numpy实现全连接网络进行mnist训练测试

import numpy as np
import matplotlib.pyplot as plt
from sklearn.utils import  shuffle
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import minmax_scale
import pickle
import struct
import os"""定义几个要用到的函数"""
def load_mnist(path, kind='train'):"""Load MNIST data from `path`"""labels_path = os.path.join(path, '%s-labels-idx1-ubyte' % kind)images_path = os.path.join(path, '%s-images-idx3-ubyte' % kind)with open(labels_path, 'rb') as lbpath:magic, n = struct.unpack('>II', lbpath.read(8))labels = np.fromfile(lbpath, dtype=np.uint8)with open(images_path, 'rb') as imgpath:magic, num, rows, cols = struct.unpack('>IIII', imgpath.read(16))images = np.fromfile(imgpath, dtype=np.uint8).reshape(len(labels), 784)return images, labelsdef softmax(x, axis=1):# 计算每行的最大值row_max = x.max(axis=axis)# 每行元素都需要减去对应的最大值，否则求exp(x)会溢出，导致inf情况row_max=row_max.reshape(-1, 1)x = x - row_max# 计算e的指数次幂x_exp = np.exp(x)x_sum = np.sum(x_exp, axis=axis, keepdims=True)s = x_exp / x_sumreturn sdef sigmoid(x, deriv=False):if deriv==True:return sigmoid(x)*(1-sigmoid(x))return 1/(1+np.exp(-x))def transform_one_hot(labels):# 转为one-hot编码n_labels = np.max(labels) + 1one_hot = np.eye(n_labels)[labels]return one_hot"""加载数据"""
images, labels = load_mnist(r'G:\demo\daily\raw')
Y = transform_one_hot(labels.astype(np.int)).astype(np.float16)
X = minmax_scale(images, feature_range=(0, 1), axis=1)
X, Y = shuffle(X, Y)
X_train, X_test, y_train, y_test = train_test_split(X, Y, test_size = 1/6)"""设置网络参数"""
hidden_layer = 100
learning_rate = 0.01
M = 1.05
batch_size = 200
epoch = 30"""初始化网络参数"""
# np.random.seed(2)
sample_size = X_train.shape[0]
input_layer_neurons = X_train.shape[1]
output_layer_neurons = y_train.shape[1]
w1 = 2*np.random.random((input_layer_neurons, hidden_layer))-1
w2 = 2*np.random.random((hidden_layer, output_layer_neurons))-1
b1 = np.zeros((1, hidden_layer))
b2 = np.zeros((1, output_layer_neurons))"""训练网络"""
losses = []
learning_accs = []
validation_accs = []
for i in range(epoch):alpha = M**i*learning_ratefor j in range(sample_size//batch_size):x_batch = X_train[j*batch_size:(j+1)*batch_size]y_batch = y_train[j*batch_size:(j+1)*batch_size]# 前向传播a1 = np.dot(x_batch, w1) + b1h1 = sigmoid(a1)a2 = np.dot(h1, w2) + b2# 计算损失loss = -y_batch.squeeze()*np.log(softmax(a2))# 反向传播dlda2 = softmax(a2) - y_batchdldw2 = np.dot(h1.T, dlda2)dldb2 = np.mean(dlda2, axis=0).reshape(1, -1)dldh1 = np.dot(dlda2, w2.T)dlda1 = dldh1 * sigmoid(a1, deriv=True)dldw1 = np.dot(x_batch.T, dlda1)dldb1 = np.mean(dlda1, axis=0).reshape(1, -1)# 更新权重w1 -= alpha * dldw1w2 -= alpha * dldw2b1 -= alpha * dldb1b2 -= alpha * dldb2# 打印训练进度if(sample_size//batch_size<=25):print('\r'+"Epoch: "+str(i)+"\t"+str(round((j+1)/(sample_size//batch_size)*100,1))+"%" +"\t"+(j+1)*'>'+(sample_size//batch_size-1-j)*'.', end="", flush=True)else:if (j+1)%(sample_size//batch_size//25)==0:print("\r"+"Epoch: "+str(i)+"\t"+str(round((j+1)/(sample_size//batch_size)*100,1))+"%" +"\t"+(j+1)//(sample_size//batch_size//25)*'>'+(sample_size//batch_size-1-j)//(sample_size//batch_size//25)*'.', end="", flush=True)# 模型评估ytrain_ = np.argmax(softmax(np.dot(sigmoid(np.dot(X_train[0:10000], w1) + b1), w2) + b2),axis=1)ytest_ = np.argmax(softmax(np.dot(sigmoid(np.dot(X_test[0:10000], w1) + b1), w2) + b2),axis=1)losses.append(loss.mean())learning_acc = np.sum(np.argmax(y_train[0:10000], axis=1)==ytrain_)/len(ytrain_)validation_acc = np.sum(np.argmax(y_test[0:10000], axis=1)==ytest_)/len(ytest_)learning_accs.append(learning_acc)validation_accs.append(validation_acc)# 打印迭代结果print("\tLoss:"+str(round(loss.mean(), 5))+"\tLearning_Acc:"+str(round(learning_acc*100, 2))+"%\tValidation_Acc:"+str(round(validation_acc*100, 2))+"%")"""画图"""
fig, ax1 = plt.subplots()
ax2 = ax1.twinx()
ax2.plot(losses,  '*-', linewidth=2, color='g')
ax1.plot(learning_accs, '^-', linewidth=2, label="learning_Acc")
ax1.plot(validation_accs, 'o-', linewidth=2, label='validation_Acc')
ax1.set_ylabel('Accuracy(%)')
ax2.set_ylabel('loss')
ax1.set_xlabel('epoch')
plt.grid()
ax1.legend(loc='best')
plt.show()

