cognix-ai 0.1.1

A simplified AI library for model management, inference, and fine-tuning.

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:

Scenario	Behavior
Missing bitsandbytes	Automatically disables 4-bit/8-bit quantization and tries to load in full precision.
Invalid Image URL	Returns a clear error message instead of crashing.
Model Incompatibility	Falls back to generic AutoModel if specialized classes like AutoModelForVision2Seq fail.
Login Failure	Provides detailed feedback if the token is invalid or there's no connection.
Missing PEFT	Informs the user that peft is required only when trying to use the Trainer.

---

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