Neural Predictor

Lightweight feedforward neural network for regression, implemented in pure Python with NumPy. Vectorized training, numerically stable activations, and a simple CLI for experimentation.

Installation

git clone https://github.com/Amarifields/neaural-preditor.git
cd neaural-preditor
pip install -r requirements.txt

CLI

python main.py --input-size 3 --hidden-size 8 --epochs 1000 --lr 0.01 --train-samples 200 --test-samples 50 --seed 42

• --quiet: suppress progress
• --no-interactive: skip interactive prompt after evaluation

Example

python main.py --epochs 600 --train-samples 300 --test-samples 60

Sample output (truncated):

training
epoch 0 mae 0.090321
epoch 100 mae 0.028114
epoch 200 mae 0.019873
...
evaluation
test_mae 0.017442

Architecture

• Input: configurable
• Hidden: single layer, configurable width
• Output: single neuron
• Activation: sigmoid
• Loss monitor: mean absolute error

Requirements

• Python 3.8+
• NumPy >= 1.21.0

License

MIT