BadBob / Playground
Browser · ONNX · WASM · MIT
05.01

準備專案

安裝列出嘅相依套件,再執行測試。開啟相機之前,先試一張本機圖片。

基礎說明

npm ci · npm run build
設定相依套件並編譯專案。
npm test · npm run check
執行自動檢查,唔會開啟鏡頭。
Terminalbash
unzip badbob-playground-javascript.zip
cd badbob-playground-javascript
npm ci
npm test
npm run check
npm run build
05.02

連接模型

開啟鏡頭之前,先驗證模型同類別標籤。

基礎說明

parseLabels(text)
按模型嘅類別次序讀取名稱;拒絕空白或重複標籤。
OnnxDetector.create(model, labelsText)
載入 ONNX 模型,並驗證輸入、輸出同標籤數量。
OnnxDetector.detect(frame)
預處理影像,執行模型並傳回偵測結果。
要留意
症狀
偵測框移位,或者模型載入失敗。
原因
前處理、輸出格式或者標籤次序唔配合。
處理方法
呢個介接器只支援以下約定。Core ML 檔案同其他 ONNX 輸出格式需要另一個介接器。

呢個介接器只支援以下約定。Core ML 檔案同其他 ONNX 輸出格式需要另一個介接器。

深入了解

座標表示偵測框嘅中心點同尺寸,單位係 640×640 輸入影像嘅像素。標籤必須嚴格跟返模型嘅類別次序排列。

ONNXyaml
input:
  color: RGB
  dtype: float32
  shape: [1, 3, 640, 640]
  letterbox: 114
  scale: 1/255
output:
  dtype: float32
  shape: [1, 4+C, N]
  boxes: [cx, cy, width, height]
  class_scores: probabilities
  objectness: false
  embedded_nms: false
onnx-detector.tstypescript
import * as ort from "onnxruntime-web/wasm";
import {
    decode,
    INPUT_EDGE,
    letterbox,
    parseLabels,
    rgbaToNchw,
    validateOutputShape,
} from "./core.ts";
import type { Detection, Detector, Letterbox } from "./types.ts";

export function validateModelMetadata(
    inputs: readonly ort.InferenceSession.ValueMetadata[],
    outputs: readonly ort.InferenceSession.ValueMetadata[],
    classes: number,
): void {
    const input = inputs[0];
    const output = outputs[0];

    if (
        inputs.length !== 1 ||
        !input?.isTensor ||
        input.type !== "float32" ||
        input.shape.length !== 4 ||
        input.shape.some((dimension, index) => dimension !== [1, 3, 640, 640][index])
    ) {
        throw new Error("The model must have one static float32 RGB input [1,3,640,640].");
    }

    if (
        outputs.length !== 1 ||
        !output?.isTensor ||
        output.type !== "float32" ||
        output.shape.length !== 3 ||
        output.shape[0] !== 1 ||
        output.shape[1] !== 4 + classes
    ) {
        throw new Error(
            "The model must have one float32 output [1,4+C,N], with C matching labels.",
        );
    }

    const count = output.shape[2];

    if (typeof count === "number") {
        validateOutputShape([1, 4 + classes, count], classes);
    }
}

export class OnnxDetector implements Detector {
    private disposed = false;

    private constructor(
        private readonly session: ort.InferenceSession,
        private readonly classes: number,
    ) {}

    static async create(model: ArrayBuffer, labelsText: string): Promise<OnnxDetector> {
        const labels = parseLabels(labelsText);
        ort.env.wasm.numThreads = 1;
        ort.env.wasm.proxy = false;
        const session = await ort.InferenceSession.create(new Uint8Array(model), {
            executionProviders: ["wasm"],
        });
        const detector = new OnnxDetector(session, labels.length);

        try {
            validateModelMetadata(session.inputMetadata, session.outputMetadata, labels.length);
            const neutral = new Float32Array(3 * INPUT_EDGE * INPUT_EDGE).fill(114 / 255);
            await detector.predict(neutral, letterbox({ width: 640, height: 640 }));
            return detector;
        } catch (error: unknown) {
            await detector.dispose();
            throw error;
        }
    }

    async detect(frame: ImageBitmap): Promise<Detection[]> {
        const transform = letterbox(frame);
        const canvas = new OffscreenCanvas(INPUT_EDGE, INPUT_EDGE);
        const context = canvas.getContext("2d", { willReadFrequently: true });

        if (!context) {
            throw new Error("OffscreenCanvas 2D is unavailable in this browser's worker.");
        }

