tempest-react-sdk 0.69.0 → 0.70.0

package/dist/vision/tasks/base.js

@@ -1,6 +1,7 @@
 import { NoDetectionsError as e } from "../core/exceptions.js";
+import { RGBImage as t } from "../types.js";
 //#region src/vision/tasks/base.ts
-var t = class {
+var n = class {
 	_session;
 	constructor(e) {
 		this._session = e;
@@ -9,15 +10,29 @@ var t = class {
 		return this._session;
 	}
 };
-function n(e) {
+function r(e) {
 	return e.toFixed(6).replace(/0+$/, "").replace(/\.$/, "");
 }
-function r(t, r) {
-	if (!r.raiseOnEmpty || t > 0) return;
-	let i = r.path ? ` in ${r.path}` : "", a = r.classes === void 0 ? "" : ` among classes [${[...r.classes].sort((e, t) => e - t).join(", ")}]`;
-	throw new e(`No detections${i}${a}: nothing cleared confThreshold=${n(r.confThreshold)}.`);
+function i(t, n) {
+	if (!n.raiseOnEmpty || t > 0) return;
+	let i = n.path ? ` in ${n.path}` : "", a = n.classes === void 0 ? "" : ` among classes [${[...n.classes].sort((e, t) => e - t).join(", ")}]`;
+	throw new e(`No detections${i}${a}: nothing cleared confThreshold=${r(n.confThreshold)}.`);
+}
+function a(e, n) {
+	let [r, i, a, o] = n.asIntXyxy(), s = Math.max(0, r), c = Math.max(0, i), l = Math.min(e.width, a), u = Math.min(e.height, o);
+	if (l <= s || u <= c) return new t(/* @__PURE__ */ new Uint8Array(), 0, 0);
+	let d = l - s, f = u - c, p = new Uint8Array(d * f * 3);
+	for (let t = 0; t < f; t++) {
+		let n = ((c + t) * e.width + s) * 3;
+		p.set(e.data.subarray(n, n + d * 3), t * d * 3);
+	}
+	return new t(p, d, f);
+}
+function o(e) {
+	let t = !1, n;
+	return () => (t ||= (n = e(), !0), n);
 }
 //#endregion
-export { t as VisionTask, r as requireDetections };
+export { n as VisionTask, a as cropToBox, o as memoize, i as requireDetections };
 
 //# sourceMappingURL=base.js.map

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

@@ -1 +1 @@
-{"version":3,"file":"base.js","names":[],"sources":["../../../src/vision/tasks/base.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Common foundation for task-oriented vision SDK objects.\n *\n * The base only owns the {@link OrtSession} — label resolution lives in each\n * subclass because how `numClasses` is read from the model differs per task.\n */\n\nimport { NoDetectionsError } from \"../core/exceptions\";\nimport type { OrtSession } from \"../core/session\";\n\nexport abstract class VisionTask {\n    protected constructor(protected readonly _session: OrtSession) {}\n\n    /** The underlying {@link OrtSession} used to run inference. */\n    get session(): OrtSession {\n        return this._session;\n    }\n}\n\n/**\n * Render a confidence threshold the way the Python SDK renders it.\n *\n * JavaScript and Python disagree on how a number becomes text. A whole\n * threshold is `1` here and `1.0` there; JavaScript holds off on exponent\n * notation until `1e-7` while Python switches at `1e-5`. A fused pipeline is\n * built once and runs under both runtimes from the same file, so a message that\n * quotes the threshold has to quote it identically — otherwise the two SDKs\n * describe the same run with two different numbers.\n *\n * Six decimals with the trailing zeros trimmed covers every threshold a caller\n * can meaningfully set and agrees byte for byte with `_format_threshold` in the\n * Python SDK's `tasks/base.py`. The pairing is fixed by a shared table in both\n * test suites, so a change on one side that is not mirrored on the other fails.\n *\n * @param value The threshold to render.\n * @returns The threshold as text, without trailing zeros or a trailing dot.\n */\nfunction formatThreshold(value: number): string {\n    return value.toFixed(6).replace(/0+$/, \"\").replace(/\\.$/, \"\");\n}\n\n/**\n * Turn an empty result into an error, when the caller asked for that.\n *\n * Shared by every task that can come back with nothing — {@link Detector},\n * {@link Segmenter} and {@link DetectClassify} — so the three agree on when\n * they throw and on what the message says. The message names the two settings\n * that decide the outcome, because \"no detections\" on its own leaves the reader\n * unable to tell a blank image from a threshold set too high.\n *\n * @param count How many detections survived every filter.\n * @param options The flag for this call, the threshold actually applied (after\n *   any per-call override), the class allowlist if one narrowed the search, and\n *   the source path when the input was one.\n * @throws {@link NoDetectionsError} when the flag is set and `count` is zero.\n */\nexport function requireDetections(\n    count: number,\n    options: {\n        readonly raiseOnEmpty: boolean;\n        readonly confThreshold: number;\n        readonly classes: readonly number[] | undefined;\n        readonly path: string | null;\n    },\n): void {\n    if (!options.raiseOnEmpty || count > 0) return;\n    const where = options.path ? ` in ${options.path}` : \"\";\n    const narrowed =\n        options.classes === undefined\n            ? \"\"\n            : ` among classes [${[...options.classes].sort((a, b) => a - b).join(\", \")}]`;\n    throw new NoDetectionsError(\n        `No detections${where}${narrowed}: nothing cleared confThreshold=${formatThreshold(options.confThreshold)}.`,\n    );\n}\n"],"mappings":";;AAWA,IAAsB,IAAtB,MAAiC;CACY;CAAzC,YAAsB,GAAyC;EAAtB,KAAA,WAAA;CAAuB;CAGhE,IAAI,UAAsB;EACtB,OAAO,KAAK;CAChB;AACJ;AAoBA,SAAS,EAAgB,GAAuB;CAC5C,OAAO,EAAM,QAAQ,CAAC,CAAC,CAAC,QAAQ,OAAO,EAAE,CAAC,CAAC,QAAQ,OAAO,EAAE;AAChE;AAiBA,SAAgB,EACZ,GACA,GAMI;CACJ,IAAI,CAAC,EAAQ,gBAAgB,IAAQ,GAAG;CACxC,IAAM,IAAQ,EAAQ,OAAO,OAAO,EAAQ,SAAS,IAC/C,IACF,EAAQ,YAAY,KAAA,IACd,KACA,mBAAmB,CAAC,GAAG,EAAQ,OAAO,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC,CAAC,CAAC,KAAK,IAAI,EAAE;CACnF,MAAM,IAAI,EACN,gBAAgB,IAAQ,EAAS,kCAAkC,EAAgB,EAAQ,aAAa,EAAE,EAC9G;AACJ"}
+{"version":3,"file":"base.js","names":[],"sources":["../../../src/vision/tasks/base.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Common foundation for task-oriented vision SDK objects.\n *\n * The base only owns the {@link OrtSession} — label resolution lives in each\n * subclass because how `numClasses` is read from the model differs per task.\n */\n\nimport { NoDetectionsError } from \"../core/exceptions\";\nimport type { OrtSession } from \"../core/session\";\nimport { type BoundingBox, RGBImage } from \"../types\";\n\nexport abstract class VisionTask {\n    protected constructor(protected readonly _session: OrtSession) {}\n\n    /** The underlying {@link OrtSession} used to run inference. */\n    get session(): OrtSession {\n        return this._session;\n    }\n}\n\n/**\n * Render a confidence threshold the way the Python SDK renders it.\n *\n * JavaScript and Python disagree on how a number becomes text. A whole\n * threshold is `1` here and `1.0` there; JavaScript holds off on exponent\n * notation until `1e-7` while Python switches at `1e-5`. A fused pipeline is\n * built once and runs under both runtimes from the same file, so a message that\n * quotes the threshold has to quote it identically — otherwise the two SDKs\n * describe the same run with two different numbers.\n *\n * Six decimals with the trailing zeros trimmed covers every threshold a caller\n * can meaningfully set and agrees byte for byte with `_format_threshold` in the\n * Python SDK's `tasks/base.py`. The pairing is fixed by a shared table in both\n * test suites, so a change on one side that is not mirrored on the other fails.\n *\n * @param value The threshold to render.\n * @returns The threshold as text, without trailing zeros or a trailing dot.\n */\nfunction formatThreshold(value: number): string {\n    return value.toFixed(6).replace(/0+$/, \"\").replace(/\\.$/, \"\");\n}\n\n/**\n * Turn an empty result into an error, when the caller asked for that.\n *\n * Shared by every task that can come back with nothing — {@link Detector},\n * {@link Segmenter} and {@link DetectClassify} — so the three agree on when\n * they throw and on what the message says. The message names the two settings\n * that decide the outcome, because \"no detections\" on its own leaves the reader\n * unable to tell a blank image from a threshold set too high.\n *\n * @param count How many detections survived every filter.\n * @param options The flag for this call, the threshold actually applied (after\n *   any per-call override), the class allowlist if one narrowed the search, and\n *   the source path when the input was one.\n * @throws {@link NoDetectionsError} when the flag is set and `count` is zero.\n */\nexport function requireDetections(\n    count: number,\n    options: {\n        readonly raiseOnEmpty: boolean;\n        readonly confThreshold: number;\n        readonly classes: readonly number[] | undefined;\n        readonly path: string | null;\n    },\n): void {\n    if (!options.raiseOnEmpty || count > 0) return;\n    const where = options.path ? ` in ${options.path}` : \"\";\n    const narrowed =\n        options.classes === undefined\n            ? \"\"\n            : ` among classes [${[...options.classes].sort((a, b) => a - b).join(\", \")}]`;\n    throw new NoDetectionsError(\n        `No detections${where}${narrowed}: nothing cleared confThreshold=${formatThreshold(options.confThreshold)}.`,\n    );\n}\n\n/**\n * Cut the box region out of an image.\n *\n * The box is truncated to whole pixels and clamped to the image, the same\n * rounding every task applies, so a crop always matches `bbox.asIntXyxy()`\n * within the frame.\n *\n * @param image The source image.\n * @param bbox The box, in the image's pixel coordinates.\n * @returns The cropped region, or a zero-sized image for a box with no area.\n */\nexport function cropToBox(image: RGBImage, bbox: BoundingBox): RGBImage {\n    const [rawX1, rawY1, rawX2, rawY2] = bbox.asIntXyxy();\n    const x1 = Math.max(0, rawX1);\n    const y1 = Math.max(0, rawY1);\n    const x2 = Math.min(image.width, rawX2);\n    const y2 = Math.min(image.height, rawY2);\n    if (x2 <= x1 || y2 <= y1) return new RGBImage(new Uint8Array(0), 0, 0);\n\n    const width = x2 - x1;\n    const height = y2 - y1;\n    const out = new Uint8Array(width * height * 3);\n    for (let row = 0; row < height; row++) {\n        const offset = ((y1 + row) * image.width + x1) * 3;\n        out.set(image.data.subarray(offset, offset + width * 3), row * width * 3);\n    }\n    return new RGBImage(out, width, height);\n}\n\n/**\n * Wrap a computation so it runs on first call and is remembered after.\n *\n * Backs the per-detection images (`croppedImage`, `segmentedImage`): most\n * callers read a box and a class and never look at the pixels, yet building\n * them eagerly copied every box region out of the frame on every\n * `predict()` — up to `maxDetections` allocations nobody asked for.\n *\n * @param compute Produces the value; called at most once.\n * @returns A getter returning the value, computing it on the first call.\n */\nexport function memoize<T>(compute: () => T): () => T {\n    let done = false;\n    let value: T;\n    return () => {\n        if (!done) {\n            value = compute();\n            done = true;\n        }\n        return value;\n    };\n}\n"],"mappings":";;;AAYA,IAAsB,IAAtB,MAAiC;CACY;CAAzC,YAAsB,GAAyC;EAAtB,KAAA,WAAA;CAAuB;CAGhE,IAAI,UAAsB;EACtB,OAAO,KAAK;CAChB;AACJ;AAoBA,SAAS,EAAgB,GAAuB;CAC5C,OAAO,EAAM,QAAQ,CAAC,CAAC,CAAC,QAAQ,OAAO,EAAE,CAAC,CAAC,QAAQ,OAAO,EAAE;AAChE;AAiBA,SAAgB,EACZ,GACA,GAMI;CACJ,IAAI,CAAC,EAAQ,gBAAgB,IAAQ,GAAG;CACxC,IAAM,IAAQ,EAAQ,OAAO,OAAO,EAAQ,SAAS,IAC/C,IACF,EAAQ,YAAY,KAAA,IACd,KACA,mBAAmB,CAAC,GAAG,EAAQ,OAAO,CAAC,CAAC,MAAM,GAAG,MAAM,IAAI,CAAC,CAAC,CAAC,KAAK,IAAI,EAAE;CACnF,MAAM,IAAI,EACN,gBAAgB,IAAQ,EAAS,kCAAkC,EAAgB,EAAQ,aAAa,EAAE,EAC9G;AACJ;AAaA,SAAgB,EAAU,GAAiB,GAA6B;CACpE,IAAM,CAAC,GAAO,GAAO,GAAO,KAAS,EAAK,UAAU,GAC9C,IAAK,KAAK,IAAI,GAAG,CAAK,GACtB,IAAK,KAAK,IAAI,GAAG,CAAK,GACtB,IAAK,KAAK,IAAI,EAAM,OAAO,CAAK,GAChC,IAAK,KAAK,IAAI,EAAM,QAAQ,CAAK;CACvC,IAAI,KAAM,KAAM,KAAM,GAAI,OAAO,IAAI,kBAAS,IAAI,WAAY,GAAG,GAAG,CAAC;CAErE,IAAM,IAAQ,IAAK,GACb,IAAS,IAAK,GACd,IAAM,IAAI,WAAW,IAAQ,IAAS,CAAC;CAC7C,KAAK,IAAI,IAAM,GAAG,IAAM,GAAQ,KAAO;EACnC,IAAM,MAAW,IAAK,KAAO,EAAM,QAAQ,KAAM;EACjD,EAAI,IAAI,EAAM,KAAK,SAAS,GAAQ,IAAS,IAAQ,CAAC,GAAG,IAAM,IAAQ,CAAC;CAC5E;CACA,OAAO,IAAI,EAAS,GAAK,GAAO,CAAM;AAC1C;AAaA,SAAgB,EAAW,GAA2B;CAClD,IAAI,IAAO,IACP;CACJ,cACI,AAEI,OADA,IAAQ,EAAQ,GACT,KAEJ;AAEf"}

