@myvillage/cli 2.16.0 → 2.16.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@myvillage/cli",
-  "version": "2.16.0",
+  "version": "2.16.1",
   "description": "MyVillageOS CLI for community developers",
   "type": "module",
   "bin": {

package/src/utils/image-backbone.js

@@ -14,7 +14,7 @@
  * datasets/models never pulls the heavy runtime in.
  */
 import { readFileSync, existsSync } from 'fs';
-import { join, dirname, extname } from 'path';
+import { join, dirname, extname, basename } from 'path';
 import { fileURLToPath } from 'url';
 import { createRequire } from 'module';
 
@@ -27,6 +27,16 @@ function requirePkg(name) {
 export const BACKBONE_ID = 'mobilenet-v3-small';
 export const INPUT_SIZE = 224;
 export const EMBED_DIM = 1024;
+/**
+ * Pictures up to this size are embedded exactly as lab-runtime embeds them
+ * (its MAX_MEGAPIXELS), so a shared model reads them the same way. Anything
+ * bigger, like a 48 MP phone photo, is box-shrunk to fit first: it keeps the
+ * float copy small and averages the detail the 224×224 resize would skip.
+ */
+const MAX_DIRECT_PIXELS = 16 * 1000 * 1000;
+/** jpeg-js's working memory cap; 512 MB stops short of 48 MP phone photos.
+ *  Its 100 MP resolution cap still applies. */
+const JPEG_MEMORY_MB = 2048;
 
 const BACKBONE_DIR = join(
   dirname(fileURLToPath(import.meta.url)),
@@ -99,18 +109,66 @@ function decodeImage(path) {
   if (!existsSync(path)) throw new Error(`Image not found: ${path}`);
   const buf = readFileSync(path);
   const ext = extname(path).toLowerCase();
-  if (ext === '.png') {
-    // Lazy require keeps these decoders out of the hot path for text/tabular.
-    const { PNG } = requirePkg('pngjs');
-    const png = PNG.sync.read(buf);
-    return { width: png.width, height: png.height, data: png.data };
+  if (ext !== '.png' && ext !== '.jpg' && ext !== '.jpeg') {
+    throw new Error(`Unsupported image type "${ext}". Use .jpg, .jpeg, or .png.`);
   }
-  if (ext === '.jpg' || ext === '.jpeg') {
+  try {
+    if (ext === '.png') {
+      // Lazy require keeps these decoders out of the hot path for text/tabular.
+      const { PNG } = requirePkg('pngjs');
+      const png = PNG.sync.read(buf);
+      return { width: png.width, height: png.height, data: png.data };
+    }
     const jpeg = requirePkg('jpeg-js');
-    const img = jpeg.decode(buf, { useTArray: true, maxMemoryUsageInMB: 512 });
+    const img = jpeg.decode(buf, { useTArray: true, maxMemoryUsageInMB: JPEG_MEMORY_MB });
     return { width: img.width, height: img.height, data: img.data };
+  } catch (e) {
+    const name = basename(path);
+    if (/maxResolutionInMP|maxMemoryUsageInMB/.test(e?.message || '')) {
+      throw new Error(`The picture "${name}" is too big to read. Make it smaller (under 100 megapixels) and add it again.`);
+    }
+    throw new Error(`The picture "${name}" couldn't be read. Try a different JPEG or PNG.`);
+  }
+}
+
+/**
+ * RGBA bytes → RGB floats in [0,1]. Pictures over MAX_DIRECT_PIXELS are
+ * averaged down by the smallest whole factor that fits (any leftover edge
+ * pixels are dropped); smaller ones are copied pixel for pixel.
+ */
+function toRgb({ width, height, data }) {
+  let k = 1;
+  while (Math.floor(width / k) * Math.floor(height / k) > MAX_DIRECT_PIXELS) k++;
+  const w = Math.floor(width / k);
+  const h = Math.floor(height / k);
+  const rgb = new Float32Array(w * h * 3);
+  if (k === 1) {
+    for (let i = 0, j = 0; i < data.length; i += 4, j += 3) {
+      rgb[j] = data[i] / 255;
+      rgb[j + 1] = data[i + 1] / 255;
+      rgb[j + 2] = data[i + 2] / 255;
+    }
+    return { rgb, w, h };
+  }
+  const scale = 1 / (k * k * 255);
+  for (let y = 0; y < h; y++) {
+    for (let x = 0; x < w; x++) {
+      let r = 0, g = 0, b = 0;
+      for (let dy = 0; dy < k; dy++) {
+        let i = ((y * k + dy) * width + x * k) * 4;
+        for (let dx = 0; dx < k; dx++, i += 4) {
+          r += data[i];
+          g += data[i + 1];
+          b += data[i + 2];
+        }
+      }
+      const j = (y * w + x) * 3;
+      rgb[j] = r * scale;
+      rgb[j + 1] = g * scale;
+      rgb[j + 2] = b * scale;
+    }
   }
-  throw new Error(`Unsupported image type "${ext}". Use .jpg, .jpeg, or .png.`);
+  return { rgb, w, h };
 }
 
 /**
@@ -120,17 +178,11 @@ function decodeImage(path) {
 export async function embedImage(path) {
   const tf = await import('@tensorflow/tfjs');
   const model = await getBackbone();
-  const { width, height, data } = decodeImage(path);
+  const { rgb, w, h } = toRgb(decodeImage(path));
 
   return tf.tidy(() => {
-    // RGBA → RGB float [0,1], shaped [h, w, 3], resized to the backbone input.
-    const rgb = new Float32Array(width * height * 3);
-    for (let i = 0, j = 0; i < data.length; i += 4, j += 3) {
-      rgb[j] = data[i] / 255;
-      rgb[j + 1] = data[i + 1] / 255;
-      rgb[j + 2] = data[i + 2] / 255;
-    }
-    const img = tf.tensor3d(rgb, [height, width, 3]);
+    // RGB float [0,1], shaped [h, w, 3], resized to the backbone input.
+    const img = tf.tensor3d(rgb, [h, w, 3]);
     const resized = tf.image.resizeBilinear(img, [INPUT_SIZE, INPUT_SIZE]);
     const batched = resized.expandDims(0); // [1, 224, 224, 3]
     const out = model.predict(batched); // [1, 1024]