        context.fillStyle = "rgb(114, 114, 114)";
        context.fillRect(0, 0, INPUT_EDGE, INPUT_EDGE);
        context.drawImage(
            frame,
            transform.left,
            transform.top,
            transform.resized.width,
            transform.resized.height,
        );
        const pixels = context.getImageData(0, 0, INPUT_EDGE, INPUT_EDGE);
        return this.predict(rgbaToNchw(pixels.data), transform);
    }

    async predict(input: Float32Array, transform: Letterbox): Promise<Detection[]> {
        if (this.disposed || input.length !== 3 * INPUT_EDGE * INPUT_EDGE) {
            throw new Error("Session is disposed or input is not [1,3,640,640].");
        }

        if (input.some((value) => !Number.isFinite(value) || value < 0 || value > 1)) {
            throw new Error("Input pixels must be finite normalized RGB values in [0,1].");
        }

        const inputName = this.session.inputNames[0];
        const outputName = this.session.outputNames[0];

        if (!inputName || !outputName) {
            throw new Error("Model tensor names are missing.");
        }

        const tensor = new ort.Tensor("float32", input, [1, 3, INPUT_EDGE, INPUT_EDGE]);
        let outputs: ort.InferenceSession.ReturnType | undefined;

        try {
            outputs = await this.session.run({ [inputName]: tensor });
            const output = outputs[outputName];

            if (
                !(output instanceof ort.Tensor) ||
                output.type !== "float32" ||
                !(output.data instanceof Float32Array)
            ) {
                throw new Error("Expected one CPU float32 output tensor.");
            }

            return decode(output.data, output.dims, this.classes, transform);
        } finally {
            tensor.dispose();

            for (const output of Object.values(outputs ?? {})) {
                if (output instanceof ort.Tensor) {
                    output.dispose();
                }
            }
        }
    }

    async dispose(): Promise<void> {
        if (this.disposed) {
            return;
        }

        this.disposed = true;
        await this.session.release();
    }
}
05.03

接入相機

瀏覽器負責擷取畫面,背景執行緒負責推論。每次只處理一格畫面,唔會令等候處理嘅佇列越積越多。

基礎說明

CameraSource.open(signal)
要求視訊權限,並喺明確啟動之後開啟相機。
CameraSource.next()
擷取下一格畫面,或者回報相機串流已經停止。
WorkerDetector.infer(frame)
預處理影像,執行模型並傳回偵測結果。
WorkerDetector.dispose()
停止接收新結果,等進行中嘅工作完成之後釋放資源。
要留意
症狀
預覽愈來愈慢,記憶體用量不斷增加。
原因
相機產生影像快過推論處理嘅速度,令冇上限嘅佇列不斷累積。
處理方法
瀏覽器負責擷取畫面,背景執行緒負責推論。每次只處理一格畫面,唔會令等候處理嘅佇列越積越多。
深入了解

要求停止、捨棄遲到嘅結果,等進行中嘅工作完成之後先釋放資源。可唔可以取消底層執行,要睇平台支唔支援。

frame-source.tstypescript
import { validateSize } from "./core.ts";

export interface FrameSource {
    next(): Promise<ImageBitmap | null>;
    stop(): void;
}

export class ImageSource implements FrameSource {
    private constructor(private image: ImageBitmap | undefined) {}

    static async open(file: File): Promise<ImageSource> {
        const image = await createImageBitmap(file);

        try {
            validateSize(image);
            return new ImageSource(image);
        } catch (error: unknown) {
            image.close();
            throw error;
        }
    }

    async next(): Promise<ImageBitmap | null> {
        const image = this.image;
        this.image = undefined;
        return image ?? null;
    }

    stop(): void {
        this.image?.close();
        this.image = undefined;
    }
}

export class CameraSource implements FrameSource {
    private stopped = false;

    private constructor(
        private readonly stream: MediaStream,
        private readonly video: HTMLVideoElement,
    ) {}

    static async open(signal: AbortSignal): Promise<CameraSource> {
        if (!navigator.mediaDevices?.getUserMedia) {
            throw new Error(
                "Camera API unavailable. Use HTTPS or localhost in a supported browser.",
            );
        }

        const stream = await navigator.mediaDevices.getUserMedia({
            video: { width: { ideal: 1280 }, height: { ideal: 720 } },
            audio: false,
        });
        const video = document.createElement("video");
        const source = new CameraSource(stream, video);

        if (signal.aborted) {
            source.stop();
            throw new Error("Camera startup stopped.");
        }

        const abort = (): void => source.stop();
        signal.addEventListener("abort", abort, { once: true });
        video.muted = true;
        video.playsInline = true;
        video.srcObject = stream;

        try {
            await video.play();

