README.md

---
license: apache-2.0
library_name: transformers
pipeline_tag: zero-shot-image-classification
language:
- zh
- en
datasets:
- BM-6B
- ImageNet
- ImageNet-CN
- Flickr30K
- Flickr30K-CN
- COCO-CN
tags:
- onnx
- feature-extraction
- multimodal
- image-text-retrieval
- zero-shot-image-classification
- bilingual
- chinese
- english
- vision-language
- custom-code
model-index:
- name: M2-Encoder-0.4B
  results:
  - task:
      type: zero-shot-image-classification
      name: Zero-Shot Image Classification
    dataset:
      name: ImageNet
      type: ImageNet
    metrics:
    - type: accuracy
      value: 78.5
      name: Top-1 Accuracy
  - task:
      type: zero-shot-image-classification
      name: Zero-Shot Image Classification
    dataset:
      name: ImageNet-CN
      type: ImageNet-CN
    metrics:
    - type: accuracy
      value: 69.1
      name: Top-1 Accuracy
  - task:
      type: image-text-retrieval
      name: Zero-Shot Image-Text Retrieval
    dataset:
      name: Flickr30K
      type: Flickr30K
    metrics:
    - type: mean_recall
      value: 94.5
      name: MR
  - task:
      type: image-text-retrieval
      name: Zero-Shot Image-Text Retrieval
    dataset:
      name: COCO
      type: COCO
    metrics:
    - type: mean_recall
      value: 75.2
      name: MR
  - task:
      type: image-text-retrieval
      name: Zero-Shot Image-Text Retrieval
    dataset:
      name: Flickr30K-CN
      type: Flickr30K-CN
    metrics:
    - type: mean_recall
      value: 91.2
      name: MR
  - task:
      type: image-text-retrieval
      name: Zero-Shot Image-Text Retrieval
    dataset:
      name: COCO-CN
      type: COCO-CN
    metrics:
    - type: mean_recall
      value: 87.8
      name: MR
---

# M2-Encoder-0.4B

`M2-Encoder-0.4B` is a Hugging Face export of the bilingual vision-language foundation model from the paper [M2-Encoder: Advancing Bilingual Image-Text Understanding by Large-scale Efficient Pretraining](https://arxiv.org/abs/2401.15896).

It supports Chinese-English image-text retrieval, zero-shot image classification, `transformers` remote-code loading, ONNXRuntime inference, and Hugging Face Inference Endpoints via the bundled `handler.py`.

This is the smallest published M2-Encoder variant and is the best starting point for CPU demos, Spaces, and lightweight retrieval services.

![Benchmark overview](https://raw.githubusercontent.com/alipay/Ant-Multi-Modal-Framework/main/prj/M2_Encoder/pics/effect.png)

## Links

- Paper: https://arxiv.org/abs/2401.15896
- Official code: https://github.com/alipay/Ant-Multi-Modal-Framework/tree/main/prj/M2_Encoder
- ModelScope source model: `M2Cognition/M2-Encoder`
- Hugging Face repo: `malusama/M2-Encoder-0.4B`
- Hugging Face Space demo: https://huggingface.co/spaces/malusama/M2-Encoder-0.4B-Space

## At A Glance

| Item | Value |
| --- | --- |
| Variant | `M2-Encoder-0.4B` |
| Languages | Chinese, English |
| Embedding dimension | `768` |
| Image size | `224` |
| Main tasks | Image-text retrieval, zero-shot image classification, bilingual feature extraction |
| Weight format | `safetensors` |
| ONNX export | `onnx/text_encoder.onnx`, `onnx/image_encoder.onnx` |

## Files In This Repo

- `m2_encoder_0.4B.safetensors`: main `transformers` weight file
- `onnx/text_encoder.onnx`: text embedding encoder
- `onnx/image_encoder.onnx`: image embedding encoder
- `examples/run_onnx_inference.py`: runnable ONNX example
- `handler.py`: custom handler for Hugging Face Inference Endpoints

## Transformers Usage

### ModelScope-equivalent scoring

The original ModelScope sample computes probabilities from raw normalized embedding dot products:

```python
from transformers import AutoModel, AutoProcessor

repo_id = "malusama/M2-Encoder-0.4B"

model = AutoModel.from_pretrained(repo_id, trust_remote_code=True)
processor = AutoProcessor.from_pretrained(repo_id, trust_remote_code=True)

text_inputs = processor(
    text=["杰尼龟", "妙蛙种子", "小火龙", "皮卡丘"],
    return_tensors="pt",
)
image_inputs = processor(images="pokemon.jpeg", return_tensors="pt")

text_outputs = model(**text_inputs)
image_outputs = model(**image_inputs)

probs = (image_outputs.image_embeds @ text_outputs.text_embeds.t()).softmax(dim=-1)
print(probs)
```

### CLIP-style logits

`model(**inputs)` also returns `logits_per_image` and `logits_per_text`, which use the model's learned `logit_scale`.
Those logits are useful, but they are not the same computation as the raw dot product used in the original ModelScope demo.