package/dist/vision/tasks/classifier.cjs

@@ -1,2 +1,2 @@
-const e=require("../core/timing.cjs"),t=require("../results.cjs"),n=require("../labels.cjs"),r=require("../core/graph.cjs"),i=require("../core/metadata.cjs"),a=require("../core/session.cjs"),o=require("../normalization.cjs"),s=require("../io/image.cjs"),c=require("../preprocess/image.cjs"),l=require("../preprocess/pipeline.cjs"),u=require("../postprocess/classification.cjs"),d=require("./base.cjs");var f=class f extends d.VisionTask{_labels;_names;_inputSize;_mean;_std;_applySoftmax;_normalization;constructor(e,t,n,r,i,a,o,s){super(e),this._labels=t,this._names=n,this._inputSize=r,this._mean=i,this._std=a,this._applySoftmax=o,this._normalization=s}get appliesSoftmax(){return this._applySoftmax}get normalization(){return this._normalization}_pipelineCache=null;async warmup(e=1){let[t,n]=this._inputSize;for(let r=0;r<e;r++){let e=c.toFloat32Tensor(l.zeroTensorData(t,n),[1,3,n,t]);await this._session.run({[this._session.inputName]:e})}}get _pipeline(){if(this._pipelineCache===null){let[e,t]=this._inputSize;this._pipelineCache=new l.ResizePipeline(e,t,this._mean,this._std)}return this._pipelineCache}static async create(e,t={}){let s=await a.OrtSession.create(e,t),c=o.resolveNormalization(s.metadata,{normalization:t.normalization,mean:t.mean,std:t.std}),l=t.numClasses??r.classificationNumClasses(s.outputShape)??void 0,u=n.resolveLabels(t.labels??i.modelNames(s.metadata),{numClasses:l}),d={};for(let e=0;e<u.length;e++)d[e]=u[e];return new f(s,u,d,r.resolveInputSize({graphShape:s.inputShape,requested:t.inputSize,fallback:[224,224]}),c.mean,c.std,t.applySoftmax??!o.isUltralyticsClassifier(s.metadata),c.name)}get labels(){return this._labels}get names(){return this._names}get inputSize(){return this._inputSize}get numClasses(){return this._labels.length}async call(e,t={}){return this.predict(e,t)}async predict(n,r={}){let i=new e.SpeedTimer,a=typeof n==`string`?n:null,o=await s.loadImage(n);i.stage(`load`);let c=this._preprocess(o);i.stage(`preprocess`);let l=await this._session.run({[this._session.inputName]:c});this._pipeline.release(),i.stage(`inference`);let d=this._session.outputNames[0];if(d===void 0)throw Error(`Classifier model has no outputs.`);let f=l[d];if(f===void 0)throw Error(`Classifier model output ${d} missing from run() result.`);let p=this._postprocess(f.data),{indices:m,values:h}=u.topK(p,r.topK??null),g=[];for(let e=0;e<m.length;e++){let t=m[e],n=this._labels[t]??`class_${t}`;g.push({classId:t,className:n,probability:h[e],cls:t,name:n,conf:h[e]})}if(g.length===0)throw Error(`Classifier produced no probabilities (empty output).`);let _=g[0],v={classId:_.classId,className:_.className,confidence:_.probability,cls:_.classId,name:_.className,conf:_.probability,image:o,probabilities:g},y=[o.height,o.width],b=new t.Probs(p);return i.stage(`postprocess`),[new t.ClassificationResults(b,v,this._names,o,y,a,i.speed())]}_preprocess(e){let[t,n]=this._inputSize,{data:r}=this._pipeline.run(e);return c.toFloat32Tensor(r,[1,3,n,t])}_postprocess(e){return this._applySoftmax?u.softmax(e):new Float32Array(e)}};exports.Classifier=f;
+const e=require("../core/timing.cjs"),t=require("../results.cjs"),n=require("../labels.cjs"),r=require("../core/dtypes.cjs"),i=require("../core/graph.cjs"),a=require("../core/metadata.cjs"),o=require("../core/session.cjs"),s=require("../normalization.cjs"),c=require("../io/image.cjs"),l=require("../preprocess/image.cjs"),u=require("../preprocess/pipeline.cjs"),d=require("../postprocess/classification.cjs"),f=require("./base.cjs");var p=class p extends f.VisionTask{_labels;_names;_inputSize;_mean;_std;_applySoftmax;_normalization;constructor(e,t,n,r,i,a,o,s){super(e),this._labels=t,this._names=n,this._inputSize=r,this._mean=i,this._std=a,this._applySoftmax=o,this._normalization=s}get appliesSoftmax(){return this._applySoftmax}get normalization(){return this._normalization}_pipelineCache=null;async warmup(e=1){let[t,n]=this._inputSize;for(let r=0;r<e;r++){let e=l.toFloat32Tensor(u.zeroTensorData(t,n),[1,3,n,t]);await this._session.run({[this._session.inputName]:e})}}get _pipeline(){if(this._pipelineCache===null){let[e,t]=this._inputSize;this._pipelineCache=new u.ResizePipeline(e,t,this._mean,this._std)}return this._pipelineCache}static async create(e,t={}){let r=await o.OrtSession.create(e,t),c=s.resolveNormalization(r.metadata,{normalization:t.normalization,mean:t.mean,std:t.std}),l=t.numClasses??i.classificationNumClasses(r.outputShape)??void 0,u=n.resolveLabels(t.labels??a.modelNames(r.metadata),{numClasses:l}),d={};for(let e=0;e<u.length;e++)d[e]=u[e];return new p(r,u,d,i.resolveInputSize({graphShape:r.inputShape,requested:t.inputSize,fallback:[224,224]}),c.mean,c.std,t.applySoftmax??!s.isUltralyticsClassifier(r.metadata),c.name)}get labels(){return this._labels}get names(){return this._names}get inputSize(){return this._inputSize}get numClasses(){return this._labels.length}async call(e,t={}){return this.predict(e,t)}async predict(n,i={}){let a=new e.SpeedTimer,o=typeof n==`string`?n:null,{image:s,canvas:l}=await c.loadImageSource(n);a.stage(`load`);let{tensor:u,reused:f}=this._preprocess(s,l);a.stage(`preprocess`);let p;try{p=await this._session.run({[this._session.inputName]:u})}finally{this._pipeline.release({reused:f})}a.stage(`inference`);let m=this._session.outputNames[0];if(m===void 0)throw Error(`Classifier model has no outputs.`);let h=p[m];if(h===void 0)throw Error(`Classifier model output ${m} missing from run() result.`);let g=this._postprocess(r.asFloat32Array(h.data)),{indices:_,values:v}=d.topK(g,i.topK??null),y=[];for(let e=0;e<_.length;e++){let t=_[e],n=this._labels[t]??`class_${t}`;y.push({classId:t,className:n,probability:v[e],cls:t,name:n,conf:v[e]})}if(y.length===0)throw Error(`Classifier produced no probabilities (empty output).`);let b=y[0],x={classId:b.classId,className:b.className,confidence:b.probability,cls:b.classId,name:b.className,conf:b.probability,image:s,probabilities:y},S=[s.height,s.width],C=new t.Probs(g);return a.stage(`postprocess`),[new t.ClassificationResults(C,x,this._names,s,S,o,a.speed())]}_preprocess(e,t){let[n,r]=this._inputSize,{data:i,reused:a}=this._pipeline.run(e,t);return{tensor:l.toFloat32Tensor(i,[1,3,r,n]),reused:a}}_postprocess(e){return this._applySoftmax?d.softmax(e):new Float32Array(e)}};exports.Classifier=p;
 //# sourceMappingURL=classifier.cjs.map

