@camstack/addon-pipeline 0.1.18 → 0.1.20

package/dist/detection-pipeline/index.mjs

@@ -1,4 +1,4 @@
-import { h as hfModelUrl, C as COCO_TO_MACRO, g as COCO_80_LABELS, A as AUDIO_MACRO_LABELS, e as errMsg, f as createEvent, E as EventCategory, i as evaluateZoneRules, Y as YAMNET_TO_MACRO, j as APPLE_SA_TO_MACRO, k as parseJsonUnknown, B as BaseAddon, m as hydrateSchema, q as pipelineExecutorCapability, r as detectionPipelineCapability } from "../index-asZs8U_s.mjs";
+import { h as hfModelUrl, C as COCO_TO_MACRO, g as COCO_80_LABELS, A as AUDIO_MACRO_LABELS, e as errMsg, f as createEvent, E as EventCategory, i as sleep, j as evaluateZoneRules, Y as YAMNET_TO_MACRO, k as APPLE_SA_TO_MACRO, q as parseJsonUnknown, B as BaseAddon, r as hydrateSchema, t as pipelineExecutorCapability, u as detectionPipelineCapability } from "../index-5aYef068.mjs";
 import * as fs from "node:fs";
 import * as path from "node:path";
 import * as os from "node:os";
@@ -872,12 +872,12 @@ const OBJECT_DETECTION_MODELS = [
       openvino: { url: hf("objectDetection/yolo26/openvino/camstack-yolo26x.xml"), sizeMB: 213, runtimes: ["python"] }
     }
   },
-  // ── Scrypted ReLU variants (optimized for int8 quantization) ──
-  // Note: Scrypted CoreML models are compiled with 320x320 fixed input size
+  // ── ReLU variants (optimized for int8 quantization) ──
+  // Note: these CoreML models are compiled with 320x320 fixed input size
   {
     id: "scrypted-yolov9t-relu",
-    name: "YOLOv9 Tiny ReLU (Scrypted)",
-    description: "Scrypted YOLOv9t ReLU — optimized for int8 quantization",
+    name: "YOLOv9 Tiny ReLU",
+    description: "YOLOv9t ReLU — optimized for int8 quantization",
     inputSize: { width: 320, height: 320 },
     labels: [],
     preprocessMode: "letterbox",
@@ -889,8 +889,8 @@ const OBJECT_DETECTION_MODELS = [
   },
   {
     id: "scrypted-yolov9s-relu",
-    name: "YOLOv9 Small ReLU (Scrypted)",
-    description: "Scrypted YOLOv9s ReLU — improved efficiency, int8 ready",
+    name: "YOLOv9 Small ReLU",
+    description: "YOLOv9s ReLU — improved efficiency, int8 ready",
     inputSize: { width: 320, height: 320 },
     labels: [],
     preprocessMode: "letterbox",
@@ -902,9 +902,9 @@ const OBJECT_DETECTION_MODELS = [
   },
   {
     id: "scrypted-yolov9c-relu",
-    name: "YOLOv9 C ReLU (Scrypted)",
-    description: "Scrypted YOLOv9c ReLU — high-accuracy, int8 ready",
-    // Scrypted ReLU variants are all compiled with fixed 320x320 input
+    name: "YOLOv9 C ReLU",
+    description: "YOLOv9c ReLU — high-accuracy, int8 ready",
+    // ReLU variants are all compiled with fixed 320x320 input
     // (see the section header above). Validated at engine load via
     // `validateOnnxInputShape` — mismatches warn, never silently miscompute.
     inputSize: { width: 320, height: 320 },
@@ -918,9 +918,9 @@ const OBJECT_DETECTION_MODELS = [
   },
   {
     id: "scrypted-yolov9m-relu",
-    name: "YOLOv9 M ReLU (Scrypted)",
-    description: "Scrypted YOLOv9m ReLU — medium, int8 ready",
-    // Scrypted ReLU variants are all compiled with fixed 320x320 input
+    name: "YOLOv9 M ReLU",
+    description: "YOLOv9m ReLU — medium, int8 ready",
+    // ReLU variants are all compiled with fixed 320x320 input
     // (see the section header above). Validated at engine load via
     // `validateOnnxInputShape` — mismatches warn, never silently miscompute.
