[RNE Rewrite] Add text and image embeddings pipelines

.agents/skills/model-schema-validation/SKILL.md

@@ -62,6 +62,20 @@ The `SymbolicTensor` helper supports specifying **Symbolic Shapes** where intege
 
 ---
 
+## ⚙️ Runtime Dynamic Dimensions: the `get_dynamic_dims_<methodName>` companion
+
+`SymbolicTensor` string symbols let the **validator** accept a range of shapes, but an ExecuTorch `.pte`'s metadata only serializes the **static upper bound** of a dynamic dimension — not its active `[min, max, step]` range. So for an input dimension that genuinely varies at runtime (e.g. a text model's sequence length), the model must be **exported with a companion method** that re-exposes the range to the runtime:
+
+* **Name**: `get_dynamic_dims_<methodName>` — e.g. `get_dynamic_dims_forward` for `forward`.
+* **Signature**: takes no arguments.
+* **Returns**: a list of outputs, one per `Tensor` input of the target method (scalar inputs are skipped), each a **2D `int32` tensor of shape `[rank, 3]`** whose rows are `[min, max, step]` constraints for that input's dimensions — e.g. `[1, 1, 1]` for a fixed dimension and `[1, maxTokens, 1]` for the dynamic one.
+
+At load time the C++ core (`Model::parseDynamicInputShapes`) reads this companion and validates inputs against the range; `model.execute` then accepts tensors at their exact runtime length. **Without the companion, a method falls back to exact per-dimension validation** — it only accepts the single shape it was exported with. So a `.pte` whose metadata reports `[1, 512]` but is meant to accept `[1, 1..512]` MUST ship `get_dynamic_dims_forward`, or variable-length inputs are rejected at runtime.
+
+This is an **export-side contract** (the export-scripts repo provides a `build_dynamic_dims_program` helper that emits the companion). The TypeScript task only declares the symbol via `SymbolicTensor` and reads the resulting upper bound from `meta.inputTensorMeta`.
+
+---
+
 ## 📋 Common Validation Recipes
 
 ### 1. Classification

apps/computer-vision/app/_layout.tsx

@@ -27,6 +27,13 @@ export default function Layout() {
           title: 'Classification',
         }}
       />
+      <Drawer.Screen
+        name="imageEmbeddings/index"
+        options={{
+          drawerLabel: 'Image Embeddings',
+          title: 'Image Embeddings',
+        }}
+      />
       <Drawer.Screen
         name="detection/index"
         options={{

apps/computer-vision/app/classification/index.tsx

@@ -1,6 +1,7 @@
 import React, { useState } from 'react';
 import { View, Text, StyleSheet, ScrollView, Platform } from 'react-native';
-import { commonStyles, ColorPalette } from '../../theme';
+import { useSafeAreaInsets } from 'react-native-safe-area-context';
+import { commonStyles, ColorPalette, theme } from '../../theme';
 import { useImage } from '@shopify/react-native-skia';
 import { useClassifier, models } from 'react-native-executorch';
 import ScreenWrapper from '../../components/ScreenWrapper';
@@ -28,6 +29,7 @@ const MODEL_OPTIONS: ModelOption[] = [
 ];
 
 function ClassificationContent() {
+  const insets = useSafeAreaInsets();
   const [selectedModel, setSelectedModel] = useState<any>(MODEL_OPTIONS[0].value);
   const [imageUri, setImageUri] = useState<string | null>(null);
   const [isProcessing, setIsProcessing] = useState(false);
@@ -97,7 +99,10 @@ function ClassificationContent() {
   return (
     <ScrollView
       style={commonStyles.container}
-      contentContainerStyle={commonStyles.contentContainer}
+      contentContainerStyle={[
+        commonStyles.contentContainer,
+        { paddingBottom: insets.bottom + theme.spacing.large },
+      ]}
     >
       <Text style={commonStyles.description}>
         Upload or capture an image to identify objects using a classifier.