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

@@ -1 +1 @@
-{"version":3,"file":"classifier.cjs","names":[],"sources":["../../../src/vision/tasks/classifier.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Image classification task using ONNX Runtime Web.\n */\n\nimport type * as ort from \"onnxruntime-web\";\n\nimport { type ModelSource, type OrtSessionOptions, OrtSession } from \"../core/session\";\nimport { SpeedTimer } from \"../core/timing\";\nimport { type ImageInput, loadImage } from \"../io/image\";\nimport { classificationNumClasses, resolveInputSize } from \"../core/graph\";\nimport { modelNames } from \"../core/metadata\";\nimport { type LabelSpec, resolveLabels } from \"../labels\";\nimport {\n    type Normalization,\n    isUltralyticsClassifier,\n    resolveNormalization,\n} from \"../normalization\";\nimport { softmax, topK } from \"../postprocess/classification\";\nimport { toFloat32Tensor } from \"../preprocess/image\";\nimport { ResizePipeline, zeroTensorData } from \"../preprocess/pipeline\";\nimport { ClassificationResults, Probs } from \"../results\";\nimport { VisionTask } from \"./base\";\nimport { type ClassProbability, type ClassificationResult, type RGBImage } from \"../types\";\n\nexport interface ClassifierOptions extends OrtSessionOptions {\n    /**\n     * Class label spec — see {@link resolveLabels}.\n     *\n     * Optional: when omitted, the names the export baked into the model are used\n     * (Ultralytics writes them as `names` in the metadata map). Only when the\n     * model carries none does this fall back to generated `class_<id>` labels.\n     * Passing a spec always wins, for a model whose names are wrong or absent.\n     */\n    readonly labels?: LabelSpec;\n    /**\n     * Number of classes the model can predict.\n     *\n     * Optional: inferred from the classification head's declared output shape\n     * `(B, nc)`. Pass it to validate that the supplied labels match the model.\n     */\n    readonly numClasses?: number;\n    /**\n     * Model input `[width, height]` in pixels.\n     *\n     * Only used when the model's graph leaves its spatial axes dynamic: a graph\n     * that declares a static size always wins, since that is the only shape ONNX\n     * Runtime will accept. Defaults to `[224, 224]`.\n     */\n    readonly inputSize?: readonly [number, number];\n    /**\n     * Which preprocessing this model expects — see {@link Normalization}.\n     *\n     * Defaults to `\"auto\"`, which reads the model's own export metadata and picks\n     * `\"ultralytics\"` (raw `[0, 1]`) for an Ultralytics classification head and\n     * `\"imagenet\"` for everything else.\n     */\n    readonly normalization?: Normalization;\n    /** Per-channel RGB mean, overriding the preset. Defaults to the preset's. */\n    readonly mean?: readonly [number, number, number];\n    /** Per-channel RGB standard deviation, overriding the preset. */\n    readonly std?: readonly [number, number, number];\n    /**\n     * Whether the model's output still needs a softmax.\n     *\n     * Left undefined (the default), this reads the model's metadata and answers\n     * `false` for an Ultralytics classification export, whose graph already ends\n     * in one — applying a second softmax to a probability vector keeps the\n     * ranking but flattens the confidences, so the top-1 stays right while every\n     * number attached to it is wrong. Detection covers that family; for any other\n     * model that already emits probabilities, pass `false` explicitly.\n     */\n    readonly applySoftmax?: boolean;\n}\n\nexport interface ClassifierPredictOptions {\n    /**\n     * If set, the per-class probability list in `results[0].result.probabilities`\n     * is truncated to the top-K entries. The bulk `probs` view always exposes\n     * the full vector.\n     */\n    readonly topK?: number;\n}\n\n/**\n * Image classifier wrapping an ONNX model with ImageNet-style preprocessing.\n *\n * `predict()` returns `Promise<ClassificationResults[]>` (length 1 for a\n * single image), mirroring Ultralytics' API. The envelope exposes a `probs`\n * collection (`top1`, `top1conf`, `top5`, `top5conf`, `data`) plus the\n * legacy per-class probability list with names already resolved.\n *\n * Defaults: 224×224 RGB input, `float32` normalized with ImageNet mean/std,\n * NCHW layout, batch size 1, softmax applied to the raw output.\n *\n * @example\n * ```typescript\n * const clf = await Classifier.create(\"/models/resnet50.onnx\", {\n *   labels: [\"tench\", \"goldfish\", ...]  // 1000 ImageNet labels\n * });\n * const r = (await clf.predict(\"/images/dog.jpg\"))[0];\n * console.log(r.cls, r.conf, r.name);\n * console.log(r.probs.top5, r.probs.top5conf);\n * ```\n */\nexport class Classifier extends VisionTask {\n    private constructor(\n        session: OrtSession,\n        private readonly _labels: readonly string[],\n        private readonly _names: Readonly<Record<number, string>>,\n        private readonly _inputSize: readonly [number, number],\n        private readonly _mean: readonly [number, number, number],\n        private readonly _std: readonly [number, number, number],\n        private readonly _applySoftmax: boolean,\n        private readonly _normalization: string,\n    ) {\n        super(session);\n    }\n\n    /**\n     * Whether a softmax is applied to the model's output before ranking.\n     *\n     * Resolved once at construction. Worth reading when confidences look\n     * compressed: a second softmax over an already-normalized vector leaves the\n     * ordering intact and the numbers meaningless, which is invisible to any check\n     * that only looks at the predicted class.\n     */\n    get appliesSoftmax(): boolean {\n        return this._applySoftmax;\n    }\n\n    /**\n     * Which preprocessing this classifier applies to every image.\n     *\n     * One of the {@link Normalization} preset names, or `\"custom\"` when the caller\n     * supplied `mean`/`std` directly. Worth reading when a model underperforms:\n     * feeding a classifier a differently prepared tensor than it was trained on\n     * degrades it without throwing anything, so \"what does this assume\" is the\n     * first question.\n     */\n    get normalization(): string {\n        return this._normalization;\n    }\n\n    private _pipelineCache: ResizePipeline | null = null;\n\n    /**\n     * Run the model once on a zero tensor, paying one-time costs up front.\n     *\n     * The first inference of a session is not representative: WebGPU compiles its\n     * shaders on it and the WASM backend faults in its arenas. Calling this while\n     * a loading spinner is still up moves that cost somewhere the user is already\n     * waiting — which matters most for a classifier running as the last step of\n     * an on-device analysis, where the delay lands right before the answer shows.\n     *\n     * @param runs How many warm-up inferences to run. One is enough for WASM;\n     *   WebGPU sometimes settles on the second.\n     */\n    async warmup(runs: number = 1): Promise<void> {\n        const [tw, th] = this._inputSize;\n        for (let i = 0; i < runs; i++) {\n            const tensor = toFloat32Tensor(zeroTensorData(tw, th), [1, 3, th, tw]);\n            await this._session.run({ [this._session.inputName]: tensor });\n        }\n    }\n\n    /**\n     * The fused preprocessing pipeline, built on first use.\n     *\n     * Lazily, because constructing it reserves the output buffer: a task built in\n     * an environment without a canvas implementation stays constructible, and only\n     * fails if it is actually asked to preprocess something.\n     */\n    private get _pipeline(): ResizePipeline {\n        if (this._pipelineCache === null) {\n            const [tw, th] = this._inputSize;\n            this._pipelineCache = new ResizePipeline(tw, th, this._mean, this._std);\n        }\n        return this._pipelineCache;\n    }\n\n    /**\n     * Load the model, resolve labels, and settle the preprocessing.\n     *\n     * @param model The model source — a URL, or the bytes.\n     * @param options Labels, input size, normalization, and session options.\n     * @throws {RangeError} If `normalization` names an unknown preset, or names\n     *   one while `mean`/`std` are also given.\n     */\n    static async create(model: ModelSource, options: ClassifierOptions = {}): Promise<Classifier> {\n        const session = await OrtSession.create(model, options);\n        const normalization = resolveNormalization(session.metadata, {\n            normalization: options.normalization,\n            mean: options.mean,\n            std: options.std,\n        });\n        const numClasses =\n            options.numClasses ?? classificationNumClasses(session.outputShape) ?? undefined;\n        const labels = resolveLabels(options.labels ?? modelNames(session.metadata), {\n            numClasses,\n        });\n        const names: Record<number, string> = {};\n        for (let i = 0; i < labels.length; i++) {\n            names[i] = labels[i] as string;\n        }\n        return new Classifier(\n            session,\n            labels,\n            names,\n            resolveInputSize({\n                graphShape: session.inputShape,\n                requested: options.inputSize,\n                fallback: [224, 224],\n            }),\n            normalization.mean,\n            normalization.std,\n            options.applySoftmax ?? !isUltralyticsClassifier(session.metadata),\n            normalization.name,\n        );\n    }\n\n    /** Class labels indexed by class id. */\n    get labels(): readonly string[] {\n        return this._labels;\n    }\n\n    /** Class id → class name dict (matches Ultralytics' `model.names`). */\n    get names(): Readonly<Record<number, string>> {\n        return this._names;\n    }\n\n    /**\n     * The `[width, height]` this task preprocesses to.\n     *\n     * Resolved at creation time from the model's graph when it declares a static\n     * input, so reading it back tells you the resolution inference really runs at\n     * — not merely what was requested.\n     */\n    get inputSize(): readonly [number, number] {\n        return this._inputSize;\n    }\n\n    /** Number of classes the model can predict. */\n    get numClasses(): number {\n        return this._labels.length;\n    }\n\n    /** Alias for {@link predict} (parity with PyTorch `nn.Module.__call__`). */\n    async call(\n        image: ImageInput,\n        options: ClassifierPredictOptions = {},\n    ): Promise<ClassificationResults[]> {\n        return this.predict(image, options);\n    }\n\n    /** Run classification on a single image. */\n    async predict(\n        image: ImageInput,\n        options: ClassifierPredictOptions = {},\n    ): Promise<ClassificationResults[]> {\n        const timer = new SpeedTimer();\n        const path = typeof image === \"string\" ? image : null;\n        const original = await loadImage(image);\n        timer.stage(\"load\");\n        const tensor = this._preprocess(original);\n        timer.stage(\"preprocess\");\n        const outputs = await this._session.run({ [this._session.inputName]: tensor });\n        this._pipeline.release();\n        timer.stage(\"inference\");\n        const firstOutputName = this._session.outputNames[0];\n        if (firstOutputName === undefined) {\n            throw new Error(\"Classifier model has no outputs.\");\n        }\n        const raw = outputs[firstOutputName];\n        if (raw === undefined) {\n            throw new Error(\n                `Classifier model output ${firstOutputName} missing from run() result.`,\n            );\n        }\n        const fullProbs = this._postprocess(raw.data as Float32Array);\n\n        const { indices, values } = topK(fullProbs, options.topK ?? null);\n        const probabilities: ClassProbability[] = [];\n        for (let i = 0; i < indices.length; i++) {\n            const id = indices[i] as number;\n            const className = this._labels[id] ?? `class_${id}`;\n            probabilities.push({\n                classId: id,\n                className,\n                probability: values[i] as number,\n                cls: id,\n                name: className,\n                conf: values[i] as number,\n            });\n        }\n        if (probabilities.length === 0) {\n            throw new Error(\"Classifier produced no probabilities (empty output).\");\n        }\n\n        const top = probabilities[0] as ClassProbability;\n        const result: ClassificationResult = {\n            classId: top.classId,\n            className: top.className,\n            confidence: top.probability,\n            cls: top.classId,\n            name: top.className,\n            conf: top.probability,\n            image: original,\n            probabilities,\n        };\n\n        const orig: readonly [number, number] = [original.height, original.width];\n        const probs = new Probs(fullProbs);\n        timer.stage(\"postprocess\");\n        return [\n            new ClassificationResults(\n                probs,\n                result,\n                this._names,\n                original,\n                orig,\n                path,\n                timer.speed(),\n            ),\n        ];\n    }\n\n    private _preprocess(image: RGBImage): ort.Tensor {\n        const [tw, th] = this._inputSize;\n        const { data } = this._pipeline.run(image);\n        return toFloat32Tensor(data, [1, 3, th, tw]);\n    }\n\n    private _postprocess(raw: Float32Array): Float32Array {\n        return this._applySoftmax ? softmax(raw) : new Float32Array(raw);\n    }\n}\n"],"mappings":"kZAyGA,IAAa,EAAb,MAAa,UAAmB,EAAA,UAAW,CAGlB,QACA,OACA,WACA,MACA,KACA,cACA,eARrB,YACI,EACA,EACA,EACA,EACA,EACA,EACA,EACA,EACF,CACE,MAAM,CAAO,EARI,KAAA,QAAA,EACA,KAAA,OAAA,EACA,KAAA,WAAA,EACA,KAAA,MAAA,EACA,KAAA,KAAA,EACA,KAAA,cAAA,EACA,KAAA,eAAA,CAGrB,CAUA,IAAI,gBAA0B,CAC1B,OAAO,KAAK,aAChB,CAWA,IAAI,eAAwB,CACxB,OAAO,KAAK,cAChB,CAEA,eAAgD,KAchD,MAAM,OAAO,EAAe,EAAkB,CAC1C,GAAM,CAAC,EAAI,GAAM,KAAK,WACtB,IAAK,IAAI,EAAI,EAAG,EAAI,EAAM,IAAK,CAC3B,IAAM,EAAS,EAAA,gBAAgB,EAAA,eAAe,EAAI,CAAE,EAAG,CAAC,EAAG,EAAG,EAAI,CAAE,CAAC,EACrE,MAAM,KAAK,SAAS,IAAI,EAAG,KAAK,SAAS,WAAY,CAAO,CAAC,CACjE,CACJ,CASA,IAAY,WAA4B,CACpC,GAAI,KAAK,iBAAmB,KAAM,CAC9B,GAAM,CAAC,EAAI,GAAM,KAAK,WACtB,KAAK,eAAiB,IAAI,EAAA,eAAe,EAAI,EAAI,KAAK,MAAO,KAAK,IAAI,CAC1E,CACA,OAAO,KAAK,cAChB,CAUA,aAAa,OAAO,EAAoB,EAA6B,CAAC,EAAwB,CAC1F,IAAM,EAAU,MAAM,EAAA,WAAW,OAAO,EAAO,CAAO,EAChD,EAAgB,EAAA,qBAAqB,EAAQ,SAAU,CACzD,cAAe,EAAQ,cACvB,KAAM,EAAQ,KACd,IAAK,EAAQ,GACjB,CAAC,EACK,EACF,EAAQ,YAAc,EAAA,yBAAyB,EAAQ,WAAW,GAAK,IAAA,GACrE,EAAS,EAAA,cAAc,EAAQ,QAAU,EAAA,WAAW,EAAQ,QAAQ,EAAG,CACzE,YACJ,CAAC,EACK,EAAgC,CAAC,EACvC,IAAK,IAAI,EAAI,EAAG,EAAI,EAAO,OAAQ,IAC/B,EAAM,GAAK,EAAO,GAEtB,OAAO,IAAI,EACP,EACA,EACA,EACA,EAAA,iBAAiB,CACb,WAAY,EAAQ,WACpB,UAAW,EAAQ,UACnB,SAAU,CAAC,IAAK,GAAG,CACvB,CAAC,EACD,EAAc,KACd,EAAc,IACd,EAAQ,cAAgB,CAAC,EAAA,wBAAwB,EAAQ,QAAQ,EACjE,EAAc,IAClB,CACJ,CAGA,IAAI,QAA4B,CAC5B,OAAO,KAAK,OAChB,CAGA,IAAI,OAA0C,CAC1C,OAAO,KAAK,MAChB,CASA,IAAI,WAAuC,CACvC,OAAO,KAAK,UAChB,CAGA,IAAI,YAAqB,CACrB,OAAO,KAAK,QAAQ,MACxB,CAGA,MAAM,KACF,EACA,EAAoC,CAAC,EACL,CAChC,OAAO,KAAK,QAAQ,EAAO,CAAO,CACtC,CAGA,MAAM,QACF,EACA,EAAoC,CAAC,EACL,CAChC,IAAM,EAAQ,IAAI,EAAA,WACZ,EAAO,OAAO,GAAU,SAAW,EAAQ,KAC3C,EAAW,MAAM,EAAA,UAAU,CAAK,EACtC,EAAM,MAAM,MAAM,EAClB,IAAM,EAAS,KAAK,YAAY,CAAQ,EACxC,EAAM,MAAM,YAAY,EACxB,IAAM,EAAU,MAAM,KAAK,SAAS,IAAI,EAAG,KAAK,SAAS,WAAY,CAAO,CAAC,EAC7E,KAAK,UAAU,QAAQ,EACvB,EAAM,MAAM,WAAW,EACvB,IAAM,EAAkB,KAAK,SAAS,YAAY,GAClD,GAAI,IAAoB,IAAA,GACpB,MAAU,MAAM,kCAAkC,EAEtD,IAAM,EAAM,EAAQ,GACpB,GAAI,IAAQ,IAAA,GACR,MAAU,MACN,2BAA2B,EAAgB,4BAC/C,EAEJ,IAAM,EAAY,KAAK,aAAa,EAAI,IAAoB,EAEtD,CAAE,UAAS,UAAW,EAAA,KAAK,EAAW,EAAQ,MAAQ,IAAI,EAC1D,EAAoC,CAAC,EAC3C,IAAK,IAAI,EAAI,EAAG,EAAI,EAAQ,OAAQ,IAAK,CACrC,IAAM,EAAK,EAAQ,GACb,EAAY,KAAK,QAAQ,IAAO,SAAS,IAC/C,EAAc,KAAK,CACf,QAAS,EACT,YACA,YAAa,EAAO,GACpB,IAAK,EACL,KAAM,EACN,KAAM,EAAO,EACjB,CAAC,CACL,CACA,GAAI,EAAc,SAAW,EACzB,MAAU,MAAM,sDAAsD,EAG1E,IAAM,EAAM,EAAc,GACpB,EAA+B,CACjC,QAAS,EAAI,QACb,UAAW,EAAI,UACf,WAAY,EAAI,YAChB,IAAK,EAAI,QACT,KAAM,EAAI,UACV,KAAM,EAAI,YACV,MAAO,EACP,eACJ,EAEM,EAAkC,CAAC,EAAS,OAAQ,EAAS,KAAK,EAClE,EAAQ,IAAI,EAAA,MAAM,CAAS,EAEjC,OADA,EAAM,MAAM,aAAa,EAClB,CACH,IAAI,EAAA,sBACA,EACA,EACA,KAAK,OACL,EACA,EACA,EACA,EAAM,MAAM,CAChB,CACJ,CACJ,CAEA,YAAoB,EAA6B,CAC7C,GAAM,CAAC,EAAI,GAAM,KAAK,WAChB,CAAE,QAAS,KAAK,UAAU,IAAI,CAAK,EACzC,OAAO,EAAA,gBAAgB,EAAM,CAAC,EAAG,EAAG,EAAI,CAAE,CAAC,CAC/C,CAEA,aAAqB,EAAiC,CAClD,OAAO,KAAK,cAAgB,EAAA,QAAQ,CAAG,EAAI,IAAI,aAAa,CAAG,CACnE,CACJ"}