            if (signal.aborted || source.stopped) {
                throw new Error("Camera startup stopped.");
            }

            validateSize({ width: video.videoWidth, height: video.videoHeight });
            return source;
        } catch (error: unknown) {
            source.stop();
            throw error;
        } finally {
            signal.removeEventListener("abort", abort);
        }
    }

    async next(): Promise<ImageBitmap | null> {
        if (this.stopped) {
            return null;
        }

        if (!this.stream.getVideoTracks().some((track) => track.readyState === "live")) {
            throw new Error("Camera stream ended. Stop, check the device and start again.");
        }

        return createImageBitmap(this.video);
    }

    stop(): void {
        if (this.stopped) {
            return;
        }

        this.stopped = true;

        for (const track of this.stream.getTracks()) {
            track.stop();
        }

        this.video.pause();
        this.video.srcObject = null;
        this.video.removeAttribute("src");
        this.video.load();
    }
}
worker-client.tstypescript
import { isWorkerResponse } from "./protocol.ts";
import type { WorkerRequest, WorkerResponse } from "./protocol.ts";
import type { Detection } from "./types.ts";

export interface PairedResult {
    frame: ImageBitmap;
    detections: Detection[];
}

interface PendingFrame {
    id: number;
    resolve: (result: PairedResult) => void;
    reject: (error: Error) => void;
}

let nextGeneration = 1;

export class WorkerDetector {
    readonly ready: Promise<void>;
    private readonly worker = new Worker(new URL("./worker.ts", import.meta.url), {
        type: "module",
    });
    private readonly generation = nextGeneration++;
    private nextId = 1;
    private pending: PendingFrame | undefined;
    private stopping = false;
    private resolveReady!: () => void;
    private rejectReady!: (error: Error) => void;
    private readonly closed: Promise<void>;
    private resolveClosed!: () => void;
    private shutdownTimer: ReturnType<typeof setTimeout> | undefined;

    constructor(model: ArrayBuffer, labels: string) {
        this.ready = new Promise((resolve, reject) => {
            this.resolveReady = resolve;
            this.rejectReady = reject;
        });
        this.closed = new Promise((resolve) => {
            this.resolveClosed = resolve;
        });
        this.worker.addEventListener("message", (event: MessageEvent<unknown>) => {
            if (!isWorkerResponse(event.data)) {
                this.fail(new Error("Worker sent an invalid response."));
                return;
            }

            this.receive(event.data);
        });
        this.worker.addEventListener("error", (event: ErrorEvent) => {
            event.preventDefault();
            this.fail(new Error(event.message || "Inference worker failed."));
        });
        this.worker.addEventListener("messageerror", () =>
            this.fail(new Error("Worker message could not be decoded.")),
        );
        try {
            this.send({ kind: "initialize", generation: this.generation, model, labels }, [model]);
        } catch (error: unknown) {
            this.fail(error instanceof Error ? error : new Error(String(error)));
        }
    }

    infer(frame: ImageBitmap): Promise<PairedResult> {
        if (this.stopping || this.pending) {
            frame.close();
            return Promise.reject(new Error("Worker is stopping or already processing a frame."));
        }

        return new Promise((resolve, reject) => {
            const id = this.nextId++;
            this.pending = { id, resolve, reject };

            try {
                this.send({ kind: "infer", generation: this.generation, id, frame }, [frame]);
            } catch (error: unknown) {
                frame.close();
                this.pending = undefined;
                reject(error);
            }
        });
    }

    async dispose(): Promise<void> {
        if (this.stopping) {
            return this.closed;
        }

        this.stopping = true;
        const error = new Error("Inference stopped.");
        this.rejectReady(error);
        this.pending?.reject(error);
        this.pending = undefined;
        this.shutdownTimer = setTimeout(() => this.finishShutdown(), 5000);
        this.send({ kind: "stop", generation: this.generation });
        await this.closed;
    }

    private receive(message: WorkerResponse): void {
        if (message.generation !== this.generation) {
            if (message.kind === "result") {
                message.frame.close();
            }

            return;
        }

        if (message.kind === "disposed") {
            this.finishShutdown();
            return;
        }

        if (this.stopping) {
            if (message.kind === "result") {
                message.frame.close();
            }

            return;
        }

        if (message.kind === "ready") {
            this.resolveReady();
            return;
        }

        if (message.kind === "error") {
            const error = new Error(message.message);

            if (message.id === null) {
                this.rejectReady(error);
                return;
            }

            if (message.id === this.pending?.id) {
                this.pending.reject(error);
                this.pending = undefined;
            }

            return;
        }

        if (message.id !== this.pending?.id) {
            message.frame.close();
            return;
        }

        this.pending.resolve({ frame: message.frame, detections: message.detections });
        this.pending = undefined;
    }

    private send(message: WorkerRequest, transfer: Transferable[] = []): void {
        this.worker.postMessage(message, transfer);
    }

    private fail(error: Error): void {
        this.stopping = true;
        this.rejectReady(error);
        this.pending?.reject(error);
        this.pending = undefined;
        this.finishShutdown();
    }

    private finishShutdown(): void {
        clearTimeout(this.shutdownTimer);
        this.worker.terminate();
        this.resolveClosed();
    }
}
05.04

