foolnet 0.3.2

A package designed for the newer about the Neural Network.

FoolNet - v0.3.2

A package designed for the new hand about the Neural Network and Deep Learning.

Not used for the application but for the learning purpose with easy-reading code.

Introduction

Through the source code of this package, equipped with the awesome website we built, you're able to know more clearly about how the Neural Network work.

In this package, all the components of Neural Network, such as activation, loss function, BP algorithm are built from scratch, totally via the numpy! Therefore you can easily read the code and understand the meaning of this part without extra effort.

What's more, we created a website for the users to understand the mathematical logic behind the code, and why we implement it in that way, so users can understand some deeper principles, and not only 'How can I use it'.

The Package

Installation

Just use pip to install it from PypI. We maintain it for free!

pip install foolnet

Requirements: python >= 3.9 ; numpy >= 1.20.0; matplotlib >= 3.8.0

Usage

A usage example under version 0.3.2.

import numpy as np

import foolnet as fn

np.random.seed(42)

func = lambda x: np.log(1 / (np.sin(x) + 2)) # 生成具有非线性可分性的数据
dataset = fn.ClassificationDataset(x_dim=2, n_class=3, nums=999, nonlinear_fn=func)

# 两层网络
model = fn.Stack(
    fn.DenseLayer(2, 3),
    fn.ReLU(),
    fn.DenseLayer(3, 3),
    fn.Softmax()
)

lossfn = fn.CrossEntropyLoss()

for i in range(30):
    loss, acc = np.array([]), np.array([])
    for x, y in dataset.trainset:
        # 前向传播
        output = model(x)
        loss = np.append(loss, lossfn(output, y))

        # 反向传播
        lossfn.backward(output, y)
        model.backward(lossfn.dinputs)

        # 更新参数
        for pair in model.parameters():
            pair[0] += -0.5 * pair[1]

    # 在测试集上计算准确率
    for x, y in dataset.testset:
        output = model(x)
        acc = np.append(acc, np.mean(np.argmax(output, axis=1) == y))

    print(f"epoch{i} loss {np.mean(loss):.5f} acc {np.mean(acc):.5f}\n")

dataset.show()

The Website

The website is built by @CoderSerio, with the dumi framework.

Online

For online website, you may goto https://lazypool.github.io/foolnet/.

Offline

For you wish downloading the documents to your local, just clone this repository.

git clone '[e-mail]:lazypool/foolnet.git'
cd foolnet/docs/

Then run the next commands, supose that you have downloaded the npm.

$ npm install
$ npm run start

Now you should be able to visit the website at https://localhost:8000, which has the same content with the online website.