+{"version":3,"file":"classifier.cjs","names":[],"sources":["../../../src/vision/tasks/classifier.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Image classification task using ONNX Runtime Web.\n */\n\nimport type * as ort from \"onnxruntime-web\";\n\nimport { type ModelSource, type OrtSessionOptions, OrtSession } from \"../core/session\";\nimport { SpeedTimer } from \"../core/timing\";\nimport { type ImageInput, loadImageSource } from \"../io/image\";\nimport { classificationNumClasses, resolveInputSize } from \"../core/graph\";\nimport { modelNames } from \"../core/metadata\";\nimport { type LabelSpec, resolveLabels } from \"../labels\";\nimport {\n    type Normalization,\n    isUltralyticsClassifier,\n    resolveNormalization,\n} from \"../normalization\";\nimport { softmax, topK } from \"../postprocess/classification\";\nimport { asFloat32Array } from \"../core/dtypes\";\nimport { toFloat32Tensor } from \"../preprocess/image\";\nimport { ResizePipeline, zeroTensorData } from \"../preprocess/pipeline\";\nimport { ClassificationResults, Probs } from \"../results\";\nimport { VisionTask } from \"./base\";\nimport { type ClassProbability, type ClassificationResult, type RGBImage } from \"../types\";\n\nexport interface ClassifierOptions extends OrtSessionOptions {\n    /**\n     * Class label spec — see {@link resolveLabels}.\n     *\n     * Optional: when omitted, the names the export baked into the model are used\n     * (Ultralytics writes them as `names` in the metadata map). Only when the\n     * model carries none does this fall back to generated `class_<id>` labels.\n     * Passing a spec always wins, for a model whose names are wrong or absent.\n     */\n    readonly labels?: LabelSpec;\n    /**\n     * Number of classes the model can predict.\n     *\n     * Optional: inferred from the classification head's declared output shape\n     * `(B, nc)`. Pass it to validate that the supplied labels match the model.\n     */\n    readonly numClasses?: number;\n    /**\n     * Model input `[width, height]` in pixels.\n     *\n     * Only used when the model's graph leaves its spatial axes dynamic: a graph\n     * that declares a static size always wins, since that is the only shape ONNX\n     * Runtime will accept. Defaults to `[224, 224]`.\n     */\n    readonly inputSize?: readonly [number, number];\n    /**\n     * Which preprocessing this model expects — see {@link Normalization}.\n     *\n     * Defaults to `\"auto\"`, which reads the model's own export metadata and picks\n     * `\"ultralytics\"` (raw `[0, 1]`) for an Ultralytics classification head and\n     * `\"imagenet\"` for everything else.\n     */\n    readonly normalization?: Normalization;\n    /** Per-channel RGB mean, overriding the preset. Defaults to the preset's. */\n    readonly mean?: readonly [number, number, number];\n    /** Per-channel RGB standard deviation, overriding the preset. */\n    readonly std?: readonly [number, number, number];\n    /**\n     * Whether the model's output still needs a softmax.\n     *\n     * Left undefined (the default), this reads the model's metadata and answers\n     * `false` for an Ultralytics classification export, whose graph already ends\n     * in one — applying a second softmax to a probability vector keeps the\n     * ranking but flattens the confidences, so the top-1 stays right while every\n     * number attached to it is wrong. Detection covers that family; for any other\n     * model that already emits probabilities, pass `false` explicitly.\n     */\n    readonly applySoftmax?: boolean;\n}\n\nexport interface ClassifierPredictOptions {\n    /**\n     * If set, the per-class probability list in `results[0].result.probabilities`\n     * is truncated to the top-K entries. The bulk `probs` view always exposes\n     * the full vector.\n     */\n    readonly topK?: number;\n}\n\n/**\n * Image classifier wrapping an ONNX model with ImageNet-style preprocessing.\n *\n * `predict()` returns `Promise<ClassificationResults[]>` (length 1 for a\n * single image), mirroring Ultralytics' API. The envelope exposes a `probs`\n * collection (`top1`, `top1conf`, `top5`, `top5conf`, `data`) plus the\n * legacy per-class probability list with names already resolved.\n *\n * Defaults: 224×224 RGB input, `float32` normalized with ImageNet mean/std,\n * NCHW layout, batch size 1, softmax applied to the raw output.\n *\n * @example\n * ```typescript\n * const clf = await Classifier.create(\"/models/resnet50.onnx\", {\n *   labels: [\"tench\", \"goldfish\", ...]  // 1000 ImageNet labels\n * });\n * const r = (await clf.predict(\"/images/dog.jpg\"))[0];\n * console.log(r.cls, r.conf, r.name);\n * console.log(r.probs.top5, r.probs.top5conf);\n * ```\n */\nexport class Classifier extends VisionTask {\n    private constructor(\n        session: OrtSession,\n        private readonly _labels: readonly string[],\n        private readonly _names: Readonly<Record<number, string>>,\n        private readonly _inputSize: readonly [number, number],\n        private readonly _mean: readonly [number, number, number],\n        private readonly _std: readonly [number, number, number],\n        private readonly _applySoftmax: boolean,\n        private readonly _normalization: string,\n    ) {\n        super(session);\n    }\n\n    /**\n     * Whether a softmax is applied to the model's output before ranking.\n     *\n     * Resolved once at construction. Worth reading when confidences look\n     * compressed: a second softmax over an already-normalized vector leaves the\n     * ordering intact and the numbers meaningless, which is invisible to any check\n     * that only looks at the predicted class.\n     */\n    get appliesSoftmax(): boolean {\n        return this._applySoftmax;\n    }\n\n    /**\n     * Which preprocessing this classifier applies to every image.\n     *\n     * One of the {@link Normalization} preset names, or `\"custom\"` when the caller\n     * supplied `mean`/`std` directly. Worth reading when a model underperforms:\n     * feeding a classifier a differently prepared tensor than it was trained on\n     * degrades it without throwing anything, so \"what does this assume\" is the\n     * first question.\n     */\n    get normalization(): string {\n        return this._normalization;\n    }\n\n    private _pipelineCache: ResizePipeline | null = null;\n\n    /**\n     * Run the model once on a zero tensor, paying one-time costs up front.\n     *\n     * The first inference of a session is not representative: WebGPU compiles its\n     * shaders on it and the WASM backend faults in its arenas. Calling this while\n     * a loading spinner is still up moves that cost somewhere the user is already\n     * waiting — which matters most for a classifier running as the last step of\n     * an on-device analysis, where the delay lands right before the answer shows.\n     *\n     * @param runs How many warm-up inferences to run. One is enough for WASM;\n     *   WebGPU sometimes settles on the second.\n     */\n    async warmup(runs: number = 1): Promise<void> {\n        const [tw, th] = this._inputSize;\n        for (let i = 0; i < runs; i++) {\n            const tensor = toFloat32Tensor(zeroTensorData(tw, th), [1, 3, th, tw]);\n            await this._session.run({ [this._session.inputName]: tensor });\n        }\n    }\n\n    /**\n     * The fused preprocessing pipeline, built on first use.\n     *\n     * Lazily, because constructing it reserves the output buffer: a task built in\n     * an environment without a canvas implementation stays constructible, and only\n     * fails if it is actually asked to preprocess something.\n     */\n    private get _pipeline(): ResizePipeline {\n        if (this._pipelineCache === null) {\n            const [tw, th] = this._inputSize;\n            this._pipelineCache = new ResizePipeline(tw, th, this._mean, this._std);\n        }\n        return this._pipelineCache;\n    }\n\n    /**\n     * Load the model, resolve labels, and settle the preprocessing.\n     *\n     * @param model The model source — a URL, or the bytes.\n     * @param options Labels, input size, normalization, and session options.\n     * @throws {RangeError} If `normalization` names an unknown preset, or names\n     *   one while `mean`/`std` are also given.\n     */\n    static async create(model: ModelSource, options: ClassifierOptions = {}): Promise<Classifier> {\n        const session = await OrtSession.create(model, options);\n        const normalization = resolveNormalization(session.metadata, {\n            normalization: options.normalization,\n            mean: options.mean,\n            std: options.std,\n        });\n        const numClasses =\n            options.numClasses ?? classificationNumClasses(session.outputShape) ?? undefined;\n        const labels = resolveLabels(options.labels ?? modelNames(session.metadata), {\n            numClasses,\n        });\n        const names: Record<number, string> = {};\n        for (let i = 0; i < labels.length; i++) {\n            names[i] = labels[i] as string;\n        }\n        return new Classifier(\n            session,\n            labels,\n            names,\n            resolveInputSize({\n                graphShape: session.inputShape,\n                requested: options.inputSize,\n                fallback: [224, 224],\n            }),\n            normalization.mean,\n            normalization.std,\n            options.applySoftmax ?? !isUltralyticsClassifier(session.metadata),\n            normalization.name,\n        );\n    }\n\n    /** Class labels indexed by class id. */\n    get labels(): readonly string[] {\n        return this._labels;\n    }\n\n    /** Class id → class name dict (matches Ultralytics' `model.names`). */\n    get names(): Readonly<Record<number, string>> {\n        return this._names;\n    }\n\n    /**\n     * The `[width, height]` this task preprocesses to.\n     *\n     * Resolved at creation time from the model's graph when it declares a static\n     * input, so reading it back tells you the resolution inference really runs at\n     * — not merely what was requested.\n     */\n    get inputSize(): readonly [number, number] {\n        return this._inputSize;\n    }\n\n    /** Number of classes the model can predict. */\n    get numClasses(): number {\n        return this._labels.length;\n    }\n\n    /** Alias for {@link predict} (parity with PyTorch `nn.Module.__call__`). */\n    async call(\n        image: ImageInput,\n        options: ClassifierPredictOptions = {},\n    ): Promise<ClassificationResults[]> {\n        return this.predict(image, options);\n    }\n\n    /** Run classification on a single image. */\n    async predict(\n        image: ImageInput,\n        options: ClassifierPredictOptions = {},\n    ): Promise<ClassificationResults[]> {\n        const timer = new SpeedTimer();\n        const path = typeof image === \"string\" ? image : null;\n        const { image: original, canvas } = await loadImageSource(image);\n        timer.stage(\"load\");\n        const { tensor, reused } = this._preprocess(original, canvas);\n        timer.stage(\"preprocess\");\n        let outputs: Record<string, ort.Tensor>;\n        try {\n            outputs = await this._session.run({ [this._session.inputName]: tensor });\n        } finally {\n            this._pipeline.release({ reused });\n        }\n        timer.stage(\"inference\");\n        const firstOutputName = this._session.outputNames[0];\n        if (firstOutputName === undefined) {\n            throw new Error(\"Classifier model has no outputs.\");\n        }\n        const raw = outputs[firstOutputName];\n        if (raw === undefined) {\n            throw new Error(\n                `Classifier model output ${firstOutputName} missing from run() result.`,\n            );\n        }\n        const fullProbs = this._postprocess(asFloat32Array(raw.data));\n\n        const { indices, values } = topK(fullProbs, options.topK ?? null);\n        const probabilities: ClassProbability[] = [];\n        for (let i = 0; i < indices.length; i++) {\n            const id = indices[i] as number;\n            const className = this._labels[id] ?? `class_${id}`;\n            probabilities.push({\n                classId: id,\n                className,\n                probability: values[i] as number,\n                cls: id,\n                name: className,\n                conf: values[i] as number,\n            });\n        }\n        if (probabilities.length === 0) {\n            throw new Error(\"Classifier produced no probabilities (empty output).\");\n        }\n\n        const top = probabilities[0] as ClassProbability;\n        const result: ClassificationResult = {\n            classId: top.classId,\n            className: top.className,\n            confidence: top.probability,\n            cls: top.classId,\n            name: top.className,\n            conf: top.probability,\n            image: original,\n            probabilities,\n        };\n\n        const orig: readonly [number, number] = [original.height, original.width];\n        const probs = new Probs(fullProbs);\n        timer.stage(\"postprocess\");\n        return [\n            new ClassificationResults(\n                probs,\n                result,\n                this._names,\n                original,\n                orig,\n                path,\n                timer.speed(),\n            ),\n        ];\n    }\n\n    /**\n     * Resize and pack the image into the tensor the model expects.\n     *\n     * @param image The decoded input.\n     * @param canvas The opaque canvas it was decoded on, when there is one — see\n     *   {@link ResizePipeline.run}.\n     */\n    private _preprocess(\n        image: RGBImage,\n        canvas: CanvasImageSource | null,\n    ): { tensor: ort.Tensor; reused: boolean } {\n        const [tw, th] = this._inputSize;\n        const { data, reused } = this._pipeline.run(image, canvas);\n        return { tensor: toFloat32Tensor(data, [1, 3, th, tw]), reused };\n    }\n\n    private _postprocess(raw: Float32Array): Float32Array {\n        return this._applySoftmax ? softmax(raw) : new Float32Array(raw);\n    }\n}\n"],"mappings":"kbA0GA,IAAa,EAAb,MAAa,UAAmB,EAAA,UAAW,CAGlB,QACA,OACA,WACA,MACA,KACA,cACA,eARrB,YACI,EACA,EACA,EACA,EACA,EACA,EACA,EACA,EACF,CACE,MAAM,CAAO,EARI,KAAA,QAAA,EACA,KAAA,OAAA,EACA,KAAA,WAAA,EACA,KAAA,MAAA,EACA,KAAA,KAAA,EACA,KAAA,cAAA,EACA,KAAA,eAAA,CAGrB,CAUA,IAAI,gBAA0B,CAC1B,OAAO,KAAK,aAChB,CAWA,IAAI,eAAwB,CACxB,OAAO,KAAK,cAChB,CAEA,eAAgD,KAchD,MAAM,OAAO,EAAe,EAAkB,CAC1C,GAAM,CAAC,EAAI,GAAM,KAAK,WACtB,IAAK,IAAI,EAAI,EAAG,EAAI,EAAM,IAAK,CAC3B,IAAM,EAAS,EAAA,gBAAgB,EAAA,eAAe,EAAI,CAAE,EAAG,CAAC,EAAG,EAAG,EAAI,CAAE,CAAC,EACrE,MAAM,KAAK,SAAS,IAAI,EAAG,KAAK,SAAS,WAAY,CAAO,CAAC,CACjE,CACJ,CASA,IAAY,WAA4B,CACpC,GAAI,KAAK,iBAAmB,KAAM,CAC9B,GAAM,CAAC,EAAI,GAAM,KAAK,WACtB,KAAK,eAAiB,IAAI,EAAA,eAAe,EAAI,EAAI,KAAK,MAAO,KAAK,IAAI,CAC1E,CACA,OAAO,KAAK,cAChB,CAUA,aAAa,OAAO,EAAoB,EAA6B,CAAC,EAAwB,CAC1F,IAAM,EAAU,MAAM,EAAA,WAAW,OAAO,EAAO,CAAO,EAChD,EAAgB,EAAA,qBAAqB,EAAQ,SAAU,CACzD,cAAe,EAAQ,cACvB,KAAM,EAAQ,KACd,IAAK,EAAQ,GACjB,CAAC,EACK,EACF,EAAQ,YAAc,EAAA,yBAAyB,EAAQ,WAAW,GAAK,IAAA,GACrE,EAAS,EAAA,cAAc,EAAQ,QAAU,EAAA,WAAW,EAAQ,QAAQ,EAAG,CACzE,YACJ,CAAC,EACK,EAAgC,CAAC,EACvC,IAAK,IAAI,EAAI,EAAG,EAAI,EAAO,OAAQ,IAC/B,EAAM,GAAK,EAAO,GAEtB,OAAO,IAAI,EACP,EACA,EACA,EACA,EAAA,iBAAiB,CACb,WAAY,EAAQ,WACpB,UAAW,EAAQ,UACnB,SAAU,CAAC,IAAK,GAAG,CACvB,CAAC,EACD,EAAc,KACd,EAAc,IACd,EAAQ,cAAgB,CAAC,EAAA,wBAAwB,EAAQ,QAAQ,EACjE,EAAc,IAClB,CACJ,CAGA,IAAI,QAA4B,CAC5B,OAAO,KAAK,OAChB,CAGA,IAAI,OAA0C,CAC1C,OAAO,KAAK,MAChB,CASA,IAAI,WAAuC,CACvC,OAAO,KAAK,UAChB,CAGA,IAAI,YAAqB,CACrB,OAAO,KAAK,QAAQ,MACxB,CAGA,MAAM,KACF,EACA,EAAoC,CAAC,EACL,CAChC,OAAO,KAAK,QAAQ,EAAO,CAAO,CACtC,CAGA,MAAM,QACF,EACA,EAAoC,CAAC,EACL,CAChC,IAAM,EAAQ,IAAI,EAAA,WACZ,EAAO,OAAO,GAAU,SAAW,EAAQ,KAC3C,CAAE,MAAO,EAAU,UAAW,MAAM,EAAA,gBAAgB,CAAK,EAC/D,EAAM,MAAM,MAAM,EAClB,GAAM,CAAE,SAAQ,UAAW,KAAK,YAAY,EAAU,CAAM,EAC5D,EAAM,MAAM,YAAY,EACxB,IAAI,EACJ,GAAI,CACA,EAAU,MAAM,KAAK,SAAS,IAAI,EAAG,KAAK,SAAS,WAAY,CAAO,CAAC,CAC3E,QAAU,CACN,KAAK,UAAU,QAAQ,CAAE,QAAO,CAAC,CACrC,CACA,EAAM,MAAM,WAAW,EACvB,IAAM,EAAkB,KAAK,SAAS,YAAY,GAClD,GAAI,IAAoB,IAAA,GACpB,MAAU,MAAM,kCAAkC,EAEtD,IAAM,EAAM,EAAQ,GACpB,GAAI,IAAQ,IAAA,GACR,MAAU,MACN,2BAA2B,EAAgB,4BAC/C,EAEJ,IAAM,EAAY,KAAK,aAAa,EAAA,eAAe,EAAI,IAAI,CAAC,EAEtD,CAAE,UAAS,UAAW,EAAA,KAAK,EAAW,EAAQ,MAAQ,IAAI,EAC1D,EAAoC,CAAC,EAC3C,IAAK,IAAI,EAAI,EAAG,EAAI,EAAQ,OAAQ,IAAK,CACrC,IAAM,EAAK,EAAQ,GACb,EAAY,KAAK,QAAQ,IAAO,SAAS,IAC/C,EAAc,KAAK,CACf,QAAS,EACT,YACA,YAAa,EAAO,GACpB,IAAK,EACL,KAAM,EACN,KAAM,EAAO,EACjB,CAAC,CACL,CACA,GAAI,EAAc,SAAW,EACzB,MAAU,MAAM,sDAAsD,EAG1E,IAAM,EAAM,EAAc,GACpB,EAA+B,CACjC,QAAS,EAAI,QACb,UAAW,EAAI,UACf,WAAY,EAAI,YAChB,IAAK,EAAI,QACT,KAAM,EAAI,UACV,KAAM,EAAI,YACV,MAAO,EACP,eACJ,EAEM,EAAkC,CAAC,EAAS,OAAQ,EAAS,KAAK,EAClE,EAAQ,IAAI,EAAA,MAAM,CAAS,EAEjC,OADA,EAAM,MAAM,aAAa,EAClB,CACH,IAAI,EAAA,sBACA,EACA,EACA,KAAK,OACL,EACA,EACA,EACA,EAAM,MAAM,CAChB,CACJ,CACJ,CASA,YACI,EACA,EACuC,CACvC,GAAM,CAAC,EAAI,GAAM,KAAK,WAChB,CAAE,OAAM,UAAW,KAAK,UAAU,IAAI,EAAO,CAAM,EACzD,MAAO,CAAE,OAAQ,EAAA,gBAAgB,EAAM,CAAC,EAAG,EAAG,EAAI,CAAE,CAAC,EAAG,QAAO,CACnE,CAEA,aAAqB,EAAiC,CAClD,OAAO,KAAK,cAAgB,EAAA,QAAQ,CAAG,EAAI,IAAI,aAAa,CAAG,CACnE,CACJ"}

