|
|
--- |
|
|
language: |
|
|
- en |
|
|
tags: |
|
|
- pytorch |
|
|
- unified-model |
|
|
- multi-modal |
|
|
- image-captioning |
|
|
- text-to-image |
|
|
- reasoning |
|
|
license: mit |
|
|
--- |
|
|
|
|
|
# Working Unified Multi-Model (.pt) |
|
|
|
|
|
A complete unified PyTorch model that delegates to specialized child models for different AI tasks. |
|
|
|
|
|
## 🚀 Features |
|
|
|
|
|
- **Single .pt file** containing all capabilities |
|
|
- **True model delegation** to specialized child models |
|
|
- **Unified reasoning** and routing |
|
|
- **Production-ready** deployment |
|
|
|
|
|
## 📦 Model Components |
|
|
|
|
|
- **Base Reasoning Model**: `distilgpt2` (~300MB) |
|
|
- **Image Captioning Model**: `BLIP` (~990MB) |
|
|
- **Text-to-Image Model**: `Stable Diffusion v1.5` |
|
|
- **Task Classifiers**: Routing and confidence scoring |
|
|
- **Embeddings**: Task type embeddings |
|
|
|
|
|
## 🎯 Capabilities |
|
|
|
|
|
1. **Text Processing**: Q&A, summarization, text generation |
|
|
2. **Image Captioning**: Describe images using BLIP model |
|
|
3. **Text-to-Image**: Generate images using Stable Diffusion |
|
|
4. **Reasoning**: Step-by-step reasoning tasks |
|
|
|
|
|
## 📊 Model Size |
|
|
|
|
|
- **File Size**: 1.26 GB |
|
|
- **Total Parameters**: ~1.2B parameters |
|
|
- **Architecture**: Unified PyTorch model |
|
|
|
|
|
## 🔧 Usage |
|
|
|
|
|
```python |
|
|
import torch |
|
|
from working_complete_unified_model_pt import WorkingUnifiedMultiModelPT |
|
|
|
|
|
# Load the model |
|
|
model = WorkingUnifiedMultiModelPT.load_model("working_unified_multi_model.pt") |
|
|
|
|
|
# Process different types of requests |
|
|
result = model.process("What is machine learning?") |
|
|
print(f"Task: {result['task_type']}") |
|
|
print(f"Output: {result['output']}") |
|
|
|
|
|
result = model.process("Generate an image of a peaceful forest") |
|
|
print(f"Task: {result['task_type']}") |
|
|
print(f"Output: {result['output']}") |
|
|
``` |
|
|
|
|
|
## 🏗️ Architecture |
|
|
|
|
|
The model uses a unified architecture where: |
|
|
1. **Parent LLM** (distilgpt2) analyzes requests and routes to appropriate child models |
|
|
2. **Child Models** handle specialized tasks: |
|
|
- BLIP for image captioning |
|
|
- Stable Diffusion for text-to-image generation |
|
|
- Base model for text processing and reasoning |
|
|
|
|
|
## 🎉 Key Innovations |
|
|
|
|
|
- **Single .pt file** for all capabilities |
|
|
- **True delegation** to specialized models |
|
|
- **Unified interface** like DeepSeek |
|
|
- **Portable** across environments |
|
|
- **Production-ready** deployment |
|
|
|
|
|
## 📄 License |
|
|
|
|
|
MIT License |
|
|
|
|
|
## 🤝 Contributing |
|
|
|
|
|
This model demonstrates the future of AI - unified, portable, and intelligent models that can handle multiple tasks through intelligent delegation. |
|
|
|
