课程
课程汇总
作业
神经网络基础
'''
@Description: 三层神经网络实现分类
@Author: Su Yunzheng
@Date: 2020-06-17 15:40:53
@LastEditTime: 2020-06-18 16:12:21
@LastEditors: Su Yunzheng
'''
import numpy as np
import sklearn
import matplotlib
import matplotlib.pyplot as plt
from getData import gen_classify_data
# 可视化
# input:sample_x(2*n), sample_y(n,)
def visualization(sample_x, sam_type, save=False, path='inference/sample-data.png'):
samp = plt.figure()
samp_ax = samp.add_subplot(1, 1, 1)
samp_ax.set_title('sample data')
type = sam_type # 给定一个label(n,)
samp_ax.scatter(sample_x[0], sample_x[1], c=type, linewidths=type)
if save:
plt.savefig(path, dpi=400, bbox_inches='tight')
plt.show()
# 可视化对比图
# 可视化
# input:sample_x(2*1000), sam_type(1000,), pred_type(1000,)
def compare_visual(sample_x, sam_type, pred_type, save=False, path='inference/inference-compare.png',
title='inference comparison', loss=0):
# 查看输出的格式
# sample_x:(2, 1000),sample_y:(3, 1000),predict_y:(3, 1000)
# print('sample_x:{},sample_y:{},predict_y:{}'.format(sample_x.shape, sample_y.shape, predict_y.shape))
comp = plt.figure()
comp.suptitle(title)
samp_ax = plt.subplot2grid((2, 2), (0, 0)) # 划分位置
pred_ax = plt.subplot2grid((2, 2), (0, 1))
comp_ax = plt.subplot2grid((2, 2), (1, 0), colspan=2)
samp_ax.set_title('sample results')
pred_ax.set_title('predict results')
comp_ax.set_title('comparison loss= %.5f' %loss)
# sam_type = np.where(sample_y.T == sample_y.max())[1] # 1000个点的类别 （1000，）
# pred_type = np.where(predict_y.T == predict_y.max())[1] # 1000个点的类别 0 1 2
# print("predict_y:{}".format(predict_y.shape)) # 3*1000
# print("pred_type :{},sam_type shape:{}".format(pred_type,sam_type.shape))
# print('sam_type:{}'.format(sam_type))
d_type = pred_type - sam_type # label和predict的差异
# print('type:{}'.format(type))
samp_ax.scatter(sample_x[0], sample_x[1], c=sam_type, linewidths=sam_type)
pred_ax.scatter(sample_x[0], sample_x[1], c=pred_type, linewidths=pred_type)
comp_ax.scatter(sample_x[0], sample_x[1], c=d_type) # 根据预测结果的不同，差异越大，其余颜色的数目就越多
if save:
plt.savefig(path, dpi=400, bbox_inches='tight')
plt.show()
# 生成作业数据数据 X为坐标，y为标签
X,y=gen_classify_data(200)
y=np.argmax(y,axis=0) # [1,0,0]-->0; [0,1,0]-->1; [0,0,1]-->2;
X=np.transpose(X)
# print("y :\n{}".format(y))
# print("y shape :\n{}".format(y.shape))
print("X :{},y :{}".format(X.shape,y.shape))
print("----------------------------------------")