numpy实现全连接网络进行mnist训练测试相关推荐

MNIST 训练测试
如下是用的是普通的前馈神经网络,输入层节点为784,隐层节点500,输出层10.这是一个标准的讲解DeepLearning中,在Tensorflow, Pytorch框架下的典型的用于识别MNIST手 ...
从一到二：利用mnist训练集生成的caffemodel对mnist测试集与自己手写的数字进行测试...
通过从零到一的教程,我们已经得到了通过mnist训练集生成的caffemodel,主要包含下面四个文件: 接下来就可以利用模型进行测试了.关于测试方法按照上篇教程还是选择bat文件,当然python. ...
(caffe入门)windows caffe 之 mnist 训练
文章目录 1. mnist 数据集下载 2. mnist 数据集转换 3. 修改网络结构文件 lenet_train_test.prototxt 和网络求解文件 lenet_solver.proto ...
深度学习实战（七）——目标检测API训练自己的数据集（R-FCN数据集制作+训练+测试）
TensorFlow提供的网络结构的预训练权重:https://cloud.tencent.com/developer/article/1006123 将voc数据集转换成.tfrecord格式供te ...
【MMDetection3D】环境搭建，使用PointPillers训练测试可视化KITTI数据集
文章目录前言 3D目标检测概述 KITTI数据集简介 MMDetection3D 环境搭建数据集准备训练测试及可视化绘制损失函数曲线参考资料前言 2D卷不动了,来卷3D,之后更多地工作会 ...
MATLAB实现基于BP神经网络的手写数字识别+GUI界面+mnist数据集测试
文章目录 MATLAB实现基于BP神经网络的手写数字识别+GUI界面+mnist数据集测试一.题目要求二.完整的目录结构说明三.Mnist数据集及数据格式转换四.BP神经网络相关知识 4.1 ...
mask-rcnn训练测试自制数据集
mask-rcnn训练测试自制数据集本项目简介本项目用于口腔模型分割,数据类型有7种,本文主要用于介绍如何使用自制数据集训练自己的模型训练环境配置操作系统:win10 GPU: GTX 108 ...
vid2vid 代码调试+训练+测试（debug+train+test）（一）测试篇
## Prerequisites - Linux or macOS - Python 3 - NVIDIA GPU + CUDA cuDNN - PyTorch 0.4 但一般的话我们为了保护已有的环 ...
什么是端到端训练测试_为什么端到端测试对您的团队很重要
什么是端到端训练测试 by Phong Huynh 由Phong Huynh 为什么端到端测试对您的团队很重要 (Why End-to-End Testing is Important for You ...

numpy实现全连接网络进行mnist训练测试

numpy实现全连接网络进行mnist训练测试相关推荐

最新文章

热门文章