     inputSize: { width: 320, height: 320 },
@@ -949,9 +949,9 @@ const FACE_DETECTION_MODELS = [
   },
   {
     id: "scrypted-yolov9t-face",
-    name: "YOLOv9t Face ReLU (Scrypted)",
-    description: "Scrypted YOLOv9t face detection — YOLO-based, fast",
-    // Scrypted YOLO face variants follow the 320x320 ReLU convention.
+    name: "YOLOv9t Face ReLU",
+    description: "YOLOv9t face detection — YOLO-based, fast, 320x320 ReLU",
+    // ReLU face variants follow the 320x320 fixed input convention.
     // Validated at engine load via `validateOnnxInputShape`.
     inputSize: { width: 320, height: 320 },
     labels: [{ id: "face", name: "Face" }],
@@ -980,8 +980,8 @@ const FACE_EMBEDDING_MODELS = [
   },
   {
     id: "inception-resnet-v1",
-    name: "Inception ResNet V1 (Scrypted)",
-    description: "FaceNet-style face recognition embeddings (512-d) — Scrypted model",
+    name: "Inception ResNet V1",
+    description: "FaceNet-style face recognition embeddings (512-d) — hosted on plugin-models HF repo",
     inputSize: { width: 160, height: 160 },
     labels: [{ id: "embedding", name: "Face Embedding" }],
     preprocessMode: "resize",
@@ -1794,8 +1794,8 @@ class EngineFactory {
    * single IPC round-trip. Returns the structured Python detection
    * dicts in input order. Throws when this factory has no pool path.
    *
-   * Used by `runPipelineBatch` for fair benchmarking against
-   * Scrypted's `detectObjects(media, {batch})`.
+   * Used by `runPipelineBatch` for fair batched benchmarking — N frames
+   * in one IPC round-trip, amortising the per-call envelope overhead.
    */
   async batchInferRaw(stepId, items, modelId) {
     if (!this.poolManager) {
@@ -3214,7 +3214,8 @@ function parseWavToAudioChunk(filePath) {
   for (let i = 0; i < pcm.length; i++) {
     float32[i] = pcm[i] / 32768;
   }
-  return { data: float32, sampleRate: 16e3, channels: 1, timestamp: Date.now() };
+  const data = new Uint8Array(float32.buffer, float32.byteOffset, float32.byteLength);
+  return { data, sampleRate: 16e3, channels: 1, timestamp: Date.now() };
 }
 function enginesEqual(a, b) {
   return a.runtime === b.runtime && a.backend === b.backend && a.format === b.format && (a.device ?? null) === (b.device ?? null);
@@ -4124,7 +4125,7 @@ class DetectionPipelineProvider {
         if (attempt < maxRetries) {
           const delayMs = 2e3 * Math.pow(2, attempt - 1);
           this.log.warn("Model download failed — retrying", { meta: { modelId: entry.id, attempt, maxRetries, error: msg, retryDelayMs: delayMs } });
-          await new Promise((resolve) => setTimeout(resolve, delayMs));
+          await sleep(delayMs);
         } else {
           throw new Error(`Failed to download model "${entry.id}" after ${maxRetries} attempts: ${msg}`, { cause: err });
         }
@@ -4199,10 +4200,10 @@ class DetectionPipelineProvider {
    * factory). Returns one minimally-shaped `FrameResult` per input
    * frame in input order.
    *
-   * Used by `scripts/bench-scrypted-style.mts` for the
-   * Scrypted-equivalent benchmark — Scrypted's `detectObjects(media,
-   * {batch})` collapses N requests into one call too, so bypassing
-   * the per-call IPC overhead is what makes the comparison fair.
+   * Used by `scripts/bench-batch-style.mts` for the batch benchmark —
+   * collapsing N requests into one call amortises the per-call IPC
+   * envelope overhead, which is what makes the batched path measurably
+   * faster than N sequential `runPipeline` calls.
