Cognix AI Library 🚀

Cognix is a powerful yet simplified Python library for model management, inference, and fine-tuning. It provides a high-level API to interact with Hugging Face models effortlessly, built for developers who want to integrate AI without the boilerplate.

Installation

pip install cognix

Features

• Simplified Model Loading: Load models for Chat, Vision, or Classification with one line.
• Robust Error Handling: Improved fallback mechanisms for ImportError and model compatibility.
• Easy Fine-Tuning: Built-in support for LoRA/QLoRA for efficient training.
• Hugging Face Integration: Secure login and seamless model downloading.
• Unified Interface: generate() for LLMs, predict() for Vision/Classification.

Quick Start

1. Login to Hugging Face

import cognix

# Method 1: Explicit token
cognix.login("your_hf_token")

# Method 2: Uses HUGGING_FACE_HUB_TOKEN environment variable
cognix.login()

2. Load and Use an LLM (Chat)

import cognix

try:
    # Load a model for chat (supports 4bit/8bit quantization)
    model = cognix.load_model("gpt2", task="chat")
    
    # Generate text
    response = model.generate("The future of AI is")
    print(f"AI: {response}")
except Exception as e:
    print(f"Error loading model: {e}")

3. Image Classification / Vision (Improved)

Cognix now handles older versions of transformers gracefully.

import cognix

# Load a vision model (e.g., BLIP or ViT)
vision_model = cognix.load_model("Salesforce/blip-image-captioning-base", task="vision")

# Predict from URL, local path, or PIL Image
result = vision_model.predict("https://example.com/image.jpg")
print(f"Description: {result}")

Error Handling & Edge Cases

Cognix is designed to fail gracefully. Here are common scenarios handled:

Fine-Tuning with LoRA

import cognix

# Load model and dataset
model = cognix.load_model("gpt2", task="chat", quantization="4bit")
dataset = cognix.load_dataset("my_data.json")

# Initialize trainer
trainer = cognix.Trainer(model, dataset, output_dir="./my_fine_tuned_model")

# Train with built-in validation!
if trainer.train(epochs=1, lr=5e-5, batch_size=4):
    trainer.save()
else:
    print("Training failed. Check logs for details.")

License

MIT

Contributing

Visit our GitHub: github.com/golemgamer/cognix