放好預測標記

還原 letterbox 縮放同填補,再將偵測框裁切到影像範圍內先繪畫。

基礎說明

letterbox(size) · unproject(...)
保持影像長寬比,將偵測框映射返原始影像。
decode(data, dims, classes, transform)
驗證預測結果,移除同一類別入面重疊嘅偵測框。
要留意

影像尺寸、裁切、方向同座標要有同一套約定。先用已知偵測框測試,再信任畫面標記。

core.tstypescript
import type { Box, Detection, Letterbox, Size } from "./types.ts";

export const INPUT_EDGE = 640;
const MAX_CANDIDATES = 100_000;

export function parseLabels(text: string): string[] {
    const normalized = text
        .replace(/^\uFEFF/, "")
        .replace(/\r\n/g, "\n")
        .replace(/\n$/, "");
    const labels = normalized.split("\n").map((line) => line.trim());

    if (labels.length > 1024 || labels.some((label) => label.length === 0 || label.length > 256)) {
        throw new Error(
            "Supply 1–1024 nonempty labels, one per line (maximum 256 characters each).",
        );
    }

    if (new Set(labels).size !== labels.length) {
        throw new Error("Labels must be distinct and in the model's exact class order.");
    }

    return labels;
}

export function validateSize(size: Size): void {
    if (
        !Number.isSafeInteger(size.width) ||
        !Number.isSafeInteger(size.height) ||
        size.width <= 0 ||
        size.height <= 0 ||
        size.width > 16384 ||
        size.height > 16384 ||
        size.width * size.height > 64_000_000
    ) {
        throw new Error(
            "Use an image with positive integer dimensions, at most 16384 per side and 64 MP.",
        );
    }
}

export function letterbox(size: Size): Letterbox {
    validateSize(size);
    const ratio = Math.min(INPUT_EDGE / size.width, INPUT_EDGE / size.height);
    const resized = {
        width: Math.max(1, Math.round(size.width * ratio)),
        height: Math.max(1, Math.round(size.height * ratio)),
    };

    return {
        original: { width: size.width, height: size.height },
        resized,
        left: Math.floor((INPUT_EDGE - resized.width) / 2),
        top: Math.floor((INPUT_EDGE - resized.height) / 2),
        edge: INPUT_EDGE,
    };
}

function validateTransform(transform: Letterbox): void {
    const expected = letterbox(transform.original);

    if (
        transform.edge !== expected.edge ||
        transform.left !== expected.left ||
        transform.top !== expected.top ||
        transform.resized.width !== expected.resized.width ||
        transform.resized.height !== expected.resized.height
    ) {
        throw new Error("Invalid letterbox geometry.");
    }
}

export function rgbaToNchw(rgba: Uint8ClampedArray): Float32Array {
    const pixels = INPUT_EDGE * INPUT_EDGE;

    if (rgba.length !== pixels * 4) {
        throw new Error("Expected a letterboxed 640×640 RGBA image.");
    }

    const result = new Float32Array(pixels * 3);

    for (let pixel = 0; pixel < pixels; pixel += 1) {
        result[pixel] = rgba[pixel * 4]! / 255;
        result[pixels + pixel] = rgba[pixel * 4 + 1]! / 255;
        result[pixels * 2 + pixel] = rgba[pixel * 4 + 2]! / 255;
    }

    return result;
}

function unproject(
    centerX: number,
    centerY: number,
    width: number,
    height: number,
    transform: Letterbox,
): Box {
    const { original, resized, left, top } = transform;
    const projectX = (x: number): number =>
        Math.max(0, Math.min(1, (x - left) / resized.width)) * original.width;
    const projectY = (y: number): number =>
        Math.max(0, Math.min(1, (y - top) / resized.height)) * original.height;
    const x = projectX(centerX - width / 2);
    const y = projectY(centerY - height / 2);

    return {
        x,
        y,
        width: projectX(centerX + width / 2) - x,
        height: projectY(centerY + height / 2) - y,
    };
}

