this.speed=c}get length(){return this.detections.length}get(e){return this.detections[e]}[Symbol.iterator](){return this.detections[Symbol.iterator]()}},o=class{probs;result;names;origImg;origShape;path;speed;constructor(t,n,r,i,a,o=null,s=e){this.probs=t,this.result=n,this.names=r,this.origImg=i,this.origShape=a,this.path=o,this.speed=s}get cls(){return this.probs.top1}get conf(){return this.probs.top1conf}get name(){return this.names[this.cls]??`class_${this.cls}`}get probabilities(){return this.result.probabilities}},s=class{boxes;masks;detections;names;origImg;origShape;path;speed;constructor(t,n,r,i,a,o,s=null,c=e){this.boxes=t,this.masks=n,this.detections=r,this.names=i,this.origImg=a,this.origShape=o,this.path=s,this.speed=c}get length(){return this.detections.length}get(e){return this.detections[e]}[Symbol.iterator](){return this.detections[Symbol.iterator]()}};exports.Boxes=t,exports.ClassificationResults=o,exports.DetectClassifyResults=a,exports.DetectionResults=i,exports.Masks=r,exports.Probs=n,exports.SegmentationResults=s;
+var e={load:0,preprocess:0,inference:0,postprocess:0},t=class{xyxy;cls;conf;origShape;constructor(e,t,n,r){this.xyxy=e,this.cls=t,this.conf=n,this.origShape=r}get length(){return this.cls.length}get shape(){return[this.length,4]}get xywh(){let e=new Float32Array(this.xyxy.length);for(let t=0;t<this.length;t++){let n=this.xyxy[t*4],r=this.xyxy[t*4+1],i=this.xyxy[t*4+2],a=this.xyxy[t*4+3];e[t*4]=(n+i)/2,e[t*4+1]=(r+a)/2,e[t*4+2]=i-n,e[t*4+3]=a-r}return e}get xyxyn(){let[e,t]=this.origShape,n=new Float32Array(this.xyxy.length);if(this.length===0||t<=0||e<=0)return n;for(let r=0;r<this.length;r++)n[r*4]=this.xyxy[r*4]/t,n[r*4+1]=this.xyxy[r*4+1]/e,n[r*4+2]=this.xyxy[r*4+2]/t,n[r*4+3]=this.xyxy[r*4+3]/e;return n}get xywhn(){let e=this.xywh,[t,n]=this.origShape;if(this.length===0||n<=0||t<=0)return e;for(let r=0;r<this.length;r++)e[r*4]=e[r*4]/n,e[r*4+1]=e[r*4+1]/t,e[r*4+2]=e[r*4+2]/n,e[r*4+3]=e[r*4+3]/t;return e}get data(){let e=new Float32Array(this.length*6);for(let t=0;t<this.length;t++)e[t*6]=this.xyxy[t*4],e[t*6+1]=this.xyxy[t*4+1],e[t*6+2]=this.xyxy[t*4+2],e[t*6+3]=this.xyxy[t*4+3],e[t*6+4]=this.conf[t],e[t*6+5]=this.cls[t];return e}},n=class{data;constructor(e){this.data=e}get length(){return this.data.length}get shape(){return[this.length]}get top1(){if(this.data.length===0)return 0;let e=0,t=this.data[0];for(let n=1;n<this.data.length;n++){let r=this.data[n];r>t&&(e=n,t=r)}return e}get top1conf(){return this.data.length===0?0:this.data[this.top1]}get top5(){return this._top(5).indices}get top5conf(){return this._top(5).values}_cache=new Map;_top(e){let t=this._cache.get(e);if(t)return t;let n=this._topK(e);return this._cache.set(e,n),n}_topK(e){let t=Math.min(e,this.data.length),n=new Int32Array(t),r=new Float32Array(t);if(t===0)return{indices:n,values:r};r.fill(-1/0);for(let e=0;e<this.data.length;e++){let i=this.data[e];if(i<=r[t-1])continue;let a=t-1;for(;a>0&&r[a-1]<i;)r[a]=r[a-1],n[a]=n[a-1],--a;r[a]=i,n[a]=e}return{indices:n,values:r}}},r=class{data;xyxy;origShape;constructor(e,t,n){this.data=e,this.xyxy=t,this.origShape=n}get length(){return this.data.length}get shape(){return[this.length]}[Symbol.iterator](){return this.data[Symbol.iterator]()}},i=class{boxes;detections;names;origImg;origShape;path;speed;constructor(t,n,r,i,a,o=null,s=e){this.boxes=t,this.detections=n,this.names=r,this.origImg=i,this.origShape=a,this.path=o,this.speed=s}get length(){return this.detections.length}get(e){return this.detections[e]}[Symbol.iterator](){return this.detections[Symbol.iterator]()}},a=class{boxes;detections;names;classifierNames;origImg;origShape;path;speed;constructor(t,n,r,i,a,o,s=null,c=e){this.boxes=t,this.detections=n,this.names=r,this.classifierNames=i,this.origImg=a,this.origShape=o,this.path=s,this.speed=c}get length(){return this.detections.length}get(e){return this.detections[e]}[Symbol.iterator](){return this.detections[Symbol.iterator]()}},o=class{probs;result;names;origImg;origShape;path;speed;constructor(t,n,r,i,a,o=null,s=e){this.probs=t,this.result=n,this.names=r,this.origImg=i,this.origShape=a,this.path=o,this.speed=s}get cls(){return this.probs.top1}get conf(){return this.probs.top1conf}get name(){return this.names[this.cls]??`class_${this.cls}`}get probabilities(){return this.result.probabilities}},s=class{boxes;masks;detections;names;origImg;origShape;path;speed;constructor(t,n,r,i,a,o,s=null,c=e){this.boxes=t,this.masks=n,this.detections=r,this.names=i,this.origImg=a,this.origShape=o,this.path=s,this.speed=c}get length(){return this.detections.length}get(e){return this.detections[e]}[Symbol.iterator](){return this.detections[Symbol.iterator]()}};exports.Boxes=t,exports.ClassificationResults=o,exports.DetectClassifyResults=a,exports.DetectionResults=i,exports.Masks=r,exports.Probs=n,exports.SegmentationResults=s;
 //# sourceMappingURL=results.cjs.map