web-sdk-pp-detection 0.2.0 → 0.3.1

package/README.md

@@ -4,15 +4,15 @@
 
 基于 ONNX Runtime Web 的浏览器端 PP-Detection 目标检测 SDK，支持 PC、移动端与各类 H5 页面。
 
-当前 SDK 版本为 **0.2.0**，完整变更见[发布说明](https://github.com/chenmohan123/web-sdk-PP-Detection/blob/main/docs/zh-CN/release-0.2.0.md)。
+当前 SDK 版本为 **0.3.1**，完整变更见[发布说明](https://github.com/chenmohan123/web-sdk-PP-Detection/blob/main/docs/zh-CN/release-0.3.1.md)。
 
 ## 安装
 
 ```bash
-pnpm add web-sdk-pp-detection@0.2.0
+pnpm add web-sdk-pp-detection@0.3.1
 ```
 
-也可以使用 `npm install web-sdk-pp-detection@0.2.0`。
+也可以使用 `npm install web-sdk-pp-detection@0.3.1`。
 
 ## 快速开始
 
@@ -82,15 +82,15 @@ const detector = await createPPDetection({
 
 A browser-first PP-Detection object detection SDK powered by ONNX Runtime Web for desktop, mobile, and H5 pages.
 
-The current SDK version is **0.2.0**. See the [release notes](https://github.com/chenmohan123/web-sdk-PP-Detection/blob/main/docs/en/release-0.2.0.md) for the complete changes.
+The current SDK version is **0.3.1**. See the [release notes](https://github.com/chenmohan123/web-sdk-PP-Detection/blob/main/docs/en/release-0.3.1.md) for the complete changes.
 
 ### Installation
 
 ```bash
-pnpm add web-sdk-pp-detection@0.2.0
+pnpm add web-sdk-pp-detection@0.3.1
 ```
 
-`npm install web-sdk-pp-detection@0.2.0` is also supported.
+`npm install web-sdk-pp-detection@0.3.1` is also supported.
 
 ### Quick start
 
@@ -149,6 +149,10 @@ From 0.2.0, without an active detector, `ModelManager.getCacheEstimate({ id, ver
 
 WeChat official-account pages and other H5/WebView integrations are supported. Native Mini Program inference is not claimed; a native Mini Program should host the H5 experience in a WebView or use server-side inference.
 
+### PP-YOLOE 稳定模型（0.3.0 起）
+
+0.3.0 配套提供 PP-YOLOE+ S 640 FP32 `0.1.0 stable` 清单，模型从 ModelScope 或 Hugging Face 的固定 revision 下载，npm 包不包含权重。见[模型接入](https://github.com/chenmohan123/web-sdk-PP-Detection/blob/v0.3.0/examples/ppyoloe-candidate/README.md)。`allowExperimental` 默认关闭，仅运行 labs 候选时显式开启；blocked 仍被拒绝。Demo 默认 PicoDet，两个模型均默认 ModelScope。
+
 ### Documentation
 
 - [Chinese documentation](https://github.com/chenmohan123/web-sdk-PP-Detection#readme)

