清华AI自强计划作业3-Part1
关于Incompatible shapes: [30,784] vs. [784,1]错误
今天在完成MNIST数据集进行简单的分类训练的作业时，遇到了这个问题，搞了一会，知道错误在哪，特地写一下心得注明一下。
第三次作业Part1
以下是修改错误之前的源码
from tensorflow.examples.tutorials.mnist import input_data # 导入 tensorflow 中的 input_data 子模块，目的是为了后续的导入读取数据
import tensorflow as tf # 导入 tensorflow 库，并且重名为 tf, 便于后面的简写 tf
import numpy as np # 导入 numpy 库，并且重名为 np, 便于后面的简写 np
#基本参数设置
batchSize = 30 #batchsize的大小，代表每次训练载入的图像张数
lr = 0.005 #学习率的大小，若后面启用learning rate decay策略，则该值为学习率的初始值
iter = 1000000 #训练的迭代次数
saveInter = 100 #保存结果的频率，即每训练100次保存一次模型训练参数及模型性能
sample_size = 55000 #学习example的总大小，MNIST中官方写60000张，实际为55000（训练）+ 5000（校验），本例中只使用了55000 train
# 对模型输出的结果进行评判，>0.5为“正”，<0.5为“负”
def predict(X): # 定义一个函数 predict， 作用是用来进行预测
num = X.shape[0] # 通过 shape 属性，得到 X 行的个数
result = [] # 定义一个空的列表 result ，后面通过 append 的方式，向里面添加元素
for i in range(num): # for循环语句， i 从0，1，2, 到 num -1
if X[i]>0.5: # 如果 X[i] 大于 0.5
result.append(1.0) # 将 1.0 添加到列表 result 中
else: # 否则，X[i] 小于或等于 0.5
result.append(0.0) # 将 0.0 添加到列表 result 中
return result # 返回 result 的结果
# 加载数据集，建议提前到官网上下载MNIST数据集，并解压到./MNIST文件夹下
# MNIST下载地址：http://yann.lecun.com/exdb/mnist/
def loadData(): # 定义一个 loadData 函数
file = "../MNIST" # 数据集 MINIST
mnist = input_data.read_data_sets(file, one_hot=True) # input_data.read_data_sets 读取数据
return mnist # 返回读取的数据 mnist
# 申请模型输入输出的占位符
def create_placeholder(n_x=784,n_y=0): # 定义一个 create_placeholder 函数
X = tf.placeholder(tf.float32,shape=[None,n_x],name='X') # 调用tf.placeholder函数，tensorflow 中定义 X
Y = tf.placeholder(tf.float32, shape=[None,], name='Y') # 调用tf.placeholder函数，tensorflow 中定义 Y
return X,Y #返回 X 和 Y 的数值
# 定义参数，W,b
def initialize_parameters(): # 定义一个 initialize_parameters 函数
W = tf.Variable(tf.zeros([784,1])) #调用tf.Variable函数，设置模型参数W，W的维度为[784,1]，且初始化为0
b = tf.Variable(tf.zeros([1,1])) #调用tf.Variable函数，设置模型参数b，b的维度为[1 ,1],且初始化为0
parameters={'W': W, # 参数权重 W
'b': b} # 参数偏置 b
return parameters # 返回参数
# 将标签转换为one-hot形式，本例中未用到该函数，是因为tensorflow中封装了one-hot功能
def convert_one_hot(Y,C): # 定义一个 convert_one_hot 函数
one_hot=np.eye(C)[Y.reshape(-1)].T # 初始化 one_hot 为对角矩阵
return one_hot # 返回 one_hot
# 定义网络模型
def forward_propagation(X,parameters): # 定义一个 forward_propagation 函数
W = parameters['W'] # 参数权重 W
b = parameters['b'] # 参数偏置 b
Z1=(X*W)+b #调用tensorflow函数，实现Z1=X*W+b
A1=tf.nn.sigmoid(Z1) #调用tf.nn.sigmoid，实现A1 = sigmoid(Z1)
A1 = tf.clip_by_value(A1,1e-3,1.0) #调用clip_by_value，将A1进行裁剪，使其在[0.001，1.0]之间，是为了避免出现接近于0的极小值，输入np.log()中出现nan的情况
return A1 # 返回 A1
# 定义loss function
def compute_cost(y_,y,W): # 定义一个 compute_cost 函数
#以下的cross_entropy经过了简单变化，在(1.0-y_)*tf.log(1.0-y)之前乘以0.1，是因为正负样本比例基本上为1：9，严重偏向负样本
#以下添加了正则，也可以尝试去掉
cross_entropy = -(1.0/batchSize)*tf.reduce_sum(y_*tf.log(y)+0.1*(1.0-y_)*tf.log(1.0-y)+tf.contrib.layers.l2_regularizer(0.1)(W)) #调用tf.reduce_sum函数，实现交叉熵函数
return cross_entropy # 返回 交叉熵函数 的数值 cross_entropy