num_examples = len(X) # 样本数 本例题为200
nn_input_dim = 2 # 输入的维度 本例题有两个特征向量
# nn_output_dim = 2 # 输出的类别个数 分为两类
nn_output_dim = 3 # 输出的类别个数 分为三类
# 梯度下降参数
epsilon = 0.01 # 学习率
reg_lambda = 0.01 # 正则化参数
# 定义损失函数，计算损失(才能用梯度下降啊...)
def calculate_loss(model,X=X,y=y,num_examples=num_examples):
'''
input: model, X, y, num_examples样本的数量(200或者1000)
'''
W1, b1, W2, b2 = model['W1'], model['b1'], model['W2'], model['b2']
# print("W1:{}\nb1:{}\nW2:{}\nb2:{}\n".format(W1,b1,W2,b2))
# print("X:{}\n".format(X.shape))
# 向前推进，前向运算
z1 = X.dot(W1) + b1 #200*2*2*10 =200*10, 200*10+200*10=200*10 ，z1:200*3
# print("X.dot(W1):{}".format(X.dot(W1).shape))
# print("z1:{}\n".format(z1.shape))
a1 = np.tanh(z1) #a1:200*10
# print("a1:{}".format(a1.shape)) # 激活
z2 = a1.dot(W2) + b2 #200*10*10*3=200*3 , b2:1*3->200*3 , z2:200*3
exp_scores = np.exp(z2) #200*3 # 全部取指数 softmax函数
# print("exp_scores.shape: {} ".format(exp_scores.shape))
probs = exp_scores / np.sum(exp_scores, axis=1, keepdims=True) #np.sum() 将矩阵按列相加，全部加到第一列 , 200*3/200*1->200*3/200*3
# print("np.sum:{}".format(np.sum(exp_scores, axis=1, keepdims=True).shape))
# print("probs:{}".format(probs)) # 概率值 200*3
# 计算损失 交叉熵函数
corect_logprobs = -np.log(probs[range(num_examples), y]) #取prob这个200*3的概率矩阵的每一行，具体是第几列是靠对应的y来确定的 #200*1
# print("y:{}".format(y))
data_loss = np.sum(corect_logprobs) #200行加在一起->1*1 一个数
# print("corect_logprobs:{}".format(corect_logprobs.shape)) #200*1
# print("data_loss:{}".format(data_loss)) #200行加在一起->1*1 一个数
# 也得加一下正则化项
data_loss += reg_lambda/2 * (np.sum(np.square(W1)) + np.sum(np.square(W2))) #W1：2*3 W2:3*2 data_loss是一个数
# print("data_loss:{}".format(data_loss))
# print("1./num_examples * data_loss:{}".format(1.*data_loss/num_examples ))
return 1./num_examples * data_loss #返回一个数，作为损失值
# 完整的训练建模函数定义
def build_model(nn_hdim, num_passes=20000, print_loss=False):
'''
参数：
1) nn_hdim: 隐层节点个数 作业为10
2）num_passes: 梯度下降迭代次数
3）print_loss: 设定为True的话，每1000次迭代输出一次loss的当前值
'''
# 随机初始化一下权重呗
np.random.seed(0) #seed只对第一组随机数起作用
W1 = np.random.randn(nn_input_dim, nn_hdim) / np.sqrt(nn_input_dim) # nn.sqrt打印nn_input_dim=2的开方，也就是1.414
# print("nn.sqrt:{}".format(np.sqrt(nn_input_dim)))
# print("W1:{}".format(W1))
# print(" np.random.randn(nn_input_dim, nn_hdim):{}",format( np.random.randn(nn_input_dim, nn_hdim)))
b1 = np.zeros((1, nn_hdim))
W2 = np.random.randn(nn_hdim, nn_output_dim) / np.sqrt(nn_hdim)
b2 = np.zeros((1, nn_output_dim))
