micro-tensor 🚀

A research-grade, first-principles deep learning framework built entirely in NumPy. Designed for educational transparency and architectural rigor.

✨ Features

• Vectorized Autograd: Support for N-dimensional tensors with automatic gradient broadcasting.
• CNN Suite: High-performance Conv2d (via as_strided), MaxPool2d, and Flatten.
• Transformer Engine: Multi-Head Self-Attention, Positional Encoding, and Layer Normalization.
• Optimizers: Integrated Adam and SGD implementations.
• Stability: Log-Sum-Exp stable CrossEntropyLoss.

📊 Benchmarks

• MNIST: 94%+ batch accuracy within 100 steps.
• GPT-Micro: Successfully performs autoregressive character-level generation.

🛠 Usage

# Clone and install
git clone [https://github.com/nibir-ai/micro-tensor](https://github.com/nibir-ai/micro-tensor)
cd micro-tensor
pip install -e .

# Run the Transformer demo
python examples/gpt_inference.py