package/dist/vision/tasks/classifier.js

@@ -1,17 +1,18 @@
 import { SpeedTimer as e } from "../core/timing.js";
 import { ClassificationResults as t, Probs as n } from "../results.js";
 import { resolveLabels as r } from "../labels.js";
-import { classificationNumClasses as i, resolveInputSize as a } from "../core/graph.js";
-import { modelNames as o } from "../core/metadata.js";
-import { OrtSession as s } from "../core/session.js";
-import { isUltralyticsClassifier as c, resolveNormalization as l } from "../normalization.js";
-import { loadImage as u } from "../io/image.js";
-import { toFloat32Tensor as d } from "../preprocess/image.js";
-import { ResizePipeline as f, zeroTensorData as p } from "../preprocess/pipeline.js";
-import { softmax as m, topK as h } from "../postprocess/classification.js";
-import { VisionTask as g } from "./base.js";
+import { asFloat32Array as i } from "../core/dtypes.js";
+import { classificationNumClasses as a, resolveInputSize as o } from "../core/graph.js";
+import { modelNames as s } from "../core/metadata.js";
+import { OrtSession as c } from "../core/session.js";
+import { isUltralyticsClassifier as l, resolveNormalization as u } from "../normalization.js";
+import { loadImageSource as d } from "../io/image.js";
+import { toFloat32Tensor as f } from "../preprocess/image.js";
+import { ResizePipeline as p, zeroTensorData as m } from "../preprocess/pipeline.js";
+import { softmax as h, topK as g } from "../postprocess/classification.js";
+import { VisionTask as _ } from "./base.js";
 //#region src/vision/tasks/classifier.ts
-var _ = class _ extends g {
+var v = class v extends _ {
 	_labels;
 	_names;
 	_inputSize;
@@ -32,7 +33,7 @@ var _ = class _ extends g {
 	async warmup(e = 1) {
 		let [t, n] = this._inputSize;
 		for (let r = 0; r < e; r++) {
-			let e = d(p(t, n), [
+			let e = f(m(t, n), [
 				1,
 				3,
 				n,
@@ -44,22 +45,22 @@ var _ = class _ extends g {
 	get _pipeline() {
 		if (this._pipelineCache === null) {
 			let [e, t] = this._inputSize;
-			this._pipelineCache = new f(e, t, this._mean, this._std);
+			this._pipelineCache = new p(e, t, this._mean, this._std);
 		}
 		return this._pipelineCache;
 	}
 	static async create(e, t = {}) {
-		let n = await s.create(e, t), u = l(n.metadata, {
+		let n = await c.create(e, t), i = u(n.metadata, {
 			normalization: t.normalization,
 			mean: t.mean,
 			std: t.std
-		}), d = t.numClasses ?? i(n.outputShape) ?? void 0, f = r(t.labels ?? o(n.metadata), { numClasses: d }), p = {};
+		}), d = t.numClasses ?? a(n.outputShape) ?? void 0, f = r(t.labels ?? s(n.metadata), { numClasses: d }), p = {};
 		for (let e = 0; e < f.length; e++) p[e] = f[e];
-		return new _(n, f, p, a({
+		return new v(n, f, p, o({
 			graphShape: n.inputShape,
 			requested: t.inputSize,
 			fallback: [224, 224]
-		}), u.mean, u.std, t.applySoftmax ?? !c(n.metadata), u.name);
+		}), i.mean, i.std, t.applySoftmax ?? !l(n.metadata), i.name);
 	}
 	get labels() {
 		return this._labels;
@@ -76,56 +77,64 @@ var _ = class _ extends g {
 	async call(e, t = {}) {
 		return this.predict(e, t);
 	}
-	async predict(r, i = {}) {
-		let a = new e(), o = typeof r == "string" ? r : null, s = await u(r);
-		a.stage("load");
-		let c = this._preprocess(s);
-		a.stage("preprocess");
-		let l = await this._session.run({ [this._session.inputName]: c });
-		this._pipeline.release(), a.stage("inference");
-		let d = this._session.outputNames[0];
-		if (d === void 0) throw Error("Classifier model has no outputs.");
-		let f = l[d];
-		if (f === void 0) throw Error(`Classifier model output ${d} missing from run() result.`);
-		let p = this._postprocess(f.data), { indices: m, values: g } = h(p, i.topK ?? null), _ = [];
-		for (let e = 0; e < m.length; e++) {
-			let t = m[e], n = this._labels[t] ?? `class_${t}`;
-			_.push({
+	async predict(r, a = {}) {
+		let o = new e(), s = typeof r == "string" ? r : null, { image: c, canvas: l } = await d(r);
+		o.stage("load");
+		let { tensor: u, reused: f } = this._preprocess(c, l);
+		o.stage("preprocess");
+		let p;
+		try {
+			p = await this._session.run({ [this._session.inputName]: u });
+		} finally {
+			this._pipeline.release({ reused: f });
+		}
+		o.stage("inference");
+		let m = this._session.outputNames[0];
+		if (m === void 0) throw Error("Classifier model has no outputs.");
+		let h = p[m];
+		if (h === void 0) throw Error(`Classifier model output ${m} missing from run() result.`);
+		let _ = this._postprocess(i(h.data)), { indices: v, values: y } = g(_, a.topK ?? null), b = [];
+		for (let e = 0; e < v.length; e++) {
+			let t = v[e], n = this._labels[t] ?? `class_${t}`;
+			b.push({
 				classId: t,
 				className: n,
-				probability: g[e],
+				probability: y[e],
 				cls: t,
 				name: n,
-				conf: g[e]
+				conf: y[e]
 			});
 		}
-		if (_.length === 0) throw Error("Classifier produced no probabilities (empty output).");
-		let v = _[0], y = {
-			classId: v.classId,
-			className: v.className,
-			confidence: v.probability,
-			cls: v.classId,
-			name: v.className,
-			conf: v.probability,
-			image: s,
-			probabilities: _
-		}, b = [s.height, s.width], x = new n(p);
-		return a.stage("postprocess"), [new t(x, y, this._names, s, b, o, a.speed())];
+		if (b.length === 0) throw Error("Classifier produced no probabilities (empty output).");
+		let x = b[0], S = {
+			classId: x.classId,
+			className: x.className,
+			confidence: x.probability,
+			cls: x.classId,
+			name: x.className,
+			conf: x.probability,
+			image: c,
+			probabilities: b
+		}, C = [c.height, c.width], w = new n(_);
+		return o.stage("postprocess"), [new t(w, S, this._names, c, C, s, o.speed())];
 	}
-	_preprocess(e) {
-		let [t, n] = this._inputSize, { data: r } = this._pipeline.run(e);
-		return d(r, [
-			1,
-			3,
-			n,
-			t
-		]);
+	_preprocess(e, t) {
+		let [n, r] = this._inputSize, { data: i, reused: a } = this._pipeline.run(e, t);
+		return {
+			tensor: f(i, [
+				1,
+				3,
+				r,
+				n
+			]),
+			reused: a
+		};
 	}
 	_postprocess(e) {
-		return this._applySoftmax ? m(e) : new Float32Array(e);
+		return this._applySoftmax ? h(e) : new Float32Array(e);
 	}
 };
 //#endregion
-export { _ as Classifier };
+export { v as Classifier };
 