# 这是咱们最后学到的模型
model = {}
loss=[] # 损失
# 开始梯度下降...
for i in range(0, num_passes):
# 前向运算计算loss
z1 = X.dot(W1) + b1
a1 = np.tanh(z1) # 激活函数 tanh
z2 = a1.dot(W2) + b2
exp_scores = np.exp(z2)
probs = exp_scores / np.sum(exp_scores, axis=1, keepdims=True) # softmax 得到概率
# print("probs shape: {}".format(probs.shape)) # 概率值 200*3
# 反向传播
delta3 = probs #### 200*2 每一列表示对应的概率
delta3[range(num_examples), y] -= 1 # 200
# print("delta3:{}".format(delta3.shape)) ################################
# print("delta3:{}".format(delta3)) ################################
# 增量
dW2 = (a1.T).dot(delta3)
db2 = np.sum(delta3, axis=0, keepdims=True)
delta2 = delta3.dot(W2.T) * (1 - np.power(a1, 2))
dW1 = np.dot(X.T, delta2)
db1 = np.sum(delta2, axis=0)
# 加上正则化项
dW2 += reg_lambda * W2
dW1 += reg_lambda * W1
# 梯度下降 更新参数
W1 += -epsilon * dW1
b1 += -epsilon * db1
W2 += -epsilon * dW2
b2 += -epsilon * db2
# 得到的模型实际上就是这些权重
model = { 'W1': W1, 'b1': b1, 'W2': W2, 'b2': b2}
loss.append(calculate_loss(model)) # 将所有的loss放在一个list中
# 如果设定print_loss了，那我们汇报一下中间状况
if print_loss and i % 1000 == 0:
print("Loss after iteration {}: {} ".format(i,calculate_loss(model)))
# print("loss:{}".format(loss)) # 输出总的loss 一个list个格式
# 绘制train-loss曲线
fig = plt.figure()
ax = fig.add_subplot(1, 1, 1)
ax.set_title('train-loss')
ax.set_ylim([0, 1])
ax.plot(loss)
plt.savefig('./inference/train-loss')
plt.show()
return model
# 判定结果的函数
def predict(model, x):
W1, b1, W2, b2 = model['W1'], model['b1'], model['W2'], model['b2']
# 前向运算
z1 = x.dot(W1) + b1
a1 = np.tanh(z1)
z2 = a1.dot(W2) + b2
exp_scores = np.exp(z2) #200*3
# 计算概率输出最大概率对应的类别
probs = exp_scores / np.sum(exp_scores, axis=1, keepdims=True) #200*3
# print('probs:{}'.format(probs)) # 输出概率矩阵
return np.argmax(probs, axis=1) #返回200行中，每一行的最大的值，得到的矩阵是一个200*1的矩阵，表示200个元素对应的类别
# 计算精度
def getAcc(pred,y):
ac=sum(pred==y)/len(pred) ## 统计两个数组相同元素的数目
# ac=sum(pred==y) ## 统计两个数组相同元素的数目
return ac*100.0
# 建立隐层有10个节点(神经元)的神经网络
model = build_model(10, print_loss=True)
print("-------------------------")
# print("model:{}".format(model))
# 评价训练集
pred=predict(model,X) # 得到预测值
# 计算损失
train_loss=calculate_loss(model,X,y)
print("train_loss:{}".format(train_loss))
# print("pred:{}".format(pred))
# print("y :{}".format(y))
# 计算精度
# print("y shape:{},pred shape:{}".format(y.shape,pred.shape)) # y shape:(200,),pred shape:(200,)
print("Train Acc:{}%".format(getAcc(pred,y)))
# 可视化训练集
visualization(np.transpose(X),y,True,path='./inference/train.png')
compare_visual(np.transpose(X),y,pred,True,path='./inference/compare_train.png',loss=train_loss)
print("------------------Test------------------")
##### 生成测试集数据 1000个点
X_test,y_test=gen_classify_data(1000)
y_test=np.argmax(y_test,axis=0) # [1,0,0]-->0; [0,1,0]-->1; [0,0,1]-->2;
X_test=np.transpose(X_test) # 2*1000 ---> 1000*2
# print("y :\n{}".format(y))
# print("y shape :\n{}".format(y.shape))
print("X_test :{},y_test :{}".format(X_test.shape,y_test.shape)) # X_test :(1000, 2),y_test :(1000,)
