Spaces:

ping98k
/

Embedding-Playground

Running

App Files Files Community

ping98k commited on Jun 28

Commit

07f99cb

1 Parent(s): 293a305

Update README and index.html for embedding playground functionality and usage instructions

Browse files

Files changed (2) hide show

README.md +25 -3
index.html +74 -20

README.md CHANGED Viewed

@@ -5,8 +5,30 @@ colorFrom: blue
 colorTo: yellow
 sdk: static
 pinned: false
-models:
-- Xenova/detr-resnet-50
 ---
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

 colorTo: yellow
 sdk: static
 pinned: false
 ---
+# Embedding WebGPU Playground
+This is a browser-based playground for exploring text embeddings and group similarity using WebGPU and ONNX models.
+## How it works
+- **Input text** in the textarea. Use single newlines (`\n`) to separate lines within a group, and triple newlines (`\n\n\n`) to separate groups.
+- For each line, the following prompt is used:
+  `Instruct: Given a textual input sentence, retrieve relevant categories that best describe it.\nQuery:{your sentence}`
+- All lines in a group are embedded in a single batch call, with pooling set to `mean` and normalization enabled.
+- Each group is averaged to a single embedding.
+- Cosine similarity is calculated between all group embeddings, resulting in a group-by-group similarity matrix.
+- The similarity matrix is visualized as a heatmap using Plotly (color range locked to 0–1).
+## Tech stack
+- [@huggingface/transformers](https://www.npmjs.com/package/@huggingface/transformers) (ESM, WebGPU)
+- [ONNX Qwen3-Embedding-0.6B-ONNX](https://huggingface.co/onnx-community/Qwen3-Embedding-0.6B-ONNX)
+- [Plotly.js](https://plotly.com/javascript/) (UMD)
+## Usage
+1. Enter or paste your text in the textarea.
+2. Separate groups with triple newlines.
+3. Click **Run** to compute and visualize group similarities.
+---

index.html CHANGED Viewed

@@ -1,29 +1,83 @@
-<!DOCTYPE html>
-<html lang="en">
 <head>
-    <meta charset="UTF-8" />
-    <link rel="stylesheet" href="style.css" />
-    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
-    <title>Transformers.js - Object Detection</title>
 </head>
 <body>
-    <h1>Object Detection w/ 🤗 Transformers.js</h1>
-    <label id="container" for="upload">
-        <svg width="25" height="25" viewBox="0 0 25 25" fill="none" xmlns="http://www.w3.org/2000/svg">
-            <path fill="#000"
-                d="M3.5 24.3a3 3 0 0 1-1.9-.8c-.5-.5-.8-1.2-.8-1.9V2.9c0-.7.3-1.3.8-1.9.6-.5 1.2-.7 2-.7h18.6c.7 0 1.3.2 1.9.7.5.6.7 1.2.7 2v18.6c0 .7-.2 1.4-.7 1.9a3 3 0 0 1-2 .8H3.6Zm0-2.7h18.7V2.9H3.5v18.7Zm2.7-2.7h13.3c.3 0 .5 0 .6-.3v-.7l-3.7-5a.6.6 0 0 0-.6-.2c-.2 0-.4 0-.5.3l-3.5 4.6-2.4-3.3a.6.6 0 0 0-.6-.3c-.2 0-.4.1-.5.3l-2.7 3.6c-.1.2-.2.4 0 .7.1.2.3.3.6.3Z">
-            </path>
-        </svg>
-        Click to upload image
-        <label id="example">(or try example)</label>
-    </label>
-    <label id="status">Loading model...</label>
-    <input id="upload" type="file" accept="image/*" />
-    <script src="index.js" type="module"></script>
 </body>
 </html>

+<!doctype html>
+<html>
 <head>
+    <meta charset="utf-8">
+    <title>WebGPU Embedding Playground</title>
+    <style>
+        body {
+            font-family: sans-serif;
+            margin: 0;
+            padding: 20px
+        }
+        textarea {
+            width: 100%;
+            height: 200px
+        }
+        button {
+            padding: 10px 20px;
+            margin-top: 10px
+        }
+        #plot {
+            width: 100%;
+            height: 600px
+        }
+    </style>
 </head>
 <body>
+    <h1>Embedding Similarity Heatmap</h1>
+    <textarea id="input"></textarea>
+    <button id="run">Run</button>
+    <div id="plot"></div>
+    <script src="https://cdn.plot.ly/plotly-2.32.0.min.js"></script>
+    <script type="module">
+        import { pipeline } from "https://cdn.jsdelivr.net/npm/@huggingface/transformers@3.5.2";
+        const embed = await pipeline(
+            "feature-extraction",
+            "onnx-community/Qwen3-Embedding-0.6B-ONNX",
+            { device: "webgpu", dtype: "q4f16" },
+        );
+        const task = "Given a textual input sentence, retrieve relevant categories that best describe it.";
+        document.getElementById("run").onclick = async () => {
+            const text = document.getElementById("input").value;
+            const groups = text.split(/\n{3,}/);
+            const groupEmbeddings = [];
+            for (const g of groups) {
+                const lines = g.split(/\n/).filter(x => x.trim() != "");
+                const prompts = lines.map(s => `Instruct: ${task}\nQuery:${s}`);
+                const out = await embed(prompts, { pooling: "mean", normalize: true });
+                const embeddings = typeof out.tolist === 'function' ? out.tolist() : out.data;
+                const dim = embeddings[0].length;
+                const avg = new Float32Array(dim);
+                for (const e of embeddings) { for (let i = 0; i < dim; i++) avg[i] += e[i]; }
+                for (let i = 0; i < dim; i++) avg[i] /= embeddings.length;
+                groupEmbeddings.push(avg);
+            }
+            const n = groupEmbeddings.length;
+            const sim = [];
+            for (let i = 0; i < n; i++) {
+                const row = [];
+                for (let j = 0; j < n; j++) {
+                    let dot = 0, na = 0, nb = 0;
+                    for (let k = 0; k < groupEmbeddings[i].length; k++) {
+                        dot += groupEmbeddings[i][k] * groupEmbeddings[j][k];
+                        na += groupEmbeddings[i][k] ** 2;
+                        nb += groupEmbeddings[j][k] ** 2;
+                    }
+                    row.push(dot / Math.sqrt(na * nb));
+                }
+                sim.push(row);
+            }
+            const data = [{ z: sim, type: "heatmap", colorscale: "Viridis", zmin: 0, zmax: 1 }];
+            Plotly.newPlot("plot", data, { xaxis: { title: "Group" }, yaxis: { title: "Group" } });
+        };
+    </script>
 </body>
 </html>