apps/computer-vision/app/imageEmbeddings/index.tsx

@@ -0,0 +1,250 @@
+import React, { useState } from 'react';
+import { View, Text, StyleSheet, ScrollView, TextInput, TouchableOpacity } from 'react-native';
+import { useSafeAreaInsets } from 'react-native-safe-area-context';
+import { commonStyles, ColorPalette } from '../../theme';
+import { useImage } from '@shopify/react-native-skia';
+import { useImageEmbedder, useTextEmbedder, models } from 'react-native-executorch';
+import ScreenWrapper from '../../components/ScreenWrapper';
+import { getImage, skImageToBuffer } from '../../utils';
+import { ModelPicker, type ModelOption } from '../../components/ModelPicker';
+import { ImageViewport } from '../../components/ImageViewport';
+import { ModelStatus } from '../../components/ModelStatus';
+import { LatencyIndicator } from '../../components/LatencyIndicator';
+import { Button } from '../../components/Button';
+
+const IMAGE_MODEL_OPTIONS: ModelOption[] = [
+  {
+    label: 'CLIP ViT-B/32 (INT8)',
+    value: models.imageEmbeddings.CLIP_VIT_BASE_PATCH32.XNNPACK_INT8,
+  },
+  {
+    label: 'CLIP ViT-B/32 (FP32)',
+    value: models.imageEmbeddings.CLIP_VIT_BASE_PATCH32.XNNPACK_FP32,
+  },
+];
+
+const DEFAULT_LABELS = [
+  'a photo of a dog',
+  'a photo of a cat',
+  'a landscape photo',
+  'a photo of food',
+  'a photo of people',
+];
+
+// CLIP text and image embeddings are L2-normalized, so their cosine similarity
+// is the dot product.
+const dot = (a: Float32Array, b: Float32Array) => {
+  let s = 0;
+  for (let i = 0; i < a.length; i++) {
+    s += a[i]! * b[i]!;
+  }
+  return s;
+};
+
+function ImageEmbeddingsContent() {
+  const [selectedImageModel, setSelectedImageModel] = useState<any>(IMAGE_MODEL_OPTIONS[0].value);
+  const [imageUri, setImageUri] = useState<string | null>(null);
+  const [labels, setLabels] = useState<string[]>(DEFAULT_LABELS);
+  const [newLabel, setNewLabel] = useState('');
+  const [results, setResults] = useState<{ label: string; score: number }[]>([]);
+  const [latency, setLatency] = useState<number | null>(null);
+  const [isProcessing, setIsProcessing] = useState(false);
+  const [error, setError] = useState<string | null>(null);
+
+  const insets = useSafeAreaInsets();
+  const skiaImage = useImage(imageUri, (err) => setError(err.message || String(err)));
+
+  // Zero-shot classification pairs a CLIP image encoder with the CLIP text
+  // encoder and scores the image against each text label by embedding similarity.
+  const imageModel = useImageEmbedder(selectedImageModel);
+  const textModel = useTextEmbedder(models.textEmbeddings.CLIP_VIT_BASE_PATCH32_TEXT);
+
+  const ready = imageModel.isReady && textModel.isReady;
+
+  const pickImage = async () => {
+    setError(null);
+    try {
+      const uri = await getImage(false);
+      if (!uri) return;
+      setImageUri(uri);
+      setResults([]);
+      setLatency(null);
+    } catch (e: any) {
+      setError(e.message || String(e));
+    }
+  };
+
+  const classify = async () => {
+    if (!skiaImage || !ready || !imageModel.embed || !textModel.embed) return;
+    setIsProcessing(true);
+    setError(null);
+    try {
+      const start = Date.now();
+      const imageEmbedding = await imageModel.embed(skImageToBuffer(skiaImage));
+      const scored: { label: string; score: number }[] = [];
+      for (const label of labels) {
+        const textEmbedding = await textModel.embed(label);
+        scored.push({ label, score: dot(imageEmbedding, textEmbedding) });
+      }
+      scored.sort((a, b) => b.score - a.score);
+      setLatency(Date.now() - start);
+      setResults(scored);
+    } catch (e: any) {
+      setError(e.message || String(e));
+    } finally {
+      setIsProcessing(false);
+    }
+  };
+
+  const addLabel = () => {
+    const trimmed = newLabel.trim();
+    if (!trimmed || labels.includes(trimmed)) return;
+    setLabels((prev) => [...prev, trimmed]);
+    setNewLabel('');
+    setResults([]);
+  };
+
+  const removeLabel = (label: string) => {
+    setLabels((prev) => prev.filter((l) => l !== label));
+    setResults((prev) => prev.filter((r) => r.label !== label));
+  };
+
+  const activeError = imageModel.error