# 得到预测值
pred_test = predict(model,X_test) # 得到预测值
# 计算损失
test_loss = calculate_loss(model,X_test,y_test,len(X_test))
print("test_loss:{}".format(test_loss))
# 计算精度
print("Test Acc:{}%".format(getAcc(pred_test,y_test)))
# 可视化训练集
visualization(np.transpose(X_test),y_test,save=True,path='inference/test.png')
compare_visual(np.transpose(X_test),y_test,pred_test,save=True,path='inference/compare_test.png',loss=test_loss)
可视化
MNIST手写数字体识别
给出三种实现方式
1、MLP
'''
@Description: mlp MNIST手写数字体识别
@Author: Su Yunzheng
@Date: 2020-06-18 16:14:32
@LastEditTime: 2020-06-18 17:18:14
@LastEditors: Su Yunzheng
'''
### 查看tensorboard:
# (tf) C:\Users\11604>tensorboard --logdir F:\睿慕课\7\苏云征-三维点云7\logs --host=127.0.0.1
import tensorflow as tf
import matplotlib.pyplot as plt
import os
import numpy as np
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("MNIST_DATA",one_hot=True) #onehot对标签的标注，非onehot是1,2,3.onehot就是只有一个1其余全是0
#超参数（学习率，batch的大小，训练的轮数，多少轮展示一下loss）
learning_rate = 0.2 # 学习率
num_step = 20 # 迭代次数
batch_size = 128 # 一个batch的size
n_batch=mnist.train.num_examples//batch_size # 有多少个batch
display_step =100 # 多少轮打印一次
#网络参数（有多少层网络，每层有多少个神经元，整个网络的输入是多少维度的，输出是多少维度的）
num_input = 784 #(28*28)
h1=1000 # 隐藏层 这个是没用到的
num_class = 10 # 输出
#图的输入
X = tf.placeholder(tf.float32,[None,num_input])
Y = tf.placeholder(tf.float32,[None,num_class])
#网络的权重和偏向,如果是两个隐层的话需要定义三个权重，包括输出层
weights={
'out':tf.Variable(tf.zeros([num_input,num_class]))
}
biase = {
'out':tf.Variable(tf.zeros([num_class]))
}
#定义网络结构
def neural_net(x):
# h1_out = tf.matmul(x,weights['h1'])+biase['h1']
out_layer = tf.matmul(x,weights['out'])+biase['out']
return out_layer
#模型输出处理
logits = neural_net(X)
prediction = tf.nn.softmax(logits) # 得到batch_size*10的概率矩阵
# 定义损失和优化器
loss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits,labels=Y))
optimizer = tf.train.GradientDescentOptimizer(learning_rate = learning_rate)
train_step = optimizer.minimize(loss_op)
#评估模型准确率
correct_pred = tf.equal(tf.argmax(prediction,1),tf.argmax(Y,1)) #结果存放在一个布尔型列表中(argmax函数返回一维张量中最大的值所在的位置)
accuracy = tf.reduce_mean(tf.cast(correct_pred,tf.float32)) #求准确率(tf.cast将布尔值转换为float型)
#初始化变量
init = tf.global_variables_initializer()
mse_summary=tf.summary.scalar("loss", loss_op)
# Create a summary to monitor accuracy tensor
acc_summary=tf.summary.scalar("accuracy", accuracy)
file_writer = tf.summary.FileWriter('./logs',tf.get_default_graph())
#开始训练
with tf.Session() as sess:
sess.run(init)
for step in range(num_step): # 总迭代
for batch in range(n_batch):
batch_x,batch_y = mnist.train.next_batch(batch_size)
_,los,ac=sess.run([train_step,mse_summary,acc_summary],feed_dict={X:batch_x,Y:batch_y})
# 记录下来每一次训练的数据
file_writer.add_summary(los,step*n_batch+batch) # loss
file_writer.add_summary(ac,step*n_batch+batch) # acc
# if step % display_step == 0 or step == 1:
# print("step:{},loss:{},acc:{}".format(step,loss,acc))
acc=sess.run(accuracy,feed_dict={X:mnist.test.images,Y:mnist.test.labels})
print("Iter:{},testing Accuracy:{}".format(step,acc))
print("优化完成!")
