张量流代码精度不增加

Question

我是Tensorflow的新手，并构建了一个具有2个隐藏层的感知器。 我的数据集有 8000 个训练示例，代码如下：-

import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
data=pd.read_csv("Churn_Modelling.csv")
X=data.iloc[:,3:13].values
Y=data.iloc[:,13].values
from sklearn.preprocessing import LabelEncoder,OneHotEncoder
le1=LabelEncoder()
X[:,1]=le1.fit_transform(X[:,1])
le2=LabelEncoder()
X[:,2]=le2.fit_transform(X[:,2])
ohe1=OneHotEncoder(categorical_features=[1])
X=ohe1.fit_transform(X).toarray()
X=X[:,1:]
from sklearn.model_selection import train_test_split
X_train,X_test,y_train,y_test=train_test_split(X,Y,test_size=0.2)
from sklearn.preprocessing import StandardScaler
sc=StandardScaler()
X_train=sc.fit_transform(X_train)
X_test=sc.transform(X_test)
y_train=np.array([[y_train[i]] for i in range(8000)])
y_test=np.array([[y_test[i]] for i in range(2000)])
import tensorflow as tf
hidden1=6
hidden2=6
batch_size=10
x=tf.placeholder(shape=[None,11],dtype=tf.float32)
y=tf.placeholder(shape=[None,1],dtype=tf.float32)
def neural_network(data):
l1={"weights":tf.Variable(tf.random_normal([11,hidden1])),
"biases":tf.Variable(tf.random_normal([hidden1]))}
l2={"weights":tf.Variable(tf.random_normal([hidden1,hidden2])),
"biases":tf.Variable(tf.random_normal([hidden2]))}
output={"weights":tf.Variable(tf.random_normal([hidden2,1])),
"biases":tf.Variable(tf.random_normal([1]))}
hl1=tf.add(tf.matmul(data,l1["weights"]),l1["biases"])
hl1=tf.nn.relu(hl1)
hl2=tf.add(tf.matmul(hl1,l2["weights"]),l2["biases"])
hl2=tf.nn.relu(hl2)
output1=tf.add(tf.matmul(hl2,output["weights"]),output["biases"])
return output1
tcost=[]
def train_neural_network(data):
prediction=neural_network(data)
cost=tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=prediction,labels=y))
#cost=-tf.reduce_mean(y*tf.log(prediction)+(1-y)*tf.log(1-prediction))
optimizer=tf.train.AdamOptimizer(0.01).minimize(cost)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
epochs=100
for epoch in range(epochs):
avg_cost=0
for i in range(batch_size):
start=i
end=start+batch_size
a=X_train[start:end,:]
b=y_train[start:end,:]
_,c=sess.run([optimizer,cost],feed_dict={x:a,y:b})
avg_cost+=c
i=i+batch_size
print("Cost",avg_cost)  
correct= tf.equal(tf.round(prediction), y)    
#            correct=tf.equal(tf.argmax(prediction,1),tf.argmax(y,1))
accuracy=tf.reduce_mean(tf.cast(correct,"float"))
##           my_acc = tf.reduce_mean(tf.cast(tf.equal(y, prediction), tf.float32))
##            print(sess.run(my_acc, feed_dict={x:X_test,y:y_test}))
print("Accuracy",accuracy.eval({x:X_train,y:y_train}))        
tcost.append(avg_cost)
plt.scatter(range(100),tcost,color="blue")
plt.show()            

train_neural_network(x)

无论我做什么，我的网络的准确性都不会超过训练集的 10%。 我尝试修改学习率以及层数和隐藏单元的数量，但无济于事，即使成本函数随着每个时代逐渐减少。

这是我最后得到的：-

成本 0.0032630344212520868
精度 0.015

我是否错误地实现了精度函数？问题出在哪里？

Answer 1

sigmoid_cross_entropy_with_logits此函数仅对训练有用，因为使用 sigmoid 和交叉熵进行了优化，对于预测，您需要在输出中使用sigmoid函数。您将logits与predictions混合在一起，这些是不同的东西。

def neural_network(data):
l1={"weights":tf.Variable(tf.random_normal([11,hidden1])),
"biases":tf.Variable(tf.random_normal([hidden1]))}
l2={"weights":tf.Variable(tf.random_normal([hidden1,hidden2])),
"biases":tf.Variable(tf.random_normal([hidden2]))}
output={"weights":tf.Variable(tf.random_normal([hidden2,1])),
"biases":tf.Variable(tf.random_normal([1]))}
hl1=tf.add(tf.matmul(data,l1["weights"]),l1["biases"])
hl1=tf.nn.relu(hl1)
hl2=tf.add(tf.matmul(hl1,l2["weights"]),l2["biases"])
hl2=tf.nn.relu(hl2)
logits=tf.add(tf.matmul(hl2,output["weights"]),output["biases"])
predictions = tf.sigmoid(logits)
return logtis, predictions
...
logits,prediction=neural_network(data)
cost=tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=logits,labels=y))