我是新来TensorFlow和神经网络在一般情况下,我想开发一个神经网络,可以预测一个属性的值(这是在Kaggle.com上开始的比赛),我知道使用神经网络可能不是解决回归问题的最佳模型,但我决定尝试一下。Tensorflow与1-hdden层预测神经网络不会改变 - 回归
当使用单层神经网络(没有隐藏层,这可能是一个线性回归)时,模型实际上预测值接近实际值,但是当我添加一个隐藏层时,预测的所有值都是相同的批次的20个输入张量:
('real', array([[ 181000.],
[ 128900.],
[ 161500.],
[ 180500.],
[ 181000.],
[ 183900.],
[ 122000.],
[ 378500.],
[ 381000.],
[ 144000.],
[ 260000.],
[ 185750.],
[ 137000.],
[ 177000.],
[ 139000.],
[ 137000.],
[ 162000.],
[ 197900.],
[ 237000.],
[ 68400.]]))
('prediction ', array([[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687],
[ 4995.10597687]]))
更新: 我注意到,预测值只反射输出层的偏见,而两者的隐含层和输出层的重量没有变化,并始终零点
为了进一步检查发生了什么问题,我生成了模型的图(一次使用隐藏层,另一次使用隐藏层)比较两个图,看看是否有某些东西丢失,不幸的是它们都是看起来是正确的我,但我还是不明白,为什么样板工程时,有没有隐藏的图层并采用了隐藏层
我的全代码如下:
# coding: utf-8
import tensorflow as tf
import numpy as np
def loadDataFromCSV(fileName , numberOfFields , numberOfOutputFields , numberOfRecords):
XsArray = np.ndarray([numberOfRecords ,(numberOfFields-numberOfOutputFields)] , dtype=np.float64)
YsArray = np.ndarray([numberOfRecords ,numberOfOutputFields] , dtype=np.float64)
fileQueue = tf.train.string_input_producer(fileName)
defaultValues = [[0]]*numberOfFields
decodedLine = [[None]]*numberOfFields
reader = tf.TextLineReader()
key , singleLine = reader.read(fileQueue)
decodedLine = tf.decode_csv(singleLine,record_defaults=defaultValues)
inputFeatures = decodedLine[0:numberOfFields-numberOfOutputFields]
outputFeatures =decodedLine[numberOfFields-numberOfOutputFields:numberOfFields]
with tf.Session() as session :
tf.global_variables_initializer().run()
coor = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coor)
for i in range(numberOfRecords) :
XsArray[i,:] ,YsArray[i,:] = session.run([inputFeatures , outputFeatures])
coor.request_stop()
coor.join(threads)
return XsArray , YsArray
x , y =loadDataFromCSV(['/Users/mousaalsulaimi/Downloads/convertcsv.csv'] , 289 , 1, 1460)
num_steps = 10000
batch_size = 20
graph = tf.Graph()
with graph.as_default() :
with tf.name_scope('input'):
inputProperties = tf.placeholder(tf.float32 , shape=(batch_size ,287))
with tf.name_scope('realPropertyValue') :
outputValues = tf.placeholder(tf.float32,shape=(batch_size,1))
with tf.name_scope('weights'):
hidden1_w = tf.Variable(tf.truncated_normal([287,1000],stddev=math.sqrt(3/(287+1000)) , dtype=tf.float32))
with tf.name_scope('baises'):
hidden1_b = tf.Variable(tf.zeros([1000] , dtype=tf.float32))
with tf.name_scope('hidden_layer'):
hidden1 =tf.matmul(inputProperties,hidden1_w) + hidden1_b
#hidden1_relu = tf.nn.relu(hidden1)
#hidden1_dropout = tf.nn.dropout(hidden1_relu,.5)
with tf.name_scope('layer2_weights'):
output_w = tf.Variable(tf.truncated_normal([1000,1],stddev=math.sqrt(3/(1000+1)) , dtype=tf.float32))
with tf.name_scope('layer2_baises'):
output_b = tf.Variable(tf.zeros([1] , dtype=tf.float32))
with tf.name_scope('layer_2_predictions'):
output =tf.matmul(hidden1,output_w) + output_b
with tf.name_scope('predictions'):
predictedValues = (output)
loss = tf.sqrt(tf.reduce_mean(tf.square(predictedValues-outputValues)))
loss_l2 = tf.nn.l2_loss(hidden1_w)
with tf.name_scope('minimization') :
minimum = tf.train.AdamOptimizer(.5).minimize(loss+.004*loss_l2)
with tf.Session(graph=graph) as session:
tf.global_variables_initializer().run()
print("Initialized")
for step in range(num_steps):
# Pick an offset within the training data, which has been randomized.
# Note: we could use better randomization across epochs.
offset = (step * batch_size) % (y.shape[0] - batch_size)
# Generate a minibatch.
batch_data = x[offset:(offset + batch_size), 1:]
batch_labels = y[offset:(offset + batch_size), :]
print("real" , batch_labels)
# Prepare a dictionary telling the session where to feed the minibatch.
# The key of the dictionary is the placeholder node of the graph to be fed,
# and the value is the numpy array to feed to it.
feed_dict = {inputProperties : batch_data, outputValues : batch_labels}
_, l, predictions , inp = session.run([minimum, loss, predictedValues ,inputProperties ], feed_dict=feed_dict)
print("prediction " , predictions)
print("loss : " , l)
print("----------")
print('+++++++++++')
也是我在的情况下上传数据文件convertcsv.csv here要看一看。
我很感谢任何帮助弄清楚我做错了什么。
谢谢
我不认为这些是导致性能不佳的原因,但我注意到了3件事:首先,您使用'hidden1'而不是'hidden_dropout'来定义'output',所以您基本上只是在做线性回归,因为层之间没有激活功能。其次,您可能想要将'output_w'的正则化添加到'loss_l2'。最后,32位通常绰绰有余,因此明确使用64位浮点数可能没有什么区别。 – Styrke
您也可以尝试权重的初始化。如果使用Xavier初始化,标准偏差应该是'sqrt(3。/(in + out))'。对于'output_w',''hidden1_w'和'sqrt(3。/(1000 + 1))'是'sqrt(3。/(287 + 1000))'。 – Styrke
谢谢Styrke,我删除了relu激活函数和退出,因为我认为他们导致问题的地方,我刚刚返回他们,我也尝试了Xavier initalization,正如您所建议的那样,但没有改变,输出层仍然不能正确预测任何事物。 –