pynet-ai 1.5.1

A Python neural network library for simple networks and educational use.

PyNet 1.5.1

A simple educational neural network library built from scratch in Python.

PyNet focuses on understanding how neural networks work internally, including layers, activations, loss functions, backpropagation, optimizers, weight initialization, and model serialization.

Features

Layers

• Dense (fully connected) layers
• Dropout layers for reducing overfitting through neuron regularization

Activations

• ReLU
• LeakyReLU
• Sigmoid
• Tanh
• Softmax
• ELU
• Swish
• GELU

Loss Functions

• MSELoss
• CrossEntropyLoss
• MAELoss
• HuberLoss
• BinaryCrossEntropyLoss

Optimizers

• SGD
• Momentum
• Adam

Weight Initialization

• He
• Xavier
• RandomNormal
• Zeros

Utilities

• Metrics
• Model saving/loading

Quick Start Guide

1. Install PyNet with pip install pynet-ai
2. Install PyGame with pip install pygame (for the example)
3. Paste example program below

import random
import pygame
from pynet import Network, Dense, LeakyReLU, Tanh, MSELoss, Adam, Xavier

WIDTH, HEIGHT = (480, 360)
FPS = 60
EPOCHS = 10000
DELAY = 1000
LEARNING_RATE = 0.0001
ACCURACY_TRIES = 50
SPEED = 100
RADIUS = 10


def normalize(x, y):
    length = (x * x + y * y) ** 0.5
    if length < 0.0001:
        return (0.0, 0.0)
    return (x / length, y / length)


def random_direction():
    return normalize(random.uniform(-1, 1), random.uniform(-1, 1))


def direction_to(x, y, target_x, target_y):
    return normalize(target_x - x, target_y - y)


def calculate_accuracy(network):
    total = 0.0
    for _ in range(ACCURACY_TRIES):
        x, y = random_direction()
        predicted_x, predicted_y = network.predict([x, y])
        predicted_x, predicted_y = normalize(predicted_x, predicted_y)
        total += (predicted_x * x + predicted_y * y + 1) / 2
    return total / ACCURACY_TRIES * 100


def train_network(network):
    loss_function = MSELoss()
    optimizer = Adam(learning_rate=LEARNING_RATE)
    for epoch in range(EPOCHS):
        x, y = random_direction()
        prediction = network.forward([x, y])
        loss = loss_function.forward(prediction, [x, y])
        gradient = loss_function.backward(prediction, [x, y])
        network.backward(gradient)
        optimizer.step(network.modules)
        if (epoch + 1) % DELAY == 0:
            print(f"Epoch {epoch + 1}/{EPOCHS} Loss: {loss:.5f} Accuracy: {calculate_accuracy(network):.2f}%")


def main():
    pygame.init()
    screen = pygame.display.set_mode((WIDTH, HEIGHT), pygame.RESIZABLE)
    pygame.display.set_caption("PyNet Follow Mouse")
    font = pygame.font.SysFont(None, 28)
    small_font = pygame.font.SysFont(None, 20)
    network = Network(Dense(2, 8, initializer=Xavier()), LeakyReLU(), Dense(8, 8, initializer=Xavier()), LeakyReLU(), Dense(8, 2, initializer=Xavier()), Tanh())
    screen.fill((20, 20, 20))
    text = font.render("Training network...", True, "white")
    screen.blit(text, text.get_rect(center=(WIDTH // 2, HEIGHT // 2)))
    pygame.display.flip()
    train_network(network)
    print(network.summary())
    accuracy = calculate_accuracy(network)
    clock = pygame.time.Clock()
    player_position = [WIDTH / 2, HEIGHT / 2]
    windowed_size = (WIDTH, HEIGHT)
    fullscreen = False
    paused = False
    running = True
    while running:
        delta_time = clock.tick(FPS) / 1000.0
        for event in pygame.event.get():
            if event.type == pygame.QUIT:
                running = False
            elif event.type == pygame.KEYDOWN:
                if event.key == pygame.K_SPACE:
                    paused = not paused
                elif event.key == pygame.K_F11:
                    fullscreen = not fullscreen
                    if fullscreen:
                        screen = pygame.display.set_mode((0, 0), pygame.FULLSCREEN)
                    else:
                        screen = pygame.display.set_mode(windowed_size, pygame.RESIZABLE)
            elif event.type == pygame.VIDEORESIZE and (not fullscreen):
                windowed_size = (event.w, event.h)
                screen = pygame.display.set_mode(windowed_size, pygame.RESIZABLE)
        mouse_position = pygame.mouse.get_pos()
        screen_width, screen_height = screen.get_size()
        if not paused:
            target_direction = direction_to(player_position[0], player_position[1], mouse_position[0], mouse_position[1])
            output_x, output_y = network.predict(list(target_direction))
            move_x, move_y = normalize(output_x, output_y)
            player_position[0] += move_x * SPEED * delta_time
            player_position[1] += move_y * SPEED * delta_time
            player_position[0] = max(RADIUS, min(screen_width - RADIUS, player_position[0]))
            player_position[1] = max(RADIUS, min(screen_height - RADIUS, player_position[1]))
        screen.fill((20, 20, 20))
        pygame.draw.line(screen, (80, 80, 80), (int(player_position[0]), int(player_position[1])), mouse_position, 1)
        pygame.draw.circle(screen, (255, 70, 70), mouse_position, RADIUS)
        pygame.draw.circle(screen, (70, 230, 120), (int(player_position[0]), int(player_position[1])), RADIUS)
        screen.blit(font.render(f"Accuracy: {accuracy:.1f}%", True, "white"), (10, 10))
        screen.blit(small_font.render("Space: Pause", True, (180, 180, 180)), (10, 42))
        status = "PAUSED" if paused else "RUNNING"
        screen.blit(font.render(status, True, "white"), (screen_width - 120, 10))
        pygame.display.flip()
    pygame.quit()


if __name__ == "__main__":
    main()

This example program will train a neural network to follow the mouse pointer.

Note: No comments or blank lines have been provided in example program.

Installation

pip install pynet-ai