+    ? String(imageModel.error)
+    : textModel.error
+      ? String(textModel.error)
+      : error;
+
+  return (
+    <ScrollView
+      style={commonStyles.container}
+      contentContainerStyle={[commonStyles.contentContainer, { paddingBottom: insets.bottom + 24 }]}
+    >
+      <Text style={commonStyles.description}>
+        Pick an image, then rank text labels by how well CLIP matches them to it (zero-shot
+        classification).
+      </Text>
+
+      <ModelPicker
+        label="Image model"
+        options={IMAGE_MODEL_OPTIONS}
+        selectedValue={selectedImageModel}
+        onValueChange={(model) => {
+          setSelectedImageModel(model);
+          setResults([]);
+          setLatency(null);
+        }}
+      />
+
+      <ModelStatus
+        isReady={ready}
+        downloadProgress={Math.min(imageModel.downloadProgress, textModel.downloadProgress)}
+        error={activeError}
+        modelTypeLabel="CLIP models"
+      />
+
+      <ImageViewport skiaImage={skiaImage} onPressPlaceholder={pickImage} />
+
+      <View style={commonStyles.buttonRow}>
+        <Button title="Pick image" onPress={pickImage} variant="secondary" />
+        <Button
+          title="Find best label"
+          onPress={classify}
+          disabled={!skiaImage || !ready || isProcessing}
+          loading={isProcessing}
+        />
+      </View>
+
+      <LatencyIndicator latency={latency} />
+
+      {results.length > 0 && (
+        <View style={styles.card}>
+          <Text style={styles.cardTitle}>Results</Text>
+          {results.map((r, i) => (
+            <View key={r.label} style={styles.row}>
+              <Text style={[styles.rowLabel, i === 0 && styles.topLabel]} numberOfLines={1}>
+                {i === 0 ? '🥇 ' : ''}
+                {r.label}
+              </Text>
+              <Text style={styles.rowScore}>{r.score.toFixed(3)}</Text>
+            </View>
+          ))}
+        </View>
+      )}
+
+      <View style={styles.card}>
+        <Text style={styles.cardTitle}>Labels</Text>
+        {labels.map((label) => (
+          <View key={label} style={styles.row}>
+            <Text style={styles.rowLabel} numberOfLines={1}>
+              {label}
+            </Text>
+            <TouchableOpacity onPress={() => removeLabel(label)} hitSlop={8}>
+              <Text style={styles.remove}>✕</Text>
+            </TouchableOpacity>
+          </View>
+        ))}
+        <View style={styles.addRow}>
+          <TextInput
+            style={styles.input}
+            placeholder="Add a label…"
+            placeholderTextColor="#94A3B8"
+            value={newLabel}
+            onChangeText={setNewLabel}
+            onSubmitEditing={addLabel}
+            returnKeyType="done"
+          />
+          <Button title="Add" onPress={addLabel} disabled={!newLabel.trim()} variant="secondary" />
+        </View>
+      </View>
+    </ScrollView>
+  );
+}
+
+export default function ImageEmbeddingsScreen() {
+  return (
+    <ScreenWrapper>
+      <ImageEmbeddingsContent />
+    </ScreenWrapper>
+  );
+}
+
+const styles = StyleSheet.create({
+  card: {
+    width: '100%',
+    backgroundColor: '#fff',
+    borderRadius: 12,
+    padding: 16,
+    borderWidth: 1,
+    borderColor: '#e9ecef',
+    marginTop: 16,
+  },
+  cardTitle: {
+    fontSize: 16,
+    fontWeight: '600',
+    color: ColorPalette.strongPrimary,
+    marginBottom: 8,
+  },
+  row: {
+    flexDirection: 'row',
+    justifyContent: 'space-between',
+    alignItems: 'center',
+    paddingVertical: 8,
+    borderBottomWidth: 1,
+    borderBottomColor: '#f1f3f5',
+  },
+  rowLabel: { fontSize: 14, color: '#334155', flex: 1, marginRight: 8 },
+  topLabel: { fontWeight: '700', color: ColorPalette.strongPrimary },
+  rowScore: { fontSize: 13, fontWeight: '600', color: ColorPalette.primary },
+  remove: { fontSize: 16, color: '#94A3B8', paddingHorizontal: 4 },
+  addRow: { flexDirection: 'row', alignItems: 'center', gap: 8, marginTop: 12 },
+  input: {
+    flex: 1,
+    backgroundColor: '#f1f3f5',
+    borderRadius: 10,
+    paddingHorizontal: 12,
+    paddingVertical: 10,
+    fontSize: 14,
+    color: '#0F172A',
+  },
+});

