GrassNet 🌱

GrassNet is a neural network library built from scratch in Python.
It provides a lightweight API for defining and training feedforward neural networks without relying on external frameworks like PyTorch or TensorFlow — just NumPy.

This project is meant as both a learning tool and a minimal framework that shows how deep learning works under the hood — forward propagation, backpropagation, optimizers, and training loops.

---

visualization.mp4

Generated via visualize.py

---

Features

• Custom API for defining feedforward networks
• Layers: Linear (Dense), activation functions (ReLU, Softplus, etc.)
• Loss functions: Mean Squared Error (MSE)
• Optimizer: AdamW (decoupled weight decay) implemented from scratch
• Example: sequence-to-sequence prediction of sin(x) in main.py

---

Example Usage

from core.layers import Linear, Softplus
from core.network import Network
import numpy as np

# Define network
input_length = 5
n = Network([
    Linear(input_length,16),
    Softplus(),
    Linear(16,16),
    Softplus(),
    Linear(16,16),
    Softplus(),
    Linear(16,1)
])   

n.MSELoss()       # Set loss function
n.InitOptimizer() # Initialize AdamW

n.tracked_run(X)  # A single run where layer outputs are tracked
n.loss_to(Y)      # Returns the loss of the latest run's output compared to Y
n.backprop(Y)     # Do one backpropagation iteration based on Y
n.run(X)          # Standard run for inference

Full example found in main.py

main.py implements a sequence-to-sequence prediction example:

1. Generates training data from sin(x)

2. Trains the model using mini-batches and AdamW

3. Uses the trained model to iteratively predict the sine sequence

4. Plots both the original and predicted wave

There is also a variant of main.py found in regression.py which implements regression prediction instead of sequence-to-sequence prediction.

Project structure

.
├── core/
│   ├── layers.py     # Linear + activations
│   ├── network.py    # Network API (forward, backprop, loss, optimizer)
│   └── optim.py      # AdamW optimizer
├── data_tools.py     # Utility functions for batching
├── main.py           # Example: sin(x) sequence prediction
├── regression.py     # Same as main.py but is regression instead
└── README.md

Try it out!

git clone https://github.com/CoffeeCoder1015/grassnet.git
cd grassnet
pip install numpy matplotlib
python main.py