 //# sourceMappingURL=classifier.js.map

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

@@ -1 +1 @@
-{"version":3,"file":"classifier.js","names":[],"sources":["../../../src/vision/tasks/classifier.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Image classification task using ONNX Runtime Web.\n */\n\nimport type * as ort from \"onnxruntime-web\";\n\nimport { type ModelSource, type OrtSessionOptions, OrtSession } from \"../core/session\";\nimport { SpeedTimer } from \"../core/timing\";\nimport { type ImageInput, loadImage } from \"../io/image\";\nimport { classificationNumClasses, resolveInputSize } from \"../core/graph\";\nimport { modelNames } from \"../core/metadata\";\nimport { type LabelSpec, resolveLabels } from \"../labels\";\nimport {\n    type Normalization,\n    isUltralyticsClassifier,\n    resolveNormalization,\n} from \"../normalization\";\nimport { softmax, topK } from \"../postprocess/classification\";\nimport { toFloat32Tensor } from \"../preprocess/image\";\nimport { ResizePipeline, zeroTensorData } from \"../preprocess/pipeline\";\nimport { ClassificationResults, Probs } from \"../results\";\nimport { VisionTask } from \"./base\";\nimport { type ClassProbability, type ClassificationResult, type RGBImage } from \"../types\";\n\nexport interface ClassifierOptions extends OrtSessionOptions {\n    /**\n     * Class label spec — see {@link resolveLabels}.\n     *\n     * Optional: when omitted, the names the export baked into the model are used\n     * (Ultralytics writes them as `names` in the metadata map). Only when the\n     * model carries none does this fall back to generated `class_<id>` labels.\n     * Passing a spec always wins, for a model whose names are wrong or absent.\n     */\n    readonly labels?: LabelSpec;\n    /**\n     * Number of classes the model can predict.\n     *\n     * Optional: inferred from the classification head's declared output shape\n     * `(B, nc)`. Pass it to validate that the supplied labels match the model.\n     */\n    readonly numClasses?: number;\n    /**\n     * Model input `[width, height]` in pixels.\n     *\n     * Only used when the model's graph leaves its spatial axes dynamic: a graph\n     * that declares a static size always wins, since that is the only shape ONNX\n     * Runtime will accept. Defaults to `[224, 224]`.\n     */\n    readonly inputSize?: readonly [number, number];\n    /**\n     * Which preprocessing this model expects — see {@link Normalization}.\n     *\n     * Defaults to `\"auto\"`, which reads the model's own export metadata and picks\n     * `\"ultralytics\"` (raw `[0, 1]`) for an Ultralytics classification head and\n     * `\"imagenet\"` for everything else.\n     */\n    readonly normalization?: Normalization;\n    /** Per-channel RGB mean, overriding the preset. Defaults to the preset's. */\n    readonly mean?: readonly [number, number, number];\n    /** Per-channel RGB standard deviation, overriding the preset. */\n    readonly std?: readonly [number, number, number];\n    /**\n     * Whether the model's output still needs a softmax.\n     *\n     * Left undefined (the default), this reads the model's metadata and answers\n     * `false` for an Ultralytics classification export, whose graph already ends\n     * in one — applying a second softmax to a probability vector keeps the\n     * ranking but flattens the confidences, so the top-1 stays right while every\n     * number attached to it is wrong. Detection covers that family; for any other\n     * model that already emits probabilities, pass `false` explicitly.\n     */\n    readonly applySoftmax?: boolean;\n}\n\nexport interface ClassifierPredictOptions {\n    /**\n     * If set, the per-class probability list in `results[0].result.probabilities`\n     * is truncated to the top-K entries. The bulk `probs` view always exposes\n     * the full vector.\n     */\n    readonly topK?: number;\n}\n\n/**\n * Image classifier wrapping an ONNX model with ImageNet-style preprocessing.\n *\n * `predict()` returns `Promise<ClassificationResults[]>` (length 1 for a\n * single image), mirroring Ultralytics' API. The envelope exposes a `probs`\n * collection (`top1`, `top1conf`, `top5`, `top5conf`, `data`) plus the\n * legacy per-class probability list with names already resolved.\n *\n * Defaults: 224×224 RGB input, `float32` normalized with ImageNet mean/std,\n * NCHW layout, batch size 1, softmax applied to the raw output.\n *\n * @example\n * ```typescript\n * const clf = await Classifier.create(\"/models/resnet50.onnx\", {\n *   labels: [\"tench\", \"goldfish\", ...]  // 1000 ImageNet labels\n * });\n * const r = (await clf.predict(\"/images/dog.jpg\"))[0];\n * console.log(r.cls, r.conf, r.name);\n * console.log(r.probs.top5, r.probs.top5conf);\n * ```\n */\nexport class Classifier extends VisionTask {\n    private constructor(\n        session: OrtSession,\n        private readonly _labels: readonly string[],\n        private readonly _names: Readonly<Record<number, string>>,\n        private readonly _inputSize: readonly [number, number],\n        private readonly _mean: readonly [number, number, number],\n        private readonly _std: readonly [number, number, number],\n        private readonly _applySoftmax: boolean,\n        private readonly _normalization: string,\n    ) {\n        super(session);\n    }\n\n    /**\n     * Whether a softmax is applied to the model's output before ranking.\n     *\n     * Resolved once at construction. Worth reading when confidences look\n     * compressed: a second softmax over an already-normalized vector leaves the\n     * ordering intact and the numbers meaningless, which is invisible to any check\n     * that only looks at the predicted class.\n     */\n    get appliesSoftmax(): boolean {\n        return this._applySoftmax;\n    }\n\n    /**\n     * Which preprocessing this classifier applies to every image.\n     *\n     * One of the {@link Normalization} preset names, or `\"custom\"` when the caller\n     * supplied `mean`/`std` directly. Worth reading when a model underperforms:\n     * feeding a classifier a differently prepared tensor than it was trained on\n     * degrades it without throwing anything, so \"what does this assume\" is the\n     * first question.\n     */\n    get normalization(): string {\n        return this._normalization;\n    }\n\n    private _pipelineCache: ResizePipeline | null = null;\n\n    /**\n     * Run the model once on a zero tensor, paying one-time costs up front.\n     *\n     * The first inference of a session is not representative: WebGPU compiles its\n     * shaders on it and the WASM backend faults in its arenas. Calling this while\n     * a loading spinner is still up moves that cost somewhere the user is already\n     * waiting — which matters most for a classifier running as the last step of\n     * an on-device analysis, where the delay lands right before the answer shows.\n     *\n     * @param runs How many warm-up inferences to run. One is enough for WASM;\n     *   WebGPU sometimes settles on the second.\n     */\n    async warmup(runs: number = 1): Promise<void> {\n        const [tw, th] = this._inputSize;\n        for (let i = 0; i < runs; i++) {\n            const tensor = toFloat32Tensor(zeroTensorData(tw, th), [1, 3, th, tw]);\n            await this._session.run({ [this._session.inputName]: tensor });\n        }\n    }\n\n    /**\n     * The fused preprocessing pipeline, built on first use.\n     *\n     * Lazily, because constructing it reserves the output buffer: a task built in\n     * an environment without a canvas implementation stays constructible, and only\n     * fails if it is actually asked to preprocess something.\n     */\n    private get _pipeline(): ResizePipeline {\n        if (this._pipelineCache === null) {\n            const [tw, th] = this._inputSize;\n            this._pipelineCache = new ResizePipeline(tw, th, this._mean, this._std);\n        }\n        return this._pipelineCache;\n    }\n\n    /**\n     * Load the model, resolve labels, and settle the preprocessing.\n     *\n     * @param model The model source — a URL, or the bytes.\n     * @param options Labels, input size, normalization, and session options.\n     * @throws {RangeError} If `normalization` names an unknown preset, or names\n     *   one while `mean`/`std` are also given.\n     */\n    static async create(model: ModelSource, options: ClassifierOptions = {}): Promise<Classifier> {\n        const session = await OrtSession.create(model, options);\n        const normalization = resolveNormalization(session.metadata, {\n            normalization: options.normalization,\n            mean: options.mean,\n            std: options.std,\n        });\n        const numClasses =\n            options.numClasses ?? classificationNumClasses(session.outputShape) ?? undefined;\n        const labels = resolveLabels(options.labels ?? modelNames(session.metadata), {\n            numClasses,\n        });\n        const names: Record<number, string> = {};\n        for (let i = 0; i < labels.length; i++) {\n            names[i] = labels[i] as string;\n        }\n        return new Classifier(\n            session,\n            labels,\n            names,\n            resolveInputSize({\n                graphShape: session.inputShape,\n                requested: options.inputSize,\n                fallback: [224, 224],\n            }),\n            normalization.mean,\n            normalization.std,\n            options.applySoftmax ?? !isUltralyticsClassifier(session.metadata),\n            normalization.name,\n        );\n    }\n\n    /** Class labels indexed by class id. */\n    get labels(): readonly string[] {\n        return this._labels;\n    }\n\n    /** Class id → class name dict (matches Ultralytics' `model.names`). */\n    get names(): Readonly<Record<number, string>> {\n        return this._names;\n    }\n\n    /**\n     * The `[width, height]` this task preprocesses to.\n     *\n     * Resolved at creation time from the model's graph when it declares a static\n     * input, so reading it back tells you the resolution inference really runs at\n     * — not merely what was requested.\n     */\n    get inputSize(): readonly [number, number] {\n        return this._inputSize;\n    }\n\n    /** Number of classes the model can predict. */\n    get numClasses(): number {\n        return this._labels.length;\n    }\n\n    /** Alias for {@link predict} (parity with PyTorch `nn.Module.__call__`). */\n    async call(\n        image: ImageInput,\n        options: ClassifierPredictOptions = {},\n    ): Promise<ClassificationResults[]> {\n        return this.predict(image, options);\n    }\n\n    /** Run classification on a single image. */\n    async predict(\n        image: ImageInput,\n        options: ClassifierPredictOptions = {},\n    ): Promise<ClassificationResults[]> {\n        const timer = new SpeedTimer();\n        const path = typeof image === \"string\" ? image : null;\n        const original = await loadImage(image);\n        timer.stage(\"load\");\n        const tensor = this._preprocess(original);\n        timer.stage(\"preprocess\");\n        const outputs = await this._session.run({ [this._session.inputName]: tensor });\n        this._pipeline.release();\n        timer.stage(\"inference\");\n        const firstOutputName = this._session.outputNames[0];\n        if (firstOutputName === undefined) {\n            throw new Error(\"Classifier model has no outputs.\");\n        }\n        const raw = outputs[firstOutputName];\n        if (raw === undefined) {\n            throw new Error(\n                `Classifier model output ${firstOutputName} missing from run() result.`,\n            );\n        }\n        const fullProbs = this._postprocess(raw.data as Float32Array);\n\n        const { indices, values } = topK(fullProbs, options.topK ?? null);\n        const probabilities: ClassProbability[] = [];\n        for (let i = 0; i < indices.length; i++) {\n            const id = indices[i] as number;\n            const className = this._labels[id] ?? `class_${id}`;\n            probabilities.push({\n                classId: id,\n                className,\n                probability: values[i] as number,\n                cls: id,\n                name: className,\n                conf: values[i] as number,\n            });\n        }\n        if (probabilities.length === 0) {\n            throw new Error(\"Classifier produced no probabilities (empty output).\");\n        }\n\n        const top = probabilities[0] as ClassProbability;\n        const result: ClassificationResult = {\n            classId: top.classId,\n            className: top.className,\n            confidence: top.probability,\n            cls: top.classId,\n            name: top.className,\n            conf: top.probability,\n            image: original,\n            probabilities,\n        };\n\n        const orig: readonly [number, number] = [original.height, original.width];\n        const probs = new Probs(fullProbs);\n        timer.stage(\"postprocess\");\n        return [\n            new ClassificationResults(\n                probs,\n                result,\n                this._names,\n                original,\n                orig,\n                path,\n                timer.speed(),\n            ),\n        ];\n    }\n\n    private _preprocess(image: RGBImage): ort.Tensor {\n        const [tw, th] = this._inputSize;\n        const { data } = this._pipeline.run(image);\n        return toFloat32Tensor(data, [1, 3, th, tw]);\n    }\n\n    private _postprocess(raw: Float32Array): Float32Array {\n        return this._applySoftmax ? softmax(raw) : new Float32Array(raw);\n    }\n}\n"],"mappings":";;;;;;;;;;;;;AAyGA,IAAa,IAAb,MAAa,UAAmB,EAAW;CAGlB;CACA;CACA;CACA;CACA;CACA;CACA;CARrB,YACI,GACA,GACA,GACA,GACA,GACA,GACA,GACA,GACF;EADmB,AAEjB,MAAM,CAAO,GARI,KAAA,UAAA,GACA,KAAA,SAAA,GACA,KAAA,aAAA,GACA,KAAA,QAAA,GACA,KAAA,OAAA,GACA,KAAA,gBAAA,GACA,KAAA,iBAAA;CAGrB;CAUA,IAAI,iBAA0B;EAC1B,OAAO,KAAK;CAChB;CAWA,IAAI,gBAAwB;EACxB,OAAO,KAAK;CAChB;CAEA,iBAAgD;CAchD,MAAM,OAAO,IAAe,GAAkB;EAC1C,IAAM,CAAC,GAAI,KAAM,KAAK;EACtB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAM,KAAK;GAC3B,IAAM,IAAS,EAAgB,EAAe,GAAI,CAAE,GAAG;IAAC;IAAG;IAAG;IAAI;GAAE,CAAC;GACrE,MAAM,KAAK,SAAS,IAAI,GAAG,KAAK,SAAS,YAAY,EAAO,CAAC;EACjE;CACJ;CASA,IAAY,YAA4B;EACpC,IAAI,KAAK,mBAAmB,MAAM;GAC9B,IAAM,CAAC,GAAI,KAAM,KAAK;GACtB,KAAK,iBAAiB,IAAI,EAAe,GAAI,GAAI,KAAK,OAAO,KAAK,IAAI;EAC1E;EACA,OAAO,KAAK;CAChB;CAUA,aAAa,OAAO,GAAoB,IAA6B,CAAC,GAAwB;EAC1F,IAAM,IAAU,MAAM,EAAW,OAAO,GAAO,CAAO,GAChD,IAAgB,EAAqB,EAAQ,UAAU;GACzD,eAAe,EAAQ;GACvB,MAAM,EAAQ;GACd,KAAK,EAAQ;EACjB,CAAC,GACK,IACF,EAAQ,cAAc,EAAyB,EAAQ,WAAW,KAAK,KAAA,GACrE,IAAS,EAAc,EAAQ,UAAU,EAAW,EAAQ,QAAQ,GAAG,EACzE,cACJ,CAAC,GACK,IAAgC,CAAC;EACvC,KAAK,IAAI,IAAI,GAAG,IAAI,EAAO,QAAQ,KAC/B,EAAM,KAAK,EAAO;EAEtB,OAAO,IAAI,EACP,GACA,GACA,GACA,EAAiB;GACb,YAAY,EAAQ;GACpB,WAAW,EAAQ;GACnB,UAAU,CAAC,KAAK,GAAG;EACvB,CAAC,GACD,EAAc,MACd,EAAc,KACd,EAAQ,gBAAgB,CAAC,EAAwB,EAAQ,QAAQ,GACjE,EAAc,IAClB;CACJ;CAGA,IAAI,SAA4B;EAC5B,OAAO,KAAK;CAChB;CAGA,IAAI,QAA0C;EAC1C,OAAO,KAAK;CAChB;CASA,IAAI,YAAuC;EACvC,OAAO,KAAK;CAChB;CAGA,IAAI,aAAqB;EACrB,OAAO,KAAK,QAAQ;CACxB;CAGA,MAAM,KACF,GACA,IAAoC,CAAC,GACL;EAChC,OAAO,KAAK,QAAQ,GAAO,CAAO;CACtC;CAGA,MAAM,QACF,GACA,IAAoC,CAAC,GACL;EAChC,IAAM,IAAQ,IAAI,EAAW,GACvB,IAAO,OAAO,KAAU,WAAW,IAAQ,MAC3C,IAAW,MAAM,EAAU,CAAK;EACtC,EAAM,MAAM,MAAM;EAClB,IAAM,IAAS,KAAK,YAAY,CAAQ;EACxC,EAAM,MAAM,YAAY;EACxB,IAAM,IAAU,MAAM,KAAK,SAAS,IAAI,GAAG,KAAK,SAAS,YAAY,EAAO,CAAC;EAE7E,AADA,KAAK,UAAU,QAAQ,GACvB,EAAM,MAAM,WAAW;EACvB,IAAM,IAAkB,KAAK,SAAS,YAAY;EAClD,IAAI,MAAoB,KAAA,GACpB,MAAU,MAAM,kCAAkC;EAEtD,IAAM,IAAM,EAAQ;EACpB,IAAI,MAAQ,KAAA,GACR,MAAU,MACN,2BAA2B,EAAgB,4BAC/C;EAEJ,IAAM,IAAY,KAAK,aAAa,EAAI,IAAoB,GAEtD,EAAE,YAAS,cAAW,EAAK,GAAW,EAAQ,QAAQ,IAAI,GAC1D,IAAoC,CAAC;EAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,EAAQ,QAAQ,KAAK;GACrC,IAAM,IAAK,EAAQ,IACb,IAAY,KAAK,QAAQ,MAAO,SAAS;GAC/C,EAAc,KAAK;IACf,SAAS;IACT;IACA,aAAa,EAAO;IACpB,KAAK;IACL,MAAM;IACN,MAAM,EAAO;GACjB,CAAC;EACL;EACA,IAAI,EAAc,WAAW,GACzB,MAAU,MAAM,sDAAsD;EAG1E,IAAM,IAAM,EAAc,IACpB,IAA+B;GACjC,SAAS,EAAI;GACb,WAAW,EAAI;GACf,YAAY,EAAI;GAChB,KAAK,EAAI;GACT,MAAM,EAAI;GACV,MAAM,EAAI;GACV,OAAO;GACP;EACJ,GAEM,IAAkC,CAAC,EAAS,QAAQ,EAAS,KAAK,GAClE,IAAQ,IAAI,EAAM,CAAS;EAEjC,OADA,EAAM,MAAM,aAAa,GAClB,CACH,IAAI,EACA,GACA,GACA,KAAK,QACL,GACA,GACA,GACA,EAAM,MAAM,CAChB,CACJ;CACJ;CAEA,YAAoB,GAA6B;EAC7C,IAAM,CAAC,GAAI,KAAM,KAAK,YAChB,EAAE,YAAS,KAAK,UAAU,IAAI,CAAK;EACzC,OAAO,EAAgB,GAAM;GAAC;GAAG;GAAG;GAAI;EAAE,CAAC;CAC/C;CAEA,aAAqB,GAAiC;EAClD,OAAO,KAAK,gBAAgB,EAAQ,CAAG,IAAI,IAAI,aAAa,CAAG;CACnE;AACJ"}