# 模型搭建、训练、存储
def model(mnist,Num): # 定义一个 model 函数
x,y_ = create_placeholder(784, 0) # 调用 create_placeholder 函数，初始化 x,y_
parameters = initialize_parameters() # 调用 initialize_parameters 函数， 初始化 参数
A1 = forward_propagation(x, parameters) # 调用 forward_propagation 函数，实现前向反馈
#设置learning rate decay策略，随着迭代次数的增加，学习率成指数逐渐减小，减小公式为：decayed_learning_rate = learning_rate * decay_rate ^ (global_step / decay_steps)
global_step = tf.Variable(0) # 调用 tf.Variable 函数， 初始化 global_step 变量
learning_rate = tf.train.exponential_decay(lr,global_step,decay_steps=sample_size/batchSize,decay_rate=0.98,staircase=True) # 设置指数衰减的 学习率，调用tf.train.exponential_decay。
cost = compute_cost(y_, A1,parameters['W']) # 调用 compute_cost 函数，计算损失函数
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost,global_step=global_step) # 调用 tf.train.GradientDescentOptimizer 函数， 实现梯度下降的优化
sess = tf.InteractiveSession() #调用tf.InteractiveSession()函数，创建Session
sess.run(tf.global_variables_initializer()) #执行tf.global_variables_initializer()，初始化参数
#利用全部样本对模型进行测试
testbatchX = mnist.train.images # 导入 mnist 数据中的训练集 图片
testbatchY = mnist.train.labels # 导入 mnist 数据中的训练集 标签
modelLast = [] # 定义一个空的列表 modelLast
logName = "./log"+str(Num)+".txt" # 新建文件名为 log"+str(Num)+".txt
#保存模型，且设定保存最大迭代次数的4个
saver = tf.train.Saver(max_to_keep=4) # 调用 tf.train.Saver 函数，保存模型
pf = open(logName, "w") # 以 写入的方式 打开文件 log"+str(Num)+".txt
for i in range(iter): # for 循环结构， 遍历 iter
#加载minibatch=50个训练样本
batch = mnist.train.next_batch(batchSize) # 调用 mnist.train.next_batch 函数，复制给 batch
batchX = batch[0] # 赋值给 batchX 为 batch 中第一个元素
batchY = batch[1] # 赋值给 batchY为 batch 中第二个元素
#执行训练
train_step.run(feed_dict={x:batchX, y_:batchY[:,Num]}) #执行tensor流图，并为其添加输入x: batchX, y_: batchY[:,Num]
#每隔saveInter次迭代，保存当前模型的状态，并测试模型精度
if i % saveInter == 0: #条件判断语句 if， 如果 i 整除 iter
[total_cross_entropy,pred,Wsum,lrr] = sess.run([cost,A1,parameters['W'],learning_rate],feed_dict={x:batchX,y_:batchY[:,Num]}) # 调用 sess.run， 启动 tensoflow
pred1 = predict(pred) # 调用 predict 函数，进行预测
#保存当前模型的学习率lr、在minibatch上的测试精度
print('lr:{:f},train Set Accuracy: {:f}'.format(lrr,(np.mean(pred1 == batchY[:,Num]) * 100))) # 输出训练集的准确率等
pf.write('lr:{:f},train Set Accuracy: {:f}\n'.format(lrr,(np.mean(pred1 == batchY[:,Num]) * 100))) # 写入训练集的准确率
#保存迭代次数、cross entropy
print("handwrite: %d, iterate times: %d , cross entropy:%g"%(Num,i,total_cross_entropy)) # 输出迭代次数，交叉熵损失函数等
pf.write("handwrite: %d, iterate times: %d , cross entropy:%g, W sum is: %g\n" %(Num,i,total_cross_entropy,np.sum(Wsum))) # 写入出迭代次数，交叉熵损失函数等
#保存当前参数状态、测试testbatch上的精度
[testpred] = sess.run([A1],feed_dict={x: testbatchX, y_: testbatchY[:, Num]}) # 调用 sess.run， 启动 tensoflow
testpred1 = predict(testpred) # 调用 predict 函数，进行预测
print('predict sum is: {:f},Testing Set Accuracy: {:f}\n'.format(np.sum(testpred1),(np.mean(testpred1 == testbatchY[:, Num]) * 100))) # 输出测试集的准确率等