function intersectionOverUnion(a: Box, b: Box): number {
    const width = Math.max(0, Math.min(a.x + a.width, b.x + b.width) - Math.max(a.x, b.x));
    const height = Math.max(0, Math.min(a.y + a.height, b.y + b.height) - Math.max(a.y, b.y));
    const intersection = width * height;
    const union = a.width * a.height + b.width * b.height - intersection;

    return union > 0 ? intersection / union : 0;
}

export function validateOutputShape(dims: readonly number[], classes: number): number {
    const count = dims[2];

    if (
        !Number.isSafeInteger(classes) ||
        classes < 1 ||
        classes > 1024 ||
        dims.length !== 3 ||
        dims[0] !== 1 ||
        dims[1] !== 4 + classes ||
        count === undefined ||
        !Number.isSafeInteger(count) ||
        count < 1 ||
        count > MAX_CANDIDATES
    ) {
        throw new Error(
            "Expected float32 [1,4+C,N], C matching labels and 1≤N≤100000. No objectness/NMS.",
        );
    }

    return count;
}

export function decode(
    data: Float32Array,
    dims: readonly number[],
    classes: number,
    transform: Letterbox,
): Detection[] {
    validateTransform(transform);
    const count = validateOutputShape(dims, classes);

    if (data.length !== (4 + classes) * count) {
        throw new Error("Output tensor data does not match its dimensions.");
    }

    const candidates: Detection[] = [];

    for (let index = 0; index < count; index += 1) {
        const centerX = data[index]!;
        const centerY = data[count + index]!;
        const width = data[count * 2 + index]!;
        const height = data[count * 3 + index]!;

        if (![centerX, centerY, width, height].every(Number.isFinite) || width < 0 || height < 0) {
            throw new Error(
                "Detector coordinates must be finite, with nonnegative width and height.",
            );
        }

        const { label, score } = bestClass(data, count, classes, index);

        if (score < 0.25) {
            continue;
        }

        const box = unproject(centerX, centerY, width, height, transform);

        if (box.width > 0 && box.height > 0) {
            candidates.push({ box, label, score });
        }
    }

    return suppressOverlaps(candidates);
}

function bestClass(
    data: Float32Array,
    count: number,
    classes: number,
    index: number,
): { label: number; score: number } {
    let label = 0;
    let score = 0;

    for (let candidate = 0; candidate < classes; candidate += 1) {
        const probability = data[(4 + candidate) * count + index]!;

        if (!Number.isFinite(probability) || probability < 0 || probability > 1) {
            throw new Error("Expected class probabilities in [0,1], not logits or objectness.");
        }

        if (probability > score) {
            score = probability;
            label = candidate;
        }
    }

    return { label, score };
}

function suppressOverlaps(candidates: Detection[]): Detection[] {
    candidates.sort((a, b) => b.score - a.score);
    const selected: Detection[] = [];

    for (const candidate of candidates.slice(0, 3000)) {
        const suppressed = selected.some(
            (previous) =>
                previous.label === candidate.label &&
                intersectionOverUnion(previous.box, candidate.box) > 0.45,
        );

        if (!suppressed) {
            selected.push(candidate);
        }

        if (selected.length === 300) {
            break;
        }
    }

    return selected;
}
05.05

建置、執行、試模型。

載入模型同標籤,再試一張本機圖片。由你決定幾時啟動相機;停止處理時,釋放相機媒體軌道同背景執行緒。

基礎說明

CameraApp.loadModel()
讀取你揀咗嘅本機檔案,並喺處理之前驗證模型同標籤。
CameraApp.start(camera)
驗證揀咗嘅來源,再交由圖片或相機流程處理。
CameraApp.stop()
停止接收新結果,等進行中嘅工作完成之後釋放資源。
要留意
症狀
停止擷取或者轉模型之後,舊偵測框又出現。
原因
較早開始嘅非同步操作,可能喺停止之後,或者揀咗新模型之後先完成。
處理方法
要求停止、捨棄遲到嘅結果,等進行中嘅工作完成之後先釋放資源。可唔可以取消底層執行,要睇平台支唔支援。
Terminalbash
npm run preview -- --port 4176

05.06 / 最佳實踐

做好細節,範例更好。

  • 擷取同推論唔好放喺介面執行緒。限制佇列長度,避免積埋舊影格。
  • 影像尺寸、裁切、方向同座標要有同一套約定。先用已知偵測框測試,再信任畫面標記。
  • 要求停止、捨棄遲到嘅結果,等進行中嘅工作完成之後先釋放資源。可唔可以取消底層執行,要睇平台支唔支援。