    */
   async runPipelineBatch(input) {
     this.log.debug("runPipelineBatch entry", {
@@ -4895,7 +4896,7 @@ class DetectionPipelineProvider {
         label: l.className,
         score: l.score
       }));
-      const sampleCount = chunk.data.length;
+      const sampleCount = chunk.data.byteLength / 4;
       const startMs = 0;
       const endMs = sampleCount / chunk.sampleRate * 1e3;
       const detections = ranked.filter((l) => l.score >= minConfidence).map((l, i) => ({
@@ -5183,7 +5184,15 @@ const TUNING_DEFAULTS = {
   // (~+30-90% on 4-8 simulated cams in `bench-pool-inprocess.mts`).
   // Higher counts saturate the single ANE channel and degrade — 2 is the
   // sweet spot on M-series.
-  coreml: { concurrency: 2, batchMode: "window", windowMs: 30, maxBatchSize: 8, numWorkers: 2 },
+  // Latency-first: one model context, one serialized predict, no cross-camera
+  // window batching. On a single shared ANE, concurrent predicts
+  // (concurrency×numWorkers) oversubscribe the accelerator, so per-frame latency
+  // scales with active-camera count; a `window` adds pure queueing delay and
+  // `model.predict([list])` does NOT fuse image-typed inputs (it loops), so the
+  // batch is no faster while every member waits for the slowest. One predict
+  // owns the ANE → each inference runs at its isolated speed. (Prev throughput-
+  // tuned default was concurrency:2, batchMode:'window', windowMs:30, numWorkers:2.)
+  coreml: { concurrency: 1, batchMode: "none", windowMs: 8, maxBatchSize: 8, numWorkers: 1 },
   openvino: { concurrency: 1, batchMode: "none", numStreams: 0 },
   onnx: { concurrency: 4, batchMode: "list", maxBatchSize: 8, intraOpThreads: 0 },
   cpu: { concurrency: 4, batchMode: "list", maxBatchSize: 8, intraOpThreads: 0 }
@@ -5506,23 +5515,17 @@ class DetectionPipelineAddon extends BaseAddon {
   resolveBackendTuning() {
     const backend = this.config.engineBackend ?? DEFAULT_CONFIG.engineBackend;
     const t = TUNING_DEFAULTS[backend] ?? TUNING_DEFAULTS["onnx"];
-    const cfg = this.config;
-    const pickNum = (userVal, defaultVal) => {
-      if (typeof userVal !== "number" || userVal <= 0) return defaultVal;
-      return userVal;
-    };
-    const pickStr = (userVal, defaultVal) => {
-      if (typeof userVal !== "string" || userVal === "") return defaultVal;
-      return userVal;
-    };
+    const tAny = t;
+    const num = (v, dflt) => typeof v === "number" && v > 0 ? v : dflt;
+    const str = (v, dflt) => typeof v === "string" && v !== "" ? v : dflt;
     return {
-      concurrency: pickNum(cfg.concurrency, t.concurrency ?? 1),
-      batchMode: pickStr(cfg.batchMode, t.batchMode ?? "window"),
-      windowMs: pickNum(cfg.windowMs, t.windowMs ?? 2),
-      maxBatchSize: pickNum(cfg.maxBatchSize, t.maxBatchSize ?? 8),
-      numStreams: pickNum(cfg.numStreams, t.numStreams ?? 0),
-      intraOpThreads: pickNum(cfg.intraOpThreads, t.intraOpThreads ?? 0),
-      numWorkers: pickNum(cfg.numWorkers, t["numWorkers"] ?? 1)
+      concurrency: num(t.concurrency, 1),
+      batchMode: str(t.batchMode, "window"),
+      windowMs: num(t.windowMs, 2),
+      maxBatchSize: num(t.maxBatchSize, 8),
+      numStreams: num(t.numStreams, 0),
+      intraOpThreads: num(t.intraOpThreads, 0),
+      numWorkers: num(tAny["numWorkers"], 1)
     };
   }
   async onInitialize() {