#训练完模型后，开始测试
# 读取图片
# print(mnist.test.images.shape) #(10000, 784)
plt.imshow(mnist.test.images[0].reshape(28,28))
plt.show()
pred_pic=mnist.test.images[0].reshape(1,-1)
pred_num=sess.run(prediction,feed_dict={X:pred_pic})
print("预测结果：{}".format( np.argmax(pred_num) ))
2、K-近邻
3、CNN
'''
@Description: cnn MNIST手写数字体识别
@Author: Su Yunzheng
@Date: 2020-06-18 16:14:39
@LastEditTime: 2020-06-18 16:14:39
@LastEditors: Su Yunzheng
'''
### 查看tensorboard:
# (tf) C:\Users\11604>tensorboard --logdir F:\睿慕课\7\苏云征-三维点云7\logs --host=127.0.0.1
import numpy as np
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import matplotlib.pyplot as plt
mnist = input_data.read_data_sets("MNIST_DATA", one_hot=True)
logs_path = './log/example/' # log存放位置
batch_size = 50
n_batch=mnist.train.num_examples//batch_size # 有多少个batch
def weight_variable(shape,namew='w'):
initial = tf.truncated_normal(shape, stddev=0.1)
return tf.Variable(initial,name=namew)
def bias_variable(shape,nameb='b'):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial,name=nameb)
def conv2d(x, W,namec='c'):
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME',name=namec)
def max_pool_2x2(x,namep='p'):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME',name=namep)
x = tf.placeholder(tf.float32, [None, 784],name='xinput')
y_ = tf.placeholder(tf.float32, [None, 10])
x_image = tf.reshape(x, [-1, 28, 28, 1])
# Conv1 Layer
W_conv1 = weight_variable([5, 5, 1, 32],'w1')
b_conv1 = bias_variable([32],'b1')
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1,'c1')
h_pool1 = max_pool_2x2(h_conv1,'p1')
# Conv2 Layer
W_conv2 = weight_variable([5, 5, 32, 64],'w2')
b_conv2 = bias_variable([64],'b2')
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)
W_fc1 = weight_variable([7 * 7 * 64, 1024],'wf1')
b_fc1 = bias_variable([1024],'bf1')
h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)
keep_prob = tf.placeholder(tf.float32,name="prob")
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
W_fc2 = weight_variable([1024, 10],'wfull2')
b_fc2 = bias_variable([10],'bf2')
y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y_conv), reduction_indices=[1]))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
init=tf.global_variables_initializer()
# tensorboard相关的
# Create a summary to monitor cost tensor
mse_summary=tf.summary.scalar("loss", cross_entropy)
# Create a summary to monitor accuracy tensor
acc_summary=tf.summary.scalar("accuracy", accuracy)
file_writer = tf.summary.FileWriter('./logs',tf.get_default_graph())
# Merge all summaries into a single op
# merged_summary_op = tf.summary.merge_all()
#开始训练
with tf.Session() as sess:
sess.run(init)
# op to write logs to Tensorboard
# summary_writer = tf.summary.FileWriter("logs/", graph=tf.get_default_graph())
saver = tf.train.Saver()
for i in range(2):
for batch_i in range(n_batch):
batch = mnist.train.next_batch(batch_size)
sess.run(train_step,feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})
summary_str = mse_summary.eval(feed_dict={x:batch[0],y_:batch[1],keep_prob:1.0})
accuracy_str=acc_summary.eval(feed_dict={x:batch[0],y_:batch[1],keep_prob:1.0})
# 记录下来每一次训练的数据
file_writer.add_summary(summary_str,i*n_batch+batch_i)
file_writer.add_summary(accuracy_str,i*n_batch+batch_i)
# 训练集精度
if (batch_i%500==0):
train_accuracy = sess.run(accuracy,feed_dict={x: batch[0], y_: batch[1], keep_prob: 1.0})
print("Iter %d, batch_i %d,training accuracy %g" % (i,batch_i, train_accuracy))
# 一次遍历完整个数据集，打印测试集精度
test_acc=sess.run(accuracy,feed_dict={x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0})
print("Iter %d, test accuracy %g" % (i, test_acc))
saver.save(sess, "Model/model.ckpt")
print("model saved!")
# 测试
#训练完模型后，开始测试
# 读取图片
# print(mnist.test.images.shape) #(10000, 784)
plt.imshow(mnist.test.images[0].reshape(28,28))
plt.show()
pred_pic=mnist.test.images[0].reshape(1,-1)
# print("pred_pic:{}".format(pred_pic.shape)) # pred_pic:(1, 784)
pred_num=sess.run(y_conv,feed_dict={x:pred_pic,keep_prob: 1.0}) # 这是一个1*10的概率矩阵 , 这里feed_dict中要有keep_prob,否则会报错
# print("概率矩阵softmax:{}".format(pred_num.shape))
print("预测结果：{}".format( np.argmax(pred_num) ))
583
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