## ONNXRuntime Usage

This repo also includes two ONNX exports:

- `onnx/text_encoder.onnx`
- `onnx/image_encoder.onnx`

Minimal example:

```python
import importlib
import json
import os
import sys

import onnxruntime as ort
from huggingface_hub import snapshot_download
from PIL import Image

repo_id = "malusama/M2-Encoder-0.4B"
model_dir = snapshot_download(repo_id=repo_id)
sys.path.insert(0, model_dir)

tokenizer_config = json.load(open(os.path.join(model_dir, "tokenizer_config.json"), "r", encoding="utf-8"))
GLMChineseTokenizer = importlib.import_module("tokenization_glm").GLMChineseTokenizer
M2EncoderImageProcessor = importlib.import_module("image_processing_m2_encoder").M2EncoderImageProcessor

tokenizer = GLMChineseTokenizer(
    vocab_file=os.path.join(model_dir, "sp.model"),
    eos_token=tokenizer_config.get("eos_token"),
    pad_token=tokenizer_config.get("pad_token"),
    cls_token=tokenizer_config.get("cls_token"),
    mask_token=tokenizer_config.get("mask_token"),
    unk_token=tokenizer_config.get("unk_token"),
)
image_processor = M2EncoderImageProcessor.from_pretrained(model_dir)

text_inputs = tokenizer(
    ["杰尼龟", "妙蛙种子", "小火龙", "皮卡丘"],
    padding="max_length",
    truncation=True,
    max_length=52,
    return_special_tokens_mask=True,
    return_tensors="np",
)
image_inputs = image_processor(Image.open("pokemon.jpeg").convert("RGB"), return_tensors="np")

text_session = ort.InferenceSession(
    os.path.join(model_dir, "onnx", "text_encoder.onnx"),
    providers=["CPUExecutionProvider"],
)
image_session = ort.InferenceSession(
    os.path.join(model_dir, "onnx", "image_encoder.onnx"),
    providers=["CPUExecutionProvider"],
)

text_embeds = text_session.run(
    None,
    {
        "input_ids": text_inputs["input_ids"],
        "attention_mask": text_inputs["attention_mask"],
    },
)[0]
image_embeds = image_session.run(
    None,
    {"pixel_values": image_inputs["pixel_values"]},
)[0]
```

Runnable script:

```bash
python examples/run_onnx_inference.py   --image pokemon.jpeg   --text 杰尼龟 妙蛙种子 小火龙 皮卡丘
```

## Inference Endpoints

This repo includes a `handler.py` for Hugging Face Inference Endpoints custom deployments.

Example request body:

```json
{
  "inputs": {
    "text": ["杰尼龟", "妙蛙种子", "小火龙", "皮卡丘"],
    "image": "https://clip-cn-beijing.oss-cn-beijing.aliyuncs.com/pokemon.jpeg"
  },
  "parameters": {
    "return_probs": true,
    "return_logits": false
  }
}
```

Example response fields:

- `text_embedding`
- `image_embedding`
- `scores`
- `probs`
- `logits_per_image` when `return_logits=true`

## Evaluation Summary

According to the official project README and paper, the M2-Encoder series is trained on the bilingual BM-6B corpus and evaluated on:

- ImageNet
- ImageNet-CN
- Flickr30K
- Flickr30K-CN
- COCO-CN

The official project reports that the M2-Encoder family sets strong bilingual retrieval and zero-shot classification results, and that the 10B variant reaches 88.5 top-1 on ImageNet and 80.7 top-1 on ImageNet-CN in the zero-shot setting. See the paper for exact cross-variant comparisons.

The structured `model-index` metadata in this card is taken from the official paper tables for this released variant. On the Hugging Face page, those results should surface in the evaluation panel once the metadata is parsed.

![Benchmark overview](https://raw.githubusercontent.com/alipay/Ant-Multi-Modal-Framework/main/prj/M2_Encoder/pics/effect.png)

## Notes

- This is a Hugging Face remote-code adapter, not a native `transformers` implementation merged upstream.
- `trust_remote_code=True` is required for `AutoModel` and `AutoProcessor`.
- This repo is intended for retrieval, classification, and embedding use cases, not text generation.
- The Hub export has been numerically checked against the official implementation for the published demo workflow.

## Citation

```bibtex
@misc{guo2024m2encoder,
  title={M2-Encoder: Advancing Bilingual Image-Text Understanding by Large-scale Efficient Pretraining},
  author={Qingpei Guo and Furong Xu and Hanxiao Zhang and Wang Ren and Ziping Ma and Lin Ju and Jian Wang and Jingdong Chen and Ming Yang},
  year={2024},
  url={https://arxiv.org/abs/2401.15896}
}
```