apps/computer-vision/app/index.tsx

@@ -14,6 +14,9 @@ export default function Home() {
         <TouchableOpacity style={styles.button} onPress={() => router.navigate('classification/')}>
           <Text style={styles.buttonText}>Classification</Text>
         </TouchableOpacity>
+        <TouchableOpacity style={styles.button} onPress={() => router.navigate('imageEmbeddings/')}>
+          <Text style={styles.buttonText}>Image Embeddings</Text>
+        </TouchableOpacity>
         <TouchableOpacity style={styles.button} onPress={() => router.navigate('detection/')}>
           <Text style={styles.buttonText}>Object Detection</Text>
         </TouchableOpacity>

apps/computer-vision/package.json

@@ -5,6 +5,8 @@
   "react-native-executorch": {
     "features": [
       "classification",
+      "imageEmbeddings",
+      "textEmbeddings",
       "instanceSegmentation",
       "keypointDetection",
       "objectDetection",

apps/computer-vision/utils.ts

@@ -1,6 +1,32 @@
 import { Alert } from 'react-native';
 import * as ImagePicker from 'expo-image-picker';
 import { ImageManipulator, SaveFormat } from 'expo-image-manipulator';
+import type { SkImage } from '@shopify/react-native-skia';
+
+/**
+ * Converts a Skia image into the raw RGBA/HWC image buffer that
+ * react-native-executorch vision tasks accept. Throws if the pixel data cannot
+ * be read.
+ * @param image - The Skia image to read pixels from.
+ * @returns The RGBA/HWC image buffer with its `data`, `width`, `height`,
+ * `format`, and `layout`.
+ */
+export const skImageToBuffer = (image: SkImage) => {
+  const pixels = image.readPixels();
+  if (!pixels) {
+    throw new Error('Failed to read pixels from image');
+  }
+  if (!(pixels instanceof Uint8Array)) {
+    throw new Error('Expected Uint8Array from readPixels');
+  }
+  return {
+    data: pixels,
+    width: image.width(),
+    height: image.height(),
+    format: 'rgba' as const,
+    layout: 'hwc' as const,
+  };
+};
 
 export const getImage = async (
   useCamera: boolean,

apps/nlp/app/_layout.tsx

@@ -27,6 +27,13 @@ export default function Layout() {
           title: 'Tokenizer',
         }}
       />
+      <Drawer.Screen
+        name="text-embeddings/index"
+        options={{
+          drawerLabel: 'Text Embeddings',
+          title: 'Text Embeddings',
+        }}
+      />
     </Drawer>
   );
 }

apps/nlp/app/index.tsx

@@ -14,6 +14,9 @@ export default function Home() {
         <TouchableOpacity style={styles.button} onPress={() => router.navigate('tokenizer/')}>
           <Text style={styles.buttonText}>Tokenizer</Text>
         </TouchableOpacity>
+        <TouchableOpacity style={styles.button} onPress={() => router.navigate('text-embeddings/')}>
+          <Text style={styles.buttonText}>Text Embeddings</Text>
+        </TouchableOpacity>
       </View>
     </View>
   );