tensorflow tf.contrib.learn DNN Regressor internals

Question

我尝试使用tensorflow从tf.contrib.learn重新创建DNNRegressor模型，但我的损失高出6个数量级。有人能指出我正确的方向吗？我不知道出了什么问题或有什么不同：/数据在这里是否有帮助http://pastebin.com/BG6r6EF6

  

tf.contrib.learn代码：

data = np.loadtxt('training.csv',
    delimiter=',',skiprows=1,usecols = (3,4,5,6,7,8,9,10,11,12,13,14,15,16,17)
   ,dtype=np.float32)

X_ = data[:,:-1]
Y_ = data[:,-1]

feature_columns = [tf.contrib.layers.real_valued_column("", dimension=14)]

classifier = tf.contrib.learn.DNNRegressor(feature_columns=feature_columns,
    hidden_units=[7],
    optimizer=tf.train.RMSPropOptimizer(learning_rate=.001),
    activation_fn=tf.nn.relu)

classifier.fit(x=X_,
               y=Y_,
               max_steps=1000)

  

tensorflow代码：

data = np.loadtxt('training.csv',
    delimiter=',',skiprows=1,usecols = (3,4,5,6,7,8,9,10,11,12,13,14,15,16,17)
    ,dtype=np.float32)

n_features = 14
hidden_units = 7
n_classes = 1
lr = .001

X = tf.placeholder(tf.float32,[None,n_features])
Y = tf.placeholder(tf.float32,[None])

W = tf.Variable(tf.truncated_normal([n_features,hidden_units]))
W2 = tf.Variable(tf.truncated_normal([hidden_units,n_classes]))    
b = tf.Variable(tf.zeros([hidden_units]))
b2 = tf.Variable(tf.zeros([n_classes]))

hidden1 = tf.nn.relu(tf.matmul(X,W) + b)
pred = tf.matmul(hidden1,W2)+b2

#I have tried a few variations of squared error loss with no luck 

loss = tf.nn.l2_loss(pred - Y)

    #loss = tf.reduce_sum(tf.pow(pred - Y,2))/(2*n_instances)
    #loss = tf.reduce_mean(tf.squared_difference(pred, Y))
    #loss = tf.reduce_sum(tf.pow(pred - Y,2))/(2*n_instances)

optimizer = tf.train.RMSPropOptimizer(lr).minimize(loss)

with tf.Session() as sess:
    init = tf.global_variables_initializer()
    sess.run(init)

    for step in range(1000):
       _, loss_value = sess.run([optimizer,loss],
                feed_dict={X: X_,Y: Y_} )

更新

我改为

loss = tf.reduce_mean(tf.squared_difference(pred, Y)) 

现在两种方法的损失大致相同（~200）。张量流模型非常不准确，但DNNRegressor输出使用验证数据时的预期。张量板图也非常不同。

DNNRegressor：

Tensorflow：

Answer 1

我将使用张量板比较两个模型的图形。你试过吗？