---
license: apple-amlr
license_name: apple-ascl
license_link: https://github.com/apple/ml-mobileclip/blob/main/LICENSE_weights_data
library_name: mobileclip
---

# 📸 MobileCLIP-B Zero-Shot Image Classifier
### Hugging Face Inference Endpoint

> **Production-ready wrapper** around Apple’s MobileCLIP-B checkpoint.
> Handles image → text similarity in a single fast call.

---

## 📑 Sidebar

- [Features](#-features)
- [Repository layout](#-repository-layout)
- [Quick start (local smoke-test)](#-quick-start-local-smoke-test)
- [Calling the deployed endpoint](#-calling-the-deployed-endpoint)
- [How it works](#-how-it-works)
- [Updating the label set](#-updating-the-label-set)
- [License](#-license)

---

## ✨ Features
|                              | This repo |
|------------------------------|-----------|
| **Model**                    | MobileCLIP-B (`datacompdr` checkpoint) |
| **Branch fusion**            | `reparameterize_model` baked in |
| **Mixed-precision**          | FP16 on GPU, FP32 on CPU |
| **Pre-computed text feats**  | One-time encoding of prompts in `items.json` |
| **Per-request work**         | _Only_ image decoding → encode_image → softmax |
| **Latency (A10G)**           | < 30 ms once the image arrives |

---

## 📁 Repository layout

| Path               | Purpose                                                          |
|--------------------|------------------------------------------------------------------|
| `handler.py`       | HF entry-point (loads model + text cache, serves requests)       |
| `reparam.py`       | 60-line stand-alone copy of Apple’s `reparameterize_model`       |
| `requirements.txt` | Minimal dep set (`torch`, `torchvision`, `open-clip-torch`)      |
| `items.json`       | Your label set (`id`, `name`, `prompt` per line)                 |
| `README.md`        | This document                                                    |

---

## 🚀 Quick start (local smoke-test)

```bash
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

python - <<'PY'
import base64, json, handler, pathlib
app = handler.EndpointHandler()

img_b64 = base64.b64encode(pathlib.Path("tests/cat.jpg").read_bytes()).decode()
print(app({"inputs": {"image": img_b64}})[:5])   # top-5 classes
PY
```

---

## 🌐 Calling the deployed endpoint

```bash
ENDPOINT="https://<your-endpoint>.aws.endpoints.huggingface.cloud"
TOKEN="hf_xxxxxxxxxxxxxxxxx"
IMG="cat.jpg"

python - <<'PY'
import base64, json, os, requests, sys
url   = os.environ["ENDPOINT"]
token = os.environ["TOKEN"]
img   = sys.argv

payload = {
    "inputs": {
        "image": base64.b64encode(open(img, "rb").read()).decode()
    }
}
resp = requests.post(
    url,
    headers={
        "Authorization": f"Bearer {token}",
        "Content-Type":  "application/json",
        "Accept":        "application/json",
    },
    json=payload,
    timeout=60,
)
print(json.dumps(resp.json()[:5], indent=2))
PY
$IMG
```

*Response example*

```json
[
  { "id": 23, "label": "cat",         "score": 0.92 },
  { "id": 11, "label": "tiger cat",   "score": 0.05 },
  { "id": 48, "label": "siamese cat", "score": 0.02 }
]
```

---

## ⚙️ How it works

1. **Startup (runs once per replica)**

   * Downloads / loads MobileCLIP-B (`datacompdr`).
   * Fuses MobileOne branches via `reparam.py`.
   * Reads `items.json` and encodes every prompt → `[N,512]` tensor.

2. **Per request**

   * Decodes base-64 JPEG/PNG.
   * Applies OpenCLIP preprocessing (224 × 224 center-crop + normalise).
   * Encodes the image, normalises, computes cosine similarity vs. cached text matrix.
   * Returns sorted `[{id, label, score}, …]`.

---

## 🔄 Updating the label set

Simply edit `items.json`, push, and redeploy.

```json
[
  { "id": 0, "name": "cat", "prompt": "a photo of a cat" },
  { "id": 1, "name": "dog", "prompt": "a photo of a dog" }
]
```

No code changes are required; the handler re-encodes prompts at start-up.

---

## ⚖️ License

* **Weights / data** — Apple AMLR (see [`LICENSE_weights_data`](./LICENSE_weights_data))
* **This wrapper code** — MIT

---

<div align="center"><sub>Maintained with ❤️ by Your-Team — Aug 2025</sub></div>