pf.write('predict sum is: {:f},Testing Set Accuracy: {:f}\n'.format(np.sum(testpred1),(np.mean(testpred1 == testbatchY[:,Num]) * 100))) # 写入测试集的准确率等
pf.write("\n") # 写入换行字符
#保存当前模型
saveName = "model/my-model-" + str(Num) # 保存模型为 "model/my-model-" + str(Num)
saver.save(sess, saveName, global_step=i) # 调用 saver.save 函数，保存模型
pf.write("save model completed\n") # 写入 save model completed
#若交叉熵出现nan（出现极值），此时停止训练，保存最新的一次模型名称
if total_cross_entropy != total_cross_entropy: # 条件判断语句 if ， 如果 total_cross_entropy 不等于 total_cross_entropy
print("is nan, stop") # 输出 is nan, stop
pf.write("is nan, stop\n") # 写入 is nan, stop
modelLast = "model/my-model-" + str(Num)+str(i-saveInter) # 模型文件名为 "model/my-model-" + str(Num)+str(i-saveInter)
break; # break 跳出循环
pf.close() # close 关闭打开的文件
return modelLast # 返回 modelLast
# 模型测试
def test_model(): # 定义 test_model 函数
mnist = loadData() # 调用 loadData 函数， 导入数据
classNum = 10 # 类别 初始化赋值为 10 ， 共有 10 类
modelNames = [] # 定义一个空的列表 modelNames
logName = "./logModelNames.txt" # 文件名为 logModelNames.txt
pf = open(logName, "w") # 以写入的方式打开 logModelNames.txt
#循环训练每个类别与其他类别的二分类器，保存10个分类器模型
for i in range(classNum): # for 循环语句， 遍历所有 classNum的类别，
modelNames.append(model(mnist,i)) # 通过 append 的方式， 向 modelNames 里面添加 model(mnist,i)
pf.write(modelNames[i]) # 写入 modelNames[i]
pf.write("\n") # 写入 换行字符
pf.close() # 关闭文件
if __name__ == '__main__': # 主程序
test_model() # 调用 test_model 函数
> 以下是错误信息 InvalidArgumentError (see above for traceback): Incompatible shapes: [30,784] vs. [784,1]
错误原因是在创建不含隐藏层的简单神经网络时tf.matmul()的参数传递问题，正确的应为 tf.matmul(x,W)
将"Z1=(X*W)+b “更改为"Z1=tf.matmul(X,W)+b”，以上代码即可运行
运行结果如下
lr:0.005000,train Set Accuracy: 16.666667
handwrite: 0, iterate times: 0 , cross entropy:4.98731
predict sum is: 55000.000000,Testing Set Accuracy: 9.898182
lr:0.005000,train Set Accuracy: 30.000000
handwrite: 0, iterate times: 100 , cross entropy:4.47434
predict sum is: 42448.000000,Testing Set Accuracy: 32.701818
lr:0.005000,train Set Accuracy: 73.333333
handwrite: 0, iterate times: 200 , cross entropy:3.06059
predict sum is: 29849.000000,Testing Set Accuracy: 55.314545
lr:0.005000,train Set Accuracy: 63.333333
handwrite: 0, iterate times: 300 , cross entropy:-0.327935
predict sum is: 32571.000000,Testing Set Accuracy: 48.732727
lr:0.005000,train Set Accuracy: 46.666667
handwrite: 0, iterate times: 400 , cross entropy:-54.8765
predict sum is: 27388.000000,Testing Set Accuracy: 55.378182
lr:0.005000,train Set Accuracy: 73.333333
handwrite: 0, iterate times: 500 , cross entropy:nan
predict sum is: 0.000000,Testing Set Accuracy: 90.101818
is nan, stop
lr:0.005000,train Set Accuracy: 13.333333
handwrite: 1, iterate times: 0 , cross entropy:4.49648
predict sum is: 55000.000000,Testing Set Accuracy: 11.234545
lr:0.005000,train Set Accuracy: 23.333333
