modu-muse 0.1.1

Modular multimodal pipeline for vision-to-LLM integration

🧠 ModuMuse

Modular Multimodal Intelligence
Plug any Hugging Face LLM and vision encoder together via a learnable projector.
Supports zero-shot inference today, and adapter-based fine-tuning tomorrow.

---

🚀 Features

• 🔌 Plug-and-play architecture for combining LLMs and vision encoders
• 🧠 Supports popular models like Qwen, Mistral, LLaMA, CLIP, XCLIP, SAM
• 🧪 Zero-shot inference with learnable projector modules
• 🛠️ Adapter-based fine-tuning (coming soon)
• 📊 Easy benchmarking and visualization tools

---

📦 Installation

pip install modu-muse

---

🧬 Quick Start

from modu_muse import Pipeline

pipe = Pipeline(
    llm_name="mistralai/Mistral-7B-Instruct-v0.2",
    vision_name="openai/clip-vit-base-patch16"
)

result = pipe.infer("path/to/image.jpg", "Describe the scene.")
print(result)

---

🧠 Architecture

[Image/Video] → [Vision Encoder] → [Projector] → [LLM]

• Vision encoder extracts features
• Projector maps visual features to LLM-compatible embeddings
• LLM generates text conditioned on visual context

---

🛠️ Fine-Tuning (Coming Soon)

Train your own projector using paired image-text datasets:

python train_adapter.py \
  --model llm=Qwen1.5 vision=xclip \
  --dataset_path ./data/relevance_dataset \
  --output_dir ./checkpoints

---

📁 Project Structure

modu_muse/
├── pipeline.py          # Main multimodal pipeline
├── projector.py         # Vision-to-LLM projector
├── models/
│   ├── llm.py           # LLM loader
│   ├── vision.py        # Vision encoder loader
├── examples/
│   └── quick_start.py   # Demo script

---

🤝 Contributing

We welcome contributions! Whether it's new model support, training scripts, or documentation improvements—open a PR or start a discussion.

---

📜 License

This project is licensed under the MIT License.
© 2025 Wissem Elkarous

---

🌐 Resources

• 🔗 Hugging Face Transformers
• 🔗 CLIP & XCLIP Models
• 🔗 Qwen LLMs

---

ModuMuse: Where vision meets language.

```