kerze

A lightweight, PyTorch-like machine learning library built from scratch in Python.

kerze combines a small NumPy-inspired multidimensional array implementation with an automatic differentiation engine and a neural-network API.

The project is designed to make the internals of numerical computing and deep learning easier to understand by keeping the implementation small, explicit, and readable.

⚠️ Warning: kerze is primarily an experimental project, not a replacement for NumPy, PyTorch, or other production-grade numerical computing frameworks and it currently supports CPU execution only. There is no GPU/CUDA support.

Installation

Install with pip

pip install kerze

From Source

git clone https://github.com/ghpratik/kerze.git
cd kerze
pip install -e .

Overview

Kerze is organized into three main layers:

Array
  ↓
Tensor + Autograd
  ↓
neural nets

• Array — NumPy-like multidimensional array and mathematical operations.
• Tensor — wraps Array and adds automatic differentiation.
• nn — PyTorch-like neural-network building blocks.

🚀 Try Kerze

Run a complete XOR neural network example in Google Colab:

Example:

from kerze import Tensor, nn

model = nn.Sequential(
    nn.Linear(3, 8),
    nn.ReLU(),
    nn.Linear(8, 1),
)

x = Tensor([[1.0, 2.0, 3.0]], requires_grad=True)
y = model(x)

loss = y.mean()
loss.backward()

Features

• NumPy-inspired multidimensional Array
• Broadcasting
• Elementwise arithmetic operations
• Matrix multiplication
• Reductions such as sum, mean, and max
• Reshape, transpose, squeeze, and other shape operations
• Automatic differentiation
• Reverse-mode backpropagation
• Computation graphs
• Gradient accumulation
• Gradient broadcasting / unbroadcasting
• Basic mathematical operations
• Activation functions
• Neural-network modules
• Learnable Parameters
• Sequential models
• Linear layers
• Loss functions
• Functional neural-network API
• Parameter initialization
• Training/evaluation modes
• PyTorch-like API design

Neural Network API

Kerze provides a small PyTorch-inspired nn API:

nn.Module
nn.Parameter
nn.Linear

nn.ReLU
nn.GELU
nn.Sigmoid
nn.Tanh

nn.MSELoss
nn.NLLLoss
nn.CrossEntropyLoss

nn.Sequential

Functional operations are available through:

from kerze.nn import functional as F

Autograd

Kerze builds a computation graph during tensor operations and traverses it backwards during:

loss.backward()

Each operation defines its own backward rule, allowing gradients to propagate through arithmetic, reductions, activations, matrix multiplication, and other supported operations.

Scope & Limitations

Kerze is intentionally small and educational.

• CPU only
• No GPU/CUDA support
• No optimized C/C++ kernels
• Limited NumPy compatibility
• Limited neural-network layers
• No optimizers yet / limited optimizer support
• No distributed training
• No production-scale performance guarantees
• Some operations support a smaller set of shapes/features than NumPy/PyTorch

The implementation prioritizes clarity and understanding of the underlying mechanics over performance and completeness.

Documentation

Detailed documentation is available in docs/:

• Array
• Tensor & Autograd
• Neural Networks

Status

Kerze is an experimental/learning project and is actively evolving.

License

Kerze is licensed under the MIT License.