handwrite: 1, iterate times: 100 , cross entropy:2.75025
predict sum is: 39640.000000,Testing Set Accuracy: 39.125455
lr:0.005000,train Set Accuracy: 60.000000
handwrite: 1, iterate times: 200 , cross entropy:2.9515
predict sum is: 27647.000000,Testing Set Accuracy: 60.712727
lr:0.005000,train Set Accuracy: 50.000000
handwrite: 1, iterate times: 300 , cross entropy:0.0803952
predict sum is: 32210.000000,Testing Set Accuracy: 52.241818
lr:0.005000,train Set Accuracy: 53.333333
handwrite: 1, iterate times: 400 , cross entropy:-81.9736
predict sum is: 30721.000000,Testing Set Accuracy: 54.490909
lr:0.005000,train Set Accuracy: 73.333333
handwrite: 1, iterate times: 500 , cross entropy:nan
predict sum is: 0.000000,Testing Set Accuracy: 88.765455
is nan, stop
lr:0.005000,train Set Accuracy: 10.000000
handwrite: 2, iterate times: 0 , cross entropy:3.94733
predict sum is: 55000.000000,Testing Set Accuracy: 9.945455
lr:0.005000,train Set Accuracy: 20.000000
handwrite: 2, iterate times: 100 , cross entropy:3.93083
predict sum is: 52604.000000,Testing Set Accuracy: 14.280000
lr:0.005000,train Set Accuracy: 66.666667
handwrite: 2, iterate times: 200 , cross entropy:2.31273
predict sum is: 25489.000000,Testing Set Accuracy: 62.140000
lr:0.005000,train Set Accuracy: 70.000000
handwrite: 2, iterate times: 300 , cross entropy:-1.51881
predict sum is: 22049.000000,Testing Set Accuracy: 66.881818
lr:0.005000,train Set Accuracy: 43.333333
handwrite: 2, iterate times: 400 , cross entropy:-59.7042
predict sum is: 29764.000000,Testing Set Accuracy: 52.989091
lr:0.005000,train Set Accuracy: 80.000000
handwrite: 2, iterate times: 500 , cross entropy:nan
predict sum is: 0.000000,Testing Set Accuracy: 90.054545
is nan, stop
lr:0.005000,train Set Accuracy: 13.333333
handwrite: 3, iterate times: 0 , cross entropy:4.51146
predict sum is: 55000.000000,Testing Set Accuracy: 10.250909
lr:0.005000,train Set Accuracy: 26.666667
handwrite: 3, iterate times: 100 , cross entropy:4.49274
predict sum is: 43234.000000,Testing Set Accuracy: 31.498182
lr:0.005000,train Set Accuracy: 53.333333
handwrite: 3, iterate times: 200 , cross entropy:3.11949
predict sum is: 33806.000000,Testing Set Accuracy: 47.530909
lr:0.005000,train Set Accuracy: 40.000000
handwrite: 3, iterate times: 300 , cross entropy:1.82847
predict sum is: 44129.000000,Testing Set Accuracy: 29.009091
lr:0.005000,train Set Accuracy: 60.000000
handwrite: 3, iterate times: 400 , cross entropy:-50.3269
predict sum is: 25410.000000,Testing Set Accuracy: 57.810909
lr:0.005000,train Set Accuracy: 76.666667
handwrite: 3, iterate times: 500 , cross entropy:nan
predict sum is: 0.000000,Testing Set Accuracy: 89.749091
is nan, stop
lr:0.005000,train Set Accuracy: 13.333333
handwrite: 4, iterate times: 0 , cross entropy:4.48332