+{"version":3,"file":"classifier.js","names":[],"sources":["../../../src/vision/tasks/classifier.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Image classification task using ONNX Runtime Web.\n */\n\nimport type * as ort from \"onnxruntime-web\";\n\nimport { type ModelSource, type OrtSessionOptions, OrtSession } from \"../core/session\";\nimport { SpeedTimer } from \"../core/timing\";\nimport { type ImageInput, loadImageSource } from \"../io/image\";\nimport { classificationNumClasses, resolveInputSize } from \"../core/graph\";\nimport { modelNames } from \"../core/metadata\";\nimport { type LabelSpec, resolveLabels } from \"../labels\";\nimport {\n    type Normalization,\n    isUltralyticsClassifier,\n    resolveNormalization,\n} from \"../normalization\";\nimport { softmax, topK } from \"../postprocess/classification\";\nimport { asFloat32Array } from \"../core/dtypes\";\nimport { toFloat32Tensor } from \"../preprocess/image\";\nimport { ResizePipeline, zeroTensorData } from \"../preprocess/pipeline\";\nimport { ClassificationResults, Probs } from \"../results\";\nimport { VisionTask } from \"./base\";\nimport { type ClassProbability, type ClassificationResult, type RGBImage } from \"../types\";\n\nexport interface ClassifierOptions extends OrtSessionOptions {\n    /**\n     * Class label spec — see {@link resolveLabels}.\n     *\n     * Optional: when omitted, the names the export baked into the model are used\n     * (Ultralytics writes them as `names` in the metadata map). Only when the\n     * model carries none does this fall back to generated `class_<id>` labels.\n     * Passing a spec always wins, for a model whose names are wrong or absent.\n     */\n    readonly labels?: LabelSpec;\n    /**\n     * Number of classes the model can predict.\n     *\n     * Optional: inferred from the classification head's declared output shape\n     * `(B, nc)`. Pass it to validate that the supplied labels match the model.\n     */\n    readonly numClasses?: number;\n    /**\n     * Model input `[width, height]` in pixels.\n     *\n     * Only used when the model's graph leaves its spatial axes dynamic: a graph\n     * that declares a static size always wins, since that is the only shape ONNX\n     * Runtime will accept. Defaults to `[224, 224]`.\n     */\n    readonly inputSize?: readonly [number, number];\n    /**\n     * Which preprocessing this model expects — see {@link Normalization}.\n     *\n     * Defaults to `\"auto\"`, which reads the model's own export metadata and picks\n     * `\"ultralytics\"` (raw `[0, 1]`) for an Ultralytics classification head and\n     * `\"imagenet\"` for everything else.\n     */\n    readonly normalization?: Normalization;\n    /** Per-channel RGB mean, overriding the preset. Defaults to the preset's. */\n    readonly mean?: readonly [number, number, number];\n    /** Per-channel RGB standard deviation, overriding the preset. */\n    readonly std?: readonly [number, number, number];\n    /**\n     * Whether the model's output still needs a softmax.\n     *\n     * Left undefined (the default), this reads the model's metadata and answers\n     * `false` for an Ultralytics classification export, whose graph already ends\n     * in one — applying a second softmax to a probability vector keeps the\n     * ranking but flattens the confidences, so the top-1 stays right while every\n     * number attached to it is wrong. Detection covers that family; for any other\n     * model that already emits probabilities, pass `false` explicitly.\n     */\n    readonly applySoftmax?: boolean;\n}\n\nexport interface ClassifierPredictOptions {\n    /**\n     * If set, the per-class probability list in `results[0].result.probabilities`\n     * is truncated to the top-K entries. The bulk `probs` view always exposes\n     * the full vector.\n     */\n    readonly topK?: number;\n}\n\n/**\n * Image classifier wrapping an ONNX model with ImageNet-style preprocessing.\n *\n * `predict()` returns `Promise<ClassificationResults[]>` (length 1 for a\n * single image), mirroring Ultralytics' API. The envelope exposes a `probs`\n * collection (`top1`, `top1conf`, `top5`, `top5conf`, `data`) plus the\n * legacy per-class probability list with names already resolved.\n *\n * Defaults: 224×224 RGB input, `float32` normalized with ImageNet mean/std,\n * NCHW layout, batch size 1, softmax applied to the raw output.\n *\n * @example\n * ```typescript\n * const clf = await Classifier.create(\"/models/resnet50.onnx\", {\n *   labels: [\"tench\", \"goldfish\", ...]  // 1000 ImageNet labels\n * });\n * const r = (await clf.predict(\"/images/dog.jpg\"))[0];\n * console.log(r.cls, r.conf, r.name);\n * console.log(r.probs.top5, r.probs.top5conf);\n * ```\n */\nexport class Classifier extends VisionTask {\n    private constructor(\n        session: OrtSession,\n        private readonly _labels: readonly string[],\n        private readonly _names: Readonly<Record<number, string>>,\n        private readonly _inputSize: readonly [number, number],\n        private readonly _mean: readonly [number, number, number],\n        private readonly _std: readonly [number, number, number],\n        private readonly _applySoftmax: boolean,\n        private readonly _normalization: string,\n    ) {\n        super(session);\n    }\n\n    /**\n     * Whether a softmax is applied to the model's output before ranking.\n     *\n     * Resolved once at construction. Worth reading when confidences look\n     * compressed: a second softmax over an already-normalized vector leaves the\n     * ordering intact and the numbers meaningless, which is invisible to any check\n     * that only looks at the predicted class.\n     */\n    get appliesSoftmax(): boolean {\n        return this._applySoftmax;\n    }\n\n    /**\n     * Which preprocessing this classifier applies to every image.\n     *\n     * One of the {@link Normalization} preset names, or `\"custom\"` when the caller\n     * supplied `mean`/`std` directly. Worth reading when a model underperforms:\n     * feeding a classifier a differently prepared tensor than it was trained on\n     * degrades it without throwing anything, so \"what does this assume\" is the\n     * first question.\n     */\n    get normalization(): string {\n        return this._normalization;\n    }\n\n    private _pipelineCache: ResizePipeline | null = null;\n\n    /**\n     * Run the model once on a zero tensor, paying one-time costs up front.\n     *\n     * The first inference of a session is not representative: WebGPU compiles its\n     * shaders on it and the WASM backend faults in its arenas. Calling this while\n     * a loading spinner is still up moves that cost somewhere the user is already\n     * waiting — which matters most for a classifier running as the last step of\n     * an on-device analysis, where the delay lands right before the answer shows.\n     *\n     * @param runs How many warm-up inferences to run. One is enough for WASM;\n     *   WebGPU sometimes settles on the second.\n     */\n    async warmup(runs: number = 1): Promise<void> {\n        const [tw, th] = this._inputSize;\n        for (let i = 0; i < runs; i++) {\n            const tensor = toFloat32Tensor(zeroTensorData(tw, th), [1, 3, th, tw]);\n            await this._session.run({ [this._session.inputName]: tensor });\n        }\n    }\n\n    /**\n     * The fused preprocessing pipeline, built on first use.\n     *\n     * Lazily, because constructing it reserves the output buffer: a task built in\n     * an environment without a canvas implementation stays constructible, and only\n     * fails if it is actually asked to preprocess something.\n     */\n    private get _pipeline(): ResizePipeline {\n        if (this._pipelineCache === null) {\n            const [tw, th] = this._inputSize;\n            this._pipelineCache = new ResizePipeline(tw, th, this._mean, this._std);\n        }\n        return this._pipelineCache;\n    }\n\n    /**\n     * Load the model, resolve labels, and settle the preprocessing.\n     *\n     * @param model The model source — a URL, or the bytes.\n     * @param options Labels, input size, normalization, and session options.\n     * @throws {RangeError} If `normalization` names an unknown preset, or names\n     *   one while `mean`/`std` are also given.\n     */\n    static async create(model: ModelSource, options: ClassifierOptions = {}): Promise<Classifier> {\n        const session = await OrtSession.create(model, options);\n        const normalization = resolveNormalization(session.metadata, {\n            normalization: options.normalization,\n            mean: options.mean,\n            std: options.std,\n        });\n        const numClasses =\n            options.numClasses ?? classificationNumClasses(session.outputShape) ?? undefined;\n        const labels = resolveLabels(options.labels ?? modelNames(session.metadata), {\n            numClasses,\n        });\n        const names: Record<number, string> = {};\n        for (let i = 0; i < labels.length; i++) {\n            names[i] = labels[i] as string;\n        }\n        return new Classifier(\n            session,\n            labels,\n            names,\n            resolveInputSize({\n                graphShape: session.inputShape,\n                requested: options.inputSize,\n                fallback: [224, 224],\n            }),\n            normalization.mean,\n            normalization.std,\n            options.applySoftmax ?? !isUltralyticsClassifier(session.metadata),\n            normalization.name,\n        );\n    }\n\n    /** Class labels indexed by class id. */\n    get labels(): readonly string[] {\n        return this._labels;\n    }\n\n    /** Class id → class name dict (matches Ultralytics' `model.names`). */\n    get names(): Readonly<Record<number, string>> {\n        return this._names;\n    }\n\n    /**\n     * The `[width, height]` this task preprocesses to.\n     *\n     * Resolved at creation time from the model's graph when it declares a static\n     * input, so reading it back tells you the resolution inference really runs at\n     * — not merely what was requested.\n     */\n    get inputSize(): readonly [number, number] {\n        return this._inputSize;\n    }\n\n    /** Number of classes the model can predict. */\n    get numClasses(): number {\n        return this._labels.length;\n    }\n\n    /** Alias for {@link predict} (parity with PyTorch `nn.Module.__call__`). */\n    async call(\n        image: ImageInput,\n        options: ClassifierPredictOptions = {},\n    ): Promise<ClassificationResults[]> {\n        return this.predict(image, options);\n    }\n\n    /** Run classification on a single image. */\n    async predict(\n        image: ImageInput,\n        options: ClassifierPredictOptions = {},\n    ): Promise<ClassificationResults[]> {\n        const timer = new SpeedTimer();\n        const path = typeof image === \"string\" ? image : null;\n        const { image: original, canvas } = await loadImageSource(image);\n        timer.stage(\"load\");\n        const { tensor, reused } = this._preprocess(original, canvas);\n        timer.stage(\"preprocess\");\n        let outputs: Record<string, ort.Tensor>;\n        try {\n            outputs = await this._session.run({ [this._session.inputName]: tensor });\n        } finally {\n            this._pipeline.release({ reused });\n        }\n        timer.stage(\"inference\");\n        const firstOutputName = this._session.outputNames[0];\n        if (firstOutputName === undefined) {\n            throw new Error(\"Classifier model has no outputs.\");\n        }\n        const raw = outputs[firstOutputName];\n        if (raw === undefined) {\n            throw new Error(\n                `Classifier model output ${firstOutputName} missing from run() result.`,\n            );\n        }\n        const fullProbs = this._postprocess(asFloat32Array(raw.data));\n\n        const { indices, values } = topK(fullProbs, options.topK ?? null);\n        const probabilities: ClassProbability[] = [];\n        for (let i = 0; i < indices.length; i++) {\n            const id = indices[i] as number;\n            const className = this._labels[id] ?? `class_${id}`;\n            probabilities.push({\n                classId: id,\n                className,\n                probability: values[i] as number,\n                cls: id,\n                name: className,\n                conf: values[i] as number,\n            });\n        }\n        if (probabilities.length === 0) {\n            throw new Error(\"Classifier produced no probabilities (empty output).\");\n        }\n\n        const top = probabilities[0] as ClassProbability;\n        const result: ClassificationResult = {\n            classId: top.classId,\n            className: top.className,\n            confidence: top.probability,\n            cls: top.classId,\n            name: top.className,\n            conf: top.probability,\n            image: original,\n            probabilities,\n        };\n\n        const orig: readonly [number, number] = [original.height, original.width];\n        const probs = new Probs(fullProbs);\n        timer.stage(\"postprocess\");\n        return [\n            new ClassificationResults(\n                probs,\n                result,\n                this._names,\n                original,\n                orig,\n                path,\n                timer.speed(),\n            ),\n        ];\n    }\n\n    /**\n     * Resize and pack the image into the tensor the model expects.\n     *\n     * @param image The decoded input.\n     * @param canvas The opaque canvas it was decoded on, when there is one — see\n     *   {@link ResizePipeline.run}.\n     */\n    private _preprocess(\n        image: RGBImage,\n        canvas: CanvasImageSource | null,\n    ): { tensor: ort.Tensor; reused: boolean } {\n        const [tw, th] = this._inputSize;\n        const { data, reused } = this._pipeline.run(image, canvas);\n        return { tensor: toFloat32Tensor(data, [1, 3, th, tw]), reused };\n    }\n\n    private _postprocess(raw: Float32Array): Float32Array {\n        return this._applySoftmax ? softmax(raw) : new Float32Array(raw);\n    }\n}\n"],"mappings":";;;;;;;;;;;;;;AA0GA,IAAa,IAAb,MAAa,UAAmB,EAAW;CAGlB;CACA;CACA;CACA;CACA;CACA;CACA;CARrB,YACI,GACA,GACA,GACA,GACA,GACA,GACA,GACA,GACF;EADmB,AAEjB,MAAM,CAAO,GARI,KAAA,UAAA,GACA,KAAA,SAAA,GACA,KAAA,aAAA,GACA,KAAA,QAAA,GACA,KAAA,OAAA,GACA,KAAA,gBAAA,GACA,KAAA,iBAAA;CAGrB;CAUA,IAAI,iBAA0B;EAC1B,OAAO,KAAK;CAChB;CAWA,IAAI,gBAAwB;EACxB,OAAO,KAAK;CAChB;CAEA,iBAAgD;CAchD,MAAM,OAAO,IAAe,GAAkB;EAC1C,IAAM,CAAC,GAAI,KAAM,KAAK;EACtB,KAAK,IAAI,IAAI,GAAG,IAAI,GAAM,KAAK;GAC3B,IAAM,IAAS,EAAgB,EAAe,GAAI,CAAE,GAAG;IAAC;IAAG;IAAG;IAAI;GAAE,CAAC;GACrE,MAAM,KAAK,SAAS,IAAI,GAAG,KAAK,SAAS,YAAY,EAAO,CAAC;EACjE;CACJ;CASA,IAAY,YAA4B;EACpC,IAAI,KAAK,mBAAmB,MAAM;GAC9B,IAAM,CAAC,GAAI,KAAM,KAAK;GACtB,KAAK,iBAAiB,IAAI,EAAe,GAAI,GAAI,KAAK,OAAO,KAAK,IAAI;EAC1E;EACA,OAAO,KAAK;CAChB;CAUA,aAAa,OAAO,GAAoB,IAA6B,CAAC,GAAwB;EAC1F,IAAM,IAAU,MAAM,EAAW,OAAO,GAAO,CAAO,GAChD,IAAgB,EAAqB,EAAQ,UAAU;GACzD,eAAe,EAAQ;GACvB,MAAM,EAAQ;GACd,KAAK,EAAQ;EACjB,CAAC,GACK,IACF,EAAQ,cAAc,EAAyB,EAAQ,WAAW,KAAK,KAAA,GACrE,IAAS,EAAc,EAAQ,UAAU,EAAW,EAAQ,QAAQ,GAAG,EACzE,cACJ,CAAC,GACK,IAAgC,CAAC;EACvC,KAAK,IAAI,IAAI,GAAG,IAAI,EAAO,QAAQ,KAC/B,EAAM,KAAK,EAAO;EAEtB,OAAO,IAAI,EACP,GACA,GACA,GACA,EAAiB;GACb,YAAY,EAAQ;GACpB,WAAW,EAAQ;GACnB,UAAU,CAAC,KAAK,GAAG;EACvB,CAAC,GACD,EAAc,MACd,EAAc,KACd,EAAQ,gBAAgB,CAAC,EAAwB,EAAQ,QAAQ,GACjE,EAAc,IAClB;CACJ;CAGA,IAAI,SAA4B;EAC5B,OAAO,KAAK;CAChB;CAGA,IAAI,QAA0C;EAC1C,OAAO,KAAK;CAChB;CASA,IAAI,YAAuC;EACvC,OAAO,KAAK;CAChB;CAGA,IAAI,aAAqB;EACrB,OAAO,KAAK,QAAQ;CACxB;CAGA,MAAM,KACF,GACA,IAAoC,CAAC,GACL;EAChC,OAAO,KAAK,QAAQ,GAAO,CAAO;CACtC;CAGA,MAAM,QACF,GACA,IAAoC,CAAC,GACL;EAChC,IAAM,IAAQ,IAAI,EAAW,GACvB,IAAO,OAAO,KAAU,WAAW,IAAQ,MAC3C,EAAE,OAAO,GAAU,cAAW,MAAM,EAAgB,CAAK;EAC/D,EAAM,MAAM,MAAM;EAClB,IAAM,EAAE,WAAQ,cAAW,KAAK,YAAY,GAAU,CAAM;EAC5D,EAAM,MAAM,YAAY;EACxB,IAAI;EACJ,IAAI;GACA,IAAU,MAAM,KAAK,SAAS,IAAI,GAAG,KAAK,SAAS,YAAY,EAAO,CAAC;EAC3E,UAAU;GACN,KAAK,UAAU,QAAQ,EAAE,UAAO,CAAC;EACrC;EACA,EAAM,MAAM,WAAW;EACvB,IAAM,IAAkB,KAAK,SAAS,YAAY;EAClD,IAAI,MAAoB,KAAA,GACpB,MAAU,MAAM,kCAAkC;EAEtD,IAAM,IAAM,EAAQ;EACpB,IAAI,MAAQ,KAAA,GACR,MAAU,MACN,2BAA2B,EAAgB,4BAC/C;EAEJ,IAAM,IAAY,KAAK,aAAa,EAAe,EAAI,IAAI,CAAC,GAEtD,EAAE,YAAS,cAAW,EAAK,GAAW,EAAQ,QAAQ,IAAI,GAC1D,IAAoC,CAAC;EAC3C,KAAK,IAAI,IAAI,GAAG,IAAI,EAAQ,QAAQ,KAAK;GACrC,IAAM,IAAK,EAAQ,IACb,IAAY,KAAK,QAAQ,MAAO,SAAS;GAC/C,EAAc,KAAK;IACf,SAAS;IACT;IACA,aAAa,EAAO;IACpB,KAAK;IACL,MAAM;IACN,MAAM,EAAO;GACjB,CAAC;EACL;EACA,IAAI,EAAc,WAAW,GACzB,MAAU,MAAM,sDAAsD;EAG1E,IAAM,IAAM,EAAc,IACpB,IAA+B;GACjC,SAAS,EAAI;GACb,WAAW,EAAI;GACf,YAAY,EAAI;GAChB,KAAK,EAAI;GACT,MAAM,EAAI;GACV,MAAM,EAAI;GACV,OAAO;GACP;EACJ,GAEM,IAAkC,CAAC,EAAS,QAAQ,EAAS,KAAK,GAClE,IAAQ,IAAI,EAAM,CAAS;EAEjC,OADA,EAAM,MAAM,aAAa,GAClB,CACH,IAAI,EACA,GACA,GACA,KAAK,QACL,GACA,GACA,GACA,EAAM,MAAM,CAChB,CACJ;CACJ;CASA,YACI,GACA,GACuC;EACvC,IAAM,CAAC,GAAI,KAAM,KAAK,YAChB,EAAE,SAAM,cAAW,KAAK,UAAU,IAAI,GAAO,CAAM;EACzD,OAAO;GAAE,QAAQ,EAAgB,GAAM;IAAC;IAAG;IAAG;IAAI;GAAE,CAAC;GAAG;EAAO;CACnE;CAEA,aAAqB,GAAiC;EAClD,OAAO,KAAK,gBAAgB,EAAQ,CAAG,IAAI,IAAI,aAAa,CAAG;CACnE;AACJ"}

