该仓库记录自己手写的两个用于数字识别的前馈型神经网络，分别是C++版本和Python版本的.

This repository reserves the ANN code, training data and testing data.

The training data are train.in and train.out(train.zip is train.in. It's so large~(102MB)). This training data contains 60000 examples. train.in has a matrix with the size of 60000*784, each row is stretched by an image with 28*28 pixels. 'train.out' has a vector with the size of 60000*1, which are the answers corresond to each row of train.in.

The testing data are test.in and test.out, which contains 10000 examples and has the same data format as the training data, but they are new data comparing to the training data.

Finally, 2022.1.30. I found the biggest problem that is the initial value, when I setup the initial value of w, b in [-1, 1] randomly. The accuracy improves to 94% in 10 minutes, it's perfect, that means my code is right!!!

使用Python实现感知器算法和前馈神经神经网络

Perceptron_main.py为感知器算法.

Input.py为数据读入的部分代码, timer.py为用于计时的部分代码, ANN.py为前馈神经网络核心代码, Main.py为主程序. 使用主函数训练前请先解压训练数据及样本的压缩包Data.rar至与代码相同的目录下.

Draw文件夹中是对应报告中的训练图片生成的代码.

更详细的解释前馈神经网络参考报告多层神经网络的训练问题.pdf

下面是核心算法实现原理图：