Upload 2 files

README.md

@@ -24,3 +24,6 @@ SentAI is a browser-based live camera application for face detection, expression
 ## Notes
 
 The app estimates visible facial expression and apparent age from camera frames. It cannot know a person's true internal feeling. Lighting, pose, camera quality, glasses, occlusion, and model bias can affect results.
+
+
+Phase 3B fixes the accuracy-pack loader by using the Transformers.js package version that the selected ONNX emotion model was published for.

index.html

@@ -485,10 +485,11 @@
   <div id="toast" class="toast" role="status" aria-live="polite"></div>
 
   <script type="module">
+    import { pipeline as xenovaPipeline, env as xenovaEnv } from "https://cdn.jsdelivr.net/npm/@xenova/transformers@2.17.2";
     const FACE_API_MODEL_URL = "https://cdn.jsdelivr.net/gh/justadudewhohacks/face-api.js@0.22.2/weights";
-    const TRANSFORMERS_CDN = "https://cdn.jsdelivr.net/npm/@huggingface/transformers@4.2.0";
+    const TRANSFORMERS_CDN = "https://cdn.jsdelivr.net/npm/@xenova/transformers@2.17.2";
     const EMOTION_MODEL_ID = "Xenova/facial_emotions_image_detection";
-    const AGE_GENDER_MODEL_ID = "onnx-community/age-gender-prediction-ONNX";
+    const AGE_GENDER_MODEL_ID = null; // Phase 3B keeps age on the stable face-api model and improves it with smoothing + calibration.
 
     const els = {
       coreDot: document.getElementById("coreDot"),
@@ -630,7 +631,7 @@
       };
 
       const nonNeutralTop = Math.max(raw.happy, raw.sad, raw.fearful, raw.angry, raw.disgusted, raw.surprised);
-      const uncertainty = clamp01((1 - nonNeutralTop) * 0.28 + raw.surprised * 0.54 + raw.neutral * 0.16);
+      const uncertainty = clamp01((1 - nonNeutralTop) * 0.10 + raw.surprised * 0.42 + raw.neutral * 0.05);
 
       return normalizeScores({
         Happy: Math.pow(raw.happy, 1.08),
@@ -638,7 +639,7 @@
         Fear: Math.max(Math.pow(raw.fearful, 1.04), raw.surprised * raw.fearful * 0.55),
         Anger: Math.pow(raw.angry, 1.03),
         Disgust: Math.pow(raw.disgusted, 1.02),
-        Confused: Math.min(0.62, uncertainty),
+        Confused: Math.min(0.42, uncertainty),
       });
     }
 
@@ -655,8 +656,8 @@
         else if (label.includes("angry") || label.includes("anger")) scores.Anger = Math.max(scores.Anger, score);
         else if (label.includes("disgust")) scores.Disgust = Math.max(scores.Disgust, score);
         else if (label.includes("surprise") || label.includes("neutral")) {
-          const scaled = label.includes("neutral") ? score * 0.30 : score * 0.72;
-          scores.Confused = Math.max(scores.Confused, Math.min(0.68, scaled));
+          const scaled = label.includes("neutral") ? score * 0.10 : score * 0.55;
+          scores.Confused = Math.max(scores.Confused, Math.min(0.48, scaled));
         }
       }
       return normalizeScores(scores);
@@ -689,9 +690,13 @@
       const second = sorted[1] || ["", 0];
       let label = top[0];
       let score = top[1];
-      if (score < 0.22 || (score - second[1] < 0.055 && label !== "Confused")) {
+      if (score < 0.18) {
         label = "Confused";
-        score = Math.max(scores.Confused || 0, 0.24);
+        score = Math.max(scores.Confused || 0, 0.18);
+      } else if (score - second[1] < 0.035 && label !== "Confused") {
+        // When the visible expression is ambiguous, mark it as confused but keep confidence modest.
+        label = "Confused";
+        score = Math.max(Math.min(scores.Confused || 0, 0.34), 0.22);
       }
       return { label, score: clamp01(score) };
     }
@@ -994,26 +999,31 @@
     }
 
     async function getTransformersModule() {
-      const mod = await import(TRANSFORMERS_CDN);
+      // Use the legacy Xenova build because the selected emotion model was published for it.
+      // This avoids the Phase 3A failure where the newer package could import but could not initialize the model.
+      const mod = { pipeline: xenovaPipeline, env: xenovaEnv };
       if (mod.env) {
         mod.env.allowLocalModels = false;
-        if (mod.env.backends && mod.env.backends.onnx && mod.env.backends.onnx.wasm) {
-          mod.env.backends.onnx.wasm.numThreads = Math.max(1, Math.min(4, navigator.hardwareConcurrency || 2));
-        }
+        mod.env.useBrowserCache = true;
+        mod.env.backends ??= {};
+        mod.env.backends.onnx ??= {};
+        mod.env.backends.onnx.wasm ??= {};
+        mod.env.backends.onnx.wasm.numThreads = Math.max(1, Math.min(4, navigator.hardwareConcurrency || 2));
       }
       return mod;
     }
 
     async function loadPipelineWithFallback(pipeline, task, modelId) {
-      const preferWebGpu = !!navigator.gpu;
-      const attempts = preferWebGpu
-        ? [{ device: "webgpu", dtype: "q8" }, { device: "wasm", dtype: "q8" }, { device: "wasm" }]
-        : [{ device: "wasm", dtype: "q8" }, { device: "wasm" }];
+      const attempts = [
+        { quantized: true, progress_callback: p => { if (p?.status) console.log("accuracy pack", p.status, p.file || ""); } },
+        { quantized: true },
+        {},
+      ];
       let lastErr = null;
       for (const opts of attempts) {
         try {
           const pipe = await pipeline(task, modelId, opts);
-          pro.device = opts.device || "wasm";
+          pro.device = "wasm";
           return pipe;
         } catch (err) {
           lastErr = err;
@@ -1050,7 +1060,7 @@
       pro.loading = true;
       pro.tried = true;
       els.accuracyBtn.disabled = true;
-      setPill(els.proDot, els.proStatus, auto ? "Accuracy pack: loading in background..." : "Accuracy pack: loading...", "loading");
+      setPill(els.proDot, els.proStatus, auto ? "Emotion accuracy pack: loading in background..." : "Emotion accuracy pack: loading...", "loading");
       if (!auto) showToast("Loading higher-accuracy transformer models. First load can take time.");
 
       try {
@@ -1064,7 +1074,7 @@
         }
 
         try {
-          if (mod.AutoModel && mod.AutoProcessor && (mod.load_image || mod.RawImage?.fromURL)) {
+          if (AGE_GENDER_MODEL_ID && mod.AutoModel && mod.AutoProcessor && (mod.load_image || mod.RawImage?.fromURL)) {
             const pair = await loadAgeGenderWithFallback(mod);
             pro.ageGenderModel = pair.model;
             pro.ageGenderProcessor = pair.processor;
@@ -1084,7 +1094,7 @@
         }
       } catch (err) {
         console.error(err);
-        setPill(els.proDot, els.proStatus, "Accuracy pack unavailable; using core model", "error");
+        setPill(els.proDot, els.proStatus, "Emotion accuracy pack unavailable; using core model", "error");
         if (!auto) showToast("Accuracy pack could not load. The app will keep using the core model.");
       } finally {
         pro.loading = false;