package/dist/vision/tasks/detectClassify.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/session.cjs"),o=require("../io/image.cjs"),s=require("../fusion.cjs"),c=require("../preprocess/image.cjs"),l=require("../preprocess/pipeline.cjs"),u=require("../postprocess/classification.cjs"),d=require("./base.cjs");var f=class f extends d.VisionTask{_spec;_labels;_names;_classifierLabels;_classifierNames;_raiseOnEmpty;constructor(e,t,n,r,i,a,o){super(e),this._spec=t,this._labels=n,this._names=r,this._classifierLabels=i,this._classifierNames=a,this._raiseOnEmpty=o}_pipelineCache=null;async warmup(e=1){let[t,n]=this._spec.inputSize;for(let r=0;r<e;r++){let e={[s.INPUT_IMAGE]:c.toFloat32Tensor(l.zeroTensorData(t,n),[1,3,n,t])};this._spec.needsSourceImage&&(e[s.INPUT_SOURCE]=c.toFloat32Tensor(l.zeroTensorData(t,n),[1,3,n,t]),e[s.INPUT_SCALE]=c.toFloat32Tensor(new Float32Array([1]),[1]),e[s.INPUT_PAD]=c.toFloat32Tensor(new Float32Array([0,0]),[2])),await this._session.run(e)}}get _pipeline(){if(this._pipelineCache===null){let[e,t]=this._spec.inputSize;this._pipelineCache=new l.LetterboxPipeline(e,t)}return this._pipelineCache}static async create(t,n={}){let r=await a.OrtSession.create(t,n),o=s.readFusionSpec(r.metadata);if(o===null)throw new e.FusionError(`This model carries no fused-pipeline metadata, so DetectClassify cannot tell how to drive it. Build one with ort_vision_sdk.compose.fuse_detect_classify, or load a plain model with Detector/Classifier instead.`);let c=i.resolveLabels(n.labels??o.detectorNames??`coco`),l=i.resolveLabels(n.classifierLabels??o.classifierNames,{numClasses:p(r)??void 0});return new f(r,o,c,m(c),l,m(l),n.raiseOnEmpty??!1)}get spec(){return this._spec}get inputSize(){return this._spec.inputSize}get labels(){return this._labels}get names(){return this._names}get classifierLabels(){return this._classifierLabels}get classifierNames(){return this._classifierNames}async call(e,t={}){return this.predict(e,t)}async predict(e,t={}){let i=new n.SpeedTimer,a=typeof e==`string`?e:null,c=await o.loadImage(e);i.stage(`load`);let{feeds:l,scale:u,padLeft:f,padTop:p}=this._preprocess(c);i.stage(`preprocess`);let m=await this._session.run(l);this._pipeline.release(),i.stage(`inference`);let v=h(m,s.OUTPUT_PROBS),S=g(m,s.OUTPUT_BOXES),C=g(m,s.OUTPUT_SCORES),w=_(m,s.OUTPUT_CLASSES),T=v.data,E=_(m,`num_detections`)[0]??0,D=Math.min(E,Math.floor(S.length/4)),O=v.dims[v.dims.length-1]??0,k=t.classes===void 0?null:new Set(t.classes),A=t.confThreshold??0,j=[];for(let e=0;e<D;e++){let n=w[e]??0,r=C[e]??0;if(r<A||k!==null&&!k.has(n))continue;let i=this._toOriginal(S,e,{scale:u,padLeft:f,padTop:p,original:c}),a=y(c,i);j.push(b(n,this._names[n]??`class_${n}`,r,i,a,this._classify(T.subarray(e*O,(e+1)*O),a,t.topK)))}d.requireDetections(j.length,{raiseOnEmpty:t.raiseOnEmpty??this._raiseOnEmpty,confThreshold:Math.max(A,this._spec.confThreshold),classes:t.classes,path:a});let M=[c.height,c.width];return i.stage(`postprocess`),[new r.DetectClassifyResults(x(j,M),j,this._names,this._classifierNames,c,M,a,i.speed())]}_preprocess(e){let[t,n]=this._spec.inputSize,r=this._pipeline.run(e),i={[s.INPUT_IMAGE]:c.toFloat32Tensor(r.data,[1,3,n,t])};return this._spec.needsSourceImage&&(i[s.INPUT_SOURCE]=v(e),i[s.INPUT_SCALE]=c.toFloat32Tensor(new Float32Array([r.scale]),[1]),i[s.INPUT_PAD]=c.toFloat32Tensor(new Float32Array([r.padLeft,r.padTop]),[2])),{feeds:i,scale:r.scale,padLeft:r.padLeft,padTop:r.padTop}}_toOriginal(e,n,r){let{scale:i,padLeft:a,padTop:o,original:s}=r,c=t=>e[n*4+t]??0,l=e=>Math.min(Math.max(e,0),s.width),u=e=>Math.min(Math.max(e,0),s.height);return new t.BoundingBox(l((c(0)-a)/i),u((c(1)-o)/i),l((c(2)-a)/i),u((c(3)-o)/i))}_classify(e,t,n){let r=this._spec.applySoftmax?u.softmax(e):e,{indices:i,values:a}=u.topK(r,n??null),o=[];for(let e=0;e<i.length;e++){let t=i[e]??0,n=a[e]??0,r=this._classifierLabels[t]??`class_${t}`;o.push({classId:t,className:r,probability:n,cls:t,name:r,conf:n})}let s=o[0]??{classId:0,className:`class_0`,probability:0,cls:0,name:`class_0`,conf:0};return{classId:s.classId,className:s.className,confidence:s.probability,cls:s.classId,name:s.className,conf:s.probability,image:t,probabilities:o}}};function p(e){let t=e.outputNames.indexOf(s.OUTPUT_PROBS);if(t<0)return null;let n=e.outputShapes[t];return n===void 0||n.length===0?null:n[n.length-1]??null}function m(e){let t={};for(let n=0;n<e.length;n++)t[n]=e[n];return t}function h(t,n){let r=t[n];if(r===void 0)throw new e.FusionError(`The fused pipeline is missing its '${n}' output.`);return r}function g(e,t){return h(e,t).data}function _(e,t){let n=h(e,t).data,r=[];for(let e=0;e<n.length;e++)r.push(Number(n[e]));return r}function v(e){let t=c.toCHW(c.toFloat32(e),e.width,e.height,3);return c.toFloat32Tensor(t,[1,3,e.height,e.width])}function y(e,n){let[r,i,a,o]=n.asIntXyxy(),s=Math.max(0,r),c=Math.max(0,i),l=Math.min(e.width,a),u=Math.min(e.height,o);if(l<=s||u<=c)return new t.RGBImage(new Uint8Array,0,0);let d=l-s,f=u-c,p=new Uint8Array(d*f*3);for(let t=0;t<f;t++){let n=((c+t)*e.width+s)*3;p.set(e.data.subarray(n,n+d*3),t*d*3)}return new t.RGBImage(p,d,f)}function b(e,t,n,r,i,a){return{classId:e,className:t,confidence:n,bbox:r,cls:e,name:t,conf:n,box:r,croppedImage:i,classification:a}}function x(e,t){let n=new Float32Array(e.length*4),i=new Int32Array(e.length),a=new Float32Array(e.length);return e.forEach((e,t)=>{let[r,o,s,c]=e.bbox.xyxy;n[t*4]=r,n[t*4+1]=o,n[t*4+2]=s,n[t*4+3]=c,i[t]=e.classId,a[t]=e.confidence}),new r.Boxes(n,i,a,t)}exports.DetectClassify=f;
+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/dtypes.cjs"),o=require("../core/session.cjs"),s=require("../io/image.cjs"),c=require("../fusion.cjs"),l=require("../preprocess/image.cjs"),u=require("../preprocess/pipeline.cjs"),d=require("../postprocess/classification.cjs"),f=require("./base.cjs");var p=class p extends f.VisionTask{_spec;_labels;_names;_classifierLabels;_classifierNames;_raiseOnEmpty;constructor(e,t,n,r,i,a,o){super(e),this._spec=t,this._labels=n,this._names=r,this._classifierLabels=i,this._classifierNames=a,this._raiseOnEmpty=o}_pipelineCache=null;async warmup(e=1){let[t,n]=this._spec.inputSize;for(let r=0;r<e;r++){let e={[c.INPUT_IMAGE]:l.toFloat32Tensor(u.zeroTensorData(t,n),[1,3,n,t])};this._spec.needsSourceImage&&(e[c.INPUT_SOURCE]=l.toFloat32Tensor(u.zeroTensorData(t,n),[1,3,n,t]),e[c.INPUT_SCALE]=l.toFloat32Tensor(new Float32Array([1]),[1]),e[c.INPUT_PAD]=l.toFloat32Tensor(new Float32Array([0,0]),[2])),await this._session.run(e)}}get _pipeline(){if(this._pipelineCache===null){let[e,t]=this._spec.inputSize;this._pipelineCache=new u.LetterboxPipeline(e,t)}return this._pipelineCache}static async create(t,n={}){let r=await o.OrtSession.create(t,n),a=c.readFusionSpec(r.metadata);if(a===null)throw new e.FusionError(`This model carries no fused-pipeline metadata, so DetectClassify cannot tell how to drive it. Build one with ort_vision_sdk.compose.fuse_detect_classify, or load a plain model with Detector/Classifier instead.`);let s=i.resolveLabels(n.labels??a.detectorNames??`coco`),l=i.resolveLabels(n.classifierLabels??a.classifierNames,{numClasses:m(r)??void 0});return new p(r,a,s,h(s),l,h(l),n.raiseOnEmpty??!1)}get spec(){return this._spec}get inputSize(){return this._spec.inputSize}get labels(){return this._labels}get names(){return this._names}get classifierLabels(){return this._classifierLabels}get classifierNames(){return this._classifierNames}async call(e,t={}){return this.predict(e,t)}async predict(e,t={}){let i=new n.SpeedTimer,o=typeof e==`string`?e:null,{image:l,canvas:u}=await s.loadImageSource(e);i.stage(`load`);let{feeds:d,scale:p,padLeft:m,padTop:h,reused:y}=this._preprocess(l,u);i.stage(`preprocess`);let C;try{C=await this._session.run(d)}finally{this._pipeline.release({reused:y})}i.stage(`inference`);let w=g(C,c.OUTPUT_PROBS),T=_(C,c.OUTPUT_BOXES),E=_(C,c.OUTPUT_SCORES),D=v(C,c.OUTPUT_CLASSES),O=a.asFloat32Array(w.data),k=v(C,`num_detections`)[0]??0,A=this._spec.hasMasks?_(C,c.OUTPUT_MASKS):null,[j,M]=this._spec.cropSize,N=Math.min(k,Math.floor(T.length/4)),P=w.dims[w.dims.length-1]??0,F=t.classes===void 0?null:new Set(t.classes),I=t.confThreshold??0,L=[];for(let e=0;e<N;e++){let n=D[e]??0,r=E[e]??0;if(r<I||F!==null&&!F.has(n))continue;let i=this._toOriginal(T,e,{scale:p,padLeft:m,padTop:h,original:l}),a=f.cropToBox(l,i);L.push(b(n,this._names[n]??`class_${n}`,r,i,a,this._classify(O.subarray(e*P,(e+1)*P),a,t.topK),A===null?null:x(A,e,j,M,a)))}f.requireDetections(L.length,{raiseOnEmpty:t.raiseOnEmpty??this._raiseOnEmpty,confThreshold:Math.max(I,this._spec.confThreshold),classes:t.classes,path:o});let R=[l.height,l.width];return i.stage(`postprocess`),[new r.DetectClassifyResults(S(L,R),L,this._names,this._classifierNames,l,R,o,i.speed())]}_preprocess(e,t){let[n,r]=this._spec.inputSize,i=this._pipeline.run(e,t),a={[c.INPUT_IMAGE]:l.toFloat32Tensor(i.data,[1,3,r,n])};return this._spec.needsSourceImage&&(a[c.INPUT_SOURCE]=y(e),a[c.INPUT_SCALE]=l.toFloat32Tensor(new Float32Array([i.scale]),[1]),a[c.INPUT_PAD]=l.toFloat32Tensor(new Float32Array([i.padLeft,i.padTop]),[2])),{feeds:a,scale:i.scale,padLeft:i.padLeft,padTop:i.padTop,reused:i.reused}}_toOriginal(e,n,r){let{scale:i,padLeft:a,padTop:o,original:s}=r,c=t=>e[n*4+t]??0,l=e=>Math.min(Math.max(e,0),s.width),u=e=>Math.min(Math.max(e,0),s.height);return new t.BoundingBox(l((c(0)-a)/i),u((c(1)-o)/i),l((c(2)-a)/i),u((c(3)-o)/i))}_classify(e,t,n){let r=this._spec.applySoftmax?d.softmax(e):e,{indices:i,values:a}=d.topK(r,n??null),o=[];for(let e=0;e<i.length;e++){let t=i[e]??0,n=a[e]??0,r=this._classifierLabels[t]??`class_${t}`;o.push({classId:t,className:r,probability:n,cls:t,name:r,conf:n})}let s=o[0]??{classId:0,className:`class_0`,probability:0,cls:0,name:`class_0`,conf:0};return{classId:s.classId,className:s.className,confidence:s.probability,cls:s.classId,name:s.className,conf:s.probability,image:t,probabilities:o}}};function m(e){let t=e.outputNames.indexOf(c.OUTPUT_PROBS);if(t<0)return null;let n=e.outputShapes[t];return n===void 0||n.length===0?null:n[n.length-1]??null}function h(e){let t={};for(let n=0;n<e.length;n++)t[n]=e[n];return t}function g(t,n){let r=t[n];if(r===void 0)throw new e.FusionError(`The fused pipeline is missing its '${n}' output.`);return r}function _(e,t){return a.asFloat32Array(g(e,t).data)}function v(e,t){let n=g(e,t).data,r=[];for(let e=0;e<n.length;e++)r.push(Number(n[e]));return r}function y(e){let t=l.toCHW(l.toFloat32(e),e.width,e.height,3);return l.toFloat32Tensor(t,[1,3,e.height,e.width])}function b(e,t,n,r,i,a,o){return{classId:e,className:t,confidence:n,bbox:r,cls:e,name:t,conf:n,box:r,croppedImage:i,classification:a,mask:o}}function x(e,n,r,i,a){let{width:o,height:s}=a;if(o<1||s<1)return null;let c=r*i,l=e.subarray(n*c,(n+1)*c),u=new Uint8Array(o*s);for(let e=0;e<s;e++){let t=Math.min(Math.floor(e*i/s),i-1);for(let n=0;n<o;n++){let i=Math.min(Math.floor(n*r/o),r-1);u[e*o+n]=(l[t*r+i]??0)>.5?255:0}}return new t.Mask(u,o,s)}function S(e,t){let n=new Float32Array(e.length*4),i=new Int32Array(e.length),a=new Float32Array(e.length);return e.forEach((e,t)=>{let[r,o,s,c]=e.bbox.xyxy;n[t*4]=r,n[t*4+1]=o,n[t*4+2]=s,n[t*4+3]=c,i[t]=e.classId,a[t]=e.confidence}),new r.Boxes(n,i,a,t)}exports.DetectClassify=p,exports.maskToBox=x;
 //# sourceMappingURL=detectClassify.cjs.map