gradzero 0.1.0

A lightweight deep learning framework from scratch

gradzero

A lightweight deep learning framework built from scratch with automatic differentiation.

Features

• Automatic Differentiation: Built-in autograd engine for computing gradients
• Tensor Operations: NumPy-based tensor with gradient tracking
• Neural Network Layers: Linear, ReLU, Sigmoid, and more
• Optimizers: SGD and Adam optimizers with momentum and weight decay support
• Easy to Use: Simple API similar to PyTorch

Installation

pip install gradzero

Quick Start

Creating Tensors with Factory Methods (实例化方法)

import gradzero as gz

# Create tensors using class method factories (实例化方法)
zeros = gz.Tensor.zeros((3, 3))          # Tensor filled with zeros
ones = gz.Tensor.ones((2, 4))            # Tensor filled with ones
randn = gz.Tensor.randn((3, 3))          # Random from normal distribution
rand = gz.Tensor.rand((2, 2))            # Random from uniform distribution
from_np = gz.Tensor.from_numpy(my_array) # From existing numpy array

# With gradient tracking
trainable = gz.Tensor.randn((3, 3), requires_grad=True)

Building a Neural Network

import gradzero as gz

# Define a model
model = gz.Sequential(
    gz.Linear(784, 128),
    gz.ReLU(),
    gz.Linear(128, 10),
)

# Loss and optimizer
criterion = gz.CrossEntropyLoss()
optimizer = gz.Adam(model.parameters(), lr=0.001)

# Training loop
for epoch in range(100):
    # Forward pass
    output = model(input_tensor)
    loss = criterion(output, target_tensor)

    # Backward pass
    model.zero_grad()
    loss.backward()

    # Update parameters
    optimizer.step()

API Reference

Tensor Factory Methods (实例化方法)

• Tensor.zeros(shape, dtype=None, requires_grad=False) - Create a tensor filled with zeros
• Tensor.ones(shape, dtype=None, requires_grad=False) - Create a tensor filled with ones
• Tensor.randn(shape, dtype=None, requires_grad=False) - Create a tensor with random values from standard normal distribution
• Tensor.rand(shape, dtype=None, requires_grad=False) - Create a tensor with random values from uniform distribution [0, 1)
• Tensor.from_numpy(array, requires_grad=False) - Create a tensor from a numpy array

Neural Network Layers

• Linear(in_features, out_features, bias=True) - Fully connected layer
• ReLU() - ReLU activation
• Sigmoid() - Sigmoid activation
• Sequential(*layers) - Container for sequential layers

Loss Functions

• MSELoss() - Mean squared error
• CrossEntropyLoss() - Cross entropy loss for classification

Optimizers

• SGD(parameters, lr=0.01, momentum=0.0, weight_decay=0.0) - Stochastic gradient descent
• Adam(parameters, lr=0.001, betas=(0.9, 0.999), eps=1e-8, weight_decay=0.0) - Adam optimizer

License

MIT License