predict sum is: 55000.000000,Testing Set Accuracy: 9.649091
lr:0.005000,train Set Accuracy: 96.666667
handwrite: 4, iterate times: 100 , cross entropy:4.01762
predict sum is: 3406.000000,Testing Set Accuracy: 93.980000
lr:0.005000,train Set Accuracy: 36.666667
handwrite: 4, iterate times: 200 , cross entropy:4.7305
predict sum is: 46868.000000,Testing Set Accuracy: 24.307273
lr:0.005000,train Set Accuracy: 36.666667
handwrite: 4, iterate times: 300 , cross entropy:0.819301
predict sum is: 32130.000000,Testing Set Accuracy: 50.052727
lr:0.005000,train Set Accuracy: 63.333333
handwrite: 4, iterate times: 400 , cross entropy:-63.146
predict sum is: 26222.000000,Testing Set Accuracy: 59.147273
lr:0.005000,train Set Accuracy: 76.666667
handwrite: 4, iterate times: 500 , cross entropy:nan
predict sum is: 0.000000,Testing Set Accuracy: 90.350909
is nan, stop
lr:0.005000,train Set Accuracy: 16.666667
handwrite: 5, iterate times: 0 , cross entropy:4.98377
predict sum is: 55000.000000,Testing Set Accuracy: 9.067273
lr:0.005000,train Set Accuracy: 56.666667
handwrite: 5, iterate times: 100 , cross entropy:2.67669
predict sum is: 24081.000000,Testing Set Accuracy: 62.716364
lr:0.005000,train Set Accuracy: 40.000000
handwrite: 5, iterate times: 200 , cross entropy:3.18606
predict sum is: 29414.000000,Testing Set Accuracy: 53.245455
lr:0.005000,train Set Accuracy: 56.666667
handwrite: 5, iterate times: 300 , cross entropy:2.15873
predict sum is: 30405.000000,Testing Set Accuracy: 51.327273
lr:0.005000,train Set Accuracy: 66.666667
handwrite: 5, iterate times: 400 , cross entropy:-39.4382
predict sum is: 28597.000000,Testing Set Accuracy: 53.967273
lr:0.005000,train Set Accuracy: 70.000000
handwrite: 5, iterate times: 500 , cross entropy:nan
predict sum is: 0.000000,Testing Set Accuracy: 90.932727
is nan, stop
lr:0.005000,train Set Accuracy: 96.666667
handwrite: 6, iterate times: 0 , cross entropy:2.54767
predict sum is: 0.000000,Testing Set Accuracy: 90.150909
lr:0.005000,train Set Accuracy: 26.666667
handwrite: 6, iterate times: 100 , cross entropy:5.00434
predict sum is: 49145.000000,Testing Set Accuracy: 20.490909
lr:0.005000,train Set Accuracy: 43.333333
handwrite: 6, iterate times: 200 , cross entropy:4.33006
predict sum is: 34103.000000,Testing Set Accuracy: 47.705455
lr:0.005000,train Set Accuracy: 70.000000
handwrite: 6, iterate times: 300 , cross entropy:1.52785
predict sum is: 21247.000000,Testing Set Accuracy: 69.029091
lr:0.005000,train Set Accuracy: 50.000000
handwrite: 6, iterate times: 400 , cross entropy:-38.3298
predict sum is: 29329.000000,Testing Set Accuracy: 54.025455
lr:0.005000,train Set Accuracy: 83.333333
handwrite: 6, iterate times: 500 , cross entropy:nan
predict sum is: 0.000000,Testing Set Accuracy: 90.150909
is nan, stop
lr:0.005000,train Set Accuracy: 100.000000
