我尝试使用tensorflow重建tf.contrib.learn中的DNNRegressor模型，但我的损失值高出六个数量级。有人能指导我正确的方向吗？我完全不知道哪里出了问题或者有什么不同之处：/ 如果有帮助的话，数据在这里 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 = 14hidden_units = 7n_classes = 1lr = .001X = 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#我尝试了几种平方误差损失的变体，但没有成功 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

我改成了

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

现在两种方法的损失值大致相同（约200）。然而，tensorflow模型的准确性非常低，而DNNRegressor在使用验证数据时输出的结果符合我的预期。Tensorboard的图表也非常不同。

DNNRegressor:

Tensorflow:

回答：

我会使用tensorboard来比较两个模型的图表。你试过吗？