package/dist/browser-global.js

@@ -14007,34 +14007,58 @@ ${u}`, c = n.createShaderModule({ code: d, label: e.name });
     const rounded = value >> BICUBIC_PRECISION_BITS;
     return Math.max(0, Math.min(255, rounded));
   }
-  function bicubicChannel(raster, resizedWidth, resizedHeight, channel) {
+  function writeBicubicChannels(raster, resizedWidth, resizedHeight, target) {
     const horizontal = createBicubicAxis(raster.width, resizedWidth);
     const vertical = createBicubicAxis(raster.height, resizedHeight);
-    const intermediate = new Uint8Array(raster.height * resizedWidth);
+    const rowStride = resizedWidth * 3;
+    const intermediate = new Uint8Array(raster.height * rowStride);
     for (let y = 0; y < raster.height; y += 1) {
+      const sourceRow = y * raster.width * 4;
+      const targetRow = y * rowStride;
       for (let x = 0; x < resizedWidth; x += 1) {
         const [start, count] = horizontal.bounds[x];
         const coefficients = horizontal.coefficients[x];
-        let sum = BICUBIC_ROUNDING;
+        let red = BICUBIC_ROUNDING;
+        let green = BICUBIC_ROUNDING;
+        let blue = BICUBIC_ROUNDING;
+        let source2 = sourceRow + start * 4;
         for (let index = 0; index < count; index += 1) {
-          sum += raster.rgba[(y * raster.width + start + index) * 4 + channel] * coefficients[index];
+          const coefficient = coefficients[index];
+          red += raster.rgba[source2] * coefficient;
+          green += raster.rgba[source2 + 1] * coefficient;
+          blue += raster.rgba[source2 + 2] * coefficient;
+          source2 += 4;
         }
-        intermediate[y * resizedWidth + x] = clipBicubic(sum);
+        const offset = targetRow + x * 3;
+        intermediate[offset] = clipBicubic(red);
+        intermediate[offset + 1] = clipBicubic(green);
+        intermediate[offset + 2] = clipBicubic(blue);
       }
     }
-    const output = new Uint8Array(resizedWidth * resizedHeight);
+    const plane = target.inputWidth * target.inputHeight;
     for (let y = 0; y < resizedHeight; y += 1) {
       const [start, count] = vertical.bounds[y];
       const coefficients = vertical.coefficients[y];
+      const sourceRow = start * rowStride;
+      const targetRow = (y + target.padTop) * target.inputWidth + target.padLeft;
       for (let x = 0; x < resizedWidth; x += 1) {
-        let sum = BICUBIC_ROUNDING;
+        let red = BICUBIC_ROUNDING;
+        let green = BICUBIC_ROUNDING;
+        let blue = BICUBIC_ROUNDING;
+        let source2 = sourceRow + x * 3;
         for (let index = 0; index < count; index += 1) {
-          sum += intermediate[(start + index) * resizedWidth + x] * coefficients[index];
+          const coefficient = coefficients[index];
+          red += intermediate[source2] * coefficient;
+          green += intermediate[source2 + 1] * coefficient;
+          blue += intermediate[source2 + 2] * coefficient;
+          source2 += rowStride;
         }
-        output[y * resizedWidth + x] = clipBicubic(sum);
+        const offset = targetRow + x;
+        target.data[offset] = target.values[clipBicubic(red)];
+        target.data[plane + offset] = target.values[256 + clipBicubic(green)];
+        target.data[plane * 2 + offset] = target.values[512 + clipBicubic(blue)];
       }
     }
-    return output;
   }
   function preprocessImage(raster, preprocessing) {
     const inputWidth = preprocessing.size.width;
@@ -14061,18 +14085,35 @@ ${u}`, c = n.createShaderModule({ code: d, label: e.name });
     const mean = normalize ? preprocessing.mean ?? [0, 0, 0] : [0, 0, 0];
     const std = normalize ? preprocessing.std ?? [1, 1, 1] : [1, 1, 1];
     const interpolation2 = preprocessing.interpolation ?? "bilinear";
+    const bicubicValues = doResize && interpolation2 === "bicubic" ? new Float32Array(256 * 3) : void 0;