handwrite: 7, iterate times: 0 , cross entropy:1.65013
predict sum is: 0.000000,Testing Set Accuracy: 89.609091
lr:0.005000,train Set Accuracy: 80.000000
handwrite: 7, iterate times: 100 , cross entropy:3.27103
predict sum is: 18348.000000,Testing Set Accuracy: 76.045455
lr:0.005000,train Set Accuracy: 33.333333
handwrite: 7, iterate times: 200 , cross entropy:2.51653
predict sum is: 35404.000000,Testing Set Accuracy: 45.529091
lr:0.005000,train Set Accuracy: 43.333333
handwrite: 7, iterate times: 300 , cross entropy:0.587267
predict sum is: 39636.000000,Testing Set Accuracy: 37.401818
lr:0.005000,train Set Accuracy: 56.666667
handwrite: 7, iterate times: 400 , cross entropy:-73.8344
predict sum is: 27259.000000,Testing Set Accuracy: 57.520000
lr:0.005000,train Set Accuracy: 76.666667
handwrite: 7, iterate times: 500 , cross entropy:nan
predict sum is: 0.000000,Testing Set Accuracy: 89.609091
is nan, stop
lr:0.005000,train Set Accuracy: 13.333333
handwrite: 8, iterate times: 0 , cross entropy:4.49435
predict sum is: 55000.000000,Testing Set Accuracy: 9.798182
lr:0.005000,train Set Accuracy: 96.666667
handwrite: 8, iterate times: 100 , cross entropy:2.4589
predict sum is: 304.000000,Testing Set Accuracy: 89.921818
lr:0.005000,train Set Accuracy: 13.333333
handwrite: 8, iterate times: 200 , cross entropy:3.77693
predict sum is: 47801.000000,Testing Set Accuracy: 22.636364
lr:0.005000,train Set Accuracy: 63.333333
handwrite: 8, iterate times: 300 , cross entropy:0.305798
predict sum is: 21461.000000,Testing Set Accuracy: 66.250909
lr:0.005000,train Set Accuracy: 36.666667
handwrite: 8, iterate times: 400 , cross entropy:-45.1836
predict sum is: 27745.000000,Testing Set Accuracy: 55.883636
lr:0.005000,train Set Accuracy: 66.666667
handwrite: 8, iterate times: 500 , cross entropy:nan
predict sum is: 0.000000,Testing Set Accuracy: 90.201818
is nan, stop
lr:0.005000,train Set Accuracy: 100.000000
handwrite: 9, iterate times: 0 , cross entropy:1.70039
predict sum is: 0.000000,Testing Set Accuracy: 90.083636
lr:0.005000,train Set Accuracy: 16.666667
handwrite: 9, iterate times: 100 , cross entropy:4.93745
predict sum is: 53646.000000,Testing Set Accuracy: 12.334545
lr:0.005000,train Set Accuracy: 53.333333
handwrite: 9, iterate times: 200 , cross entropy:3.10768
predict sum is: 39225.000000,Testing Set Accuracy: 38.001818
lr:0.005000,train Set Accuracy: 60.000000
handwrite: 9, iterate times: 300 , cross entropy:1.00401
predict sum is: 29131.000000,Testing Set Accuracy: 54.281818
lr:0.005000,train Set Accuracy: 43.333333
handwrite: 9, iterate times: 400 , cross entropy:-69.1324
predict sum is: 27424.000000,Testing Set Accuracy: 56.832727
lr:0.005000,train Set Accuracy: 76.666667
handwrite: 9, iterate times: 500 , cross entropy:nan
predict sum is: 0.000000,Testing Set Accuracy: 90.083636
is nan, stop