     for (let channel = 0; channel < 3; channel += 1) {
       const padding = normalize && mean[channel] !== 0 ? -mean[channel] / std[channel] : 0;
       data.fill(padding, channel * plane, (channel + 1) * plane);
-      const bicubic = doResize && interpolation2 === "bicubic" ? bicubicChannel(raster, resizedWidth, resizedHeight, channel) : void 0;
+      if (bicubicValues) {
+        for (let pixel = 0; pixel < 256; pixel += 1) {
+          const scaled = rescale ? pixel * preprocessing.rescaleFactor : pixel;
+          bicubicValues[channel * 256 + pixel] = normalize ? (scaled - mean[channel]) / std[channel] : scaled;
+        }
+        continue;
+      }
       for (let y = 0; y < resizedHeight; y += 1) {
         for (let x = 0; x < resizedWidth; x += 1) {
-          const pixel = bicubic ? bicubic[y * resizedWidth + x] : doResize ? bilinearChannel(raster, x, y, resizedWidth, resizedHeight, channel) : sampleChannel(raster, x, y, channel);
+          const pixel = doResize ? bilinearChannel(raster, x, y, resizedWidth, resizedHeight, channel) : sampleChannel(raster, x, y, channel);
           const scaled = rescale ? pixel * preprocessing.rescaleFactor : pixel;
           data[channel * plane + (y + padTop) * inputWidth + x + padLeft] = normalize ? (scaled - mean[channel]) / std[channel] : scaled;
         }
       }
     }
+    if (bicubicValues) {
+      writeBicubicChannels(raster, resizedWidth, resizedHeight, {
+        data,
+        inputWidth,
+        inputHeight,
+        padLeft,
+        padTop,
+        values: bicubicValues
+      });
+    }
     return {
       data,
       dims: [1, 3, inputHeight, inputWidth],
@@ -15667,13 +15708,19 @@ ${u}`, c = n.createShaderModule({ code: d, label: e.name });
     if (executionMode === "worker" && !capabilities.worker) {
       fail("CAPABILITY_UNSUPPORTED", "\u8BF7\u6C42\u7684 worker \u4E0D\u53EF\u7528", { executionMode });
     }
-    const variant3 = manifest.variants?.find(
-      (candidate) => candidate.precision === requestedPrecision && candidate.status !== "labs" && candidate.status !== "blocked"
-    );
+    const matchingVariants = manifest.variants?.filter(
+      (candidate) => candidate.precision === requestedPrecision && candidate.status !== "blocked"
+    ) ?? [];
+    const variant3 = matchingVariants.find((candidate) => candidate.status !== "labs") ?? (options.allowExperimental === true ? matchingVariants.find((candidate) => candidate.status === "labs") : void 0);
     if (!variant3) {
-      fail("MODEL_INCOMPATIBLE", `manifest \u6CA1\u6709\u53EF\u7528\u7684 ${requestedPrecision} \u7A33\u5B9A\u53D8\u4F53`, {
-        requestedPrecision
-      });
+      fail(
+        "MODEL_INCOMPATIBLE",
+        options.allowExperimental === true ? `manifest \u6CA1\u6709\u53EF\u7528\u7684 ${requestedPrecision} \u53D8\u4F53` : `manifest \u6CA1\u6709\u53EF\u7528\u7684 ${requestedPrecision} \u7A33\u5B9A\u53D8\u4F53`,
+        {
+          requestedPrecision,
+          allowExperimental: options.allowExperimental === true
+        }
+      );
     }
     const candidates = candidatesForBackend(requestedBackend, capabilities).filter(
       (backend) => variant3.backends.includes(backend)
@@ -15849,7 +15896,7 @@ ${u}`, c = n.createShaderModule({ code: d, label: e.name });
   };
 
   // src/index.ts
-  var CURRENT_SDK_VERSION = "0.2.0";
+  var CURRENT_SDK_VERSION = "0.3.1";
   function probePPDetectionCapabilities(options = {}) {
     return probeCapabilities(options);
   }
@@ -15964,6 +16011,7 @@ ${u}`, c = n.createShaderModule({ code: d, label: e.name });
     options.onProgress?.({ phase: "manifest", status: "complete" });
     const plan = selectExecutionPlan(
       {
+        allowExperimental: options.allowExperimental,
         backend: options.backend,
         precision: options.precision === "auto" ? void 0 : options.precision,
         executionMode: options.executionMode,