用于回归的tensorflow深层神经网络总是可以在一批中预测相同的结果

Question

我使用张量流来实现简单的多层感知器进行回归。该代码是从标准mnist分类器修改而来的，我只将输出成本更改为MSE（使用tf.reduce_mean(tf.square(pred-y))），并更改了一些输入，输出大小设置。但是，如果我使用回归训练网络，则在几个时期之后，输出批次将完全相同。例如：

target: 48.129, estimated: 42.634
target: 46.590, estimated: 42.634
target: 34.209, estimated: 42.634
target: 69.677, estimated: 42.634
......

我尝试使用sklearn.preprocessing.scale尝试不同的批处理大小，不同的初始化和输入归一化（我的输入范围有很大不同）。但是，它们都不起作用。我还尝试了Tensorflow中的sklearn示例之一（使用波士顿数据进行深度神经网络回归）。但是在第40行中出现了另一个错误：

“模块”对象没有属性“ infer_real_valued_columns_from_input”

有人知道问题出在哪里吗？谢谢

我的代码在下面列出，可能有点长，但是很简单：

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import tensorflow as tf
from tensorflow.contrib import learn
import matplotlib.pyplot as plt

from sklearn.pipeline import Pipeline
from sklearn import datasets, linear_model
from sklearn import cross_validation
import numpy as np

boston = learn.datasets.load_dataset('boston')
x, y = boston.data, boston.target
X_train, X_test, Y_train, Y_test = cross_validation.train_test_split(
x, y, test_size=0.2, random_state=42)

total_len = X_train.shape[0]

# Parameters
learning_rate = 0.001
training_epochs = 500
batch_size = 10
display_step = 1
dropout_rate = 0.9
# Network Parameters
n_hidden_1 = 32 # 1st layer number of features
n_hidden_2 = 200 # 2nd layer number of features
n_hidden_3 = 200
n_hidden_4 = 256
n_input = X_train.shape[1]
n_classes = 1

# tf Graph input
x = tf.placeholder("float", [None, 13])
y = tf.placeholder("float", [None])

# Create model
def multilayer_perceptron(x, weights, biases):
    # Hidden layer with RELU activation
    layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1'])
    layer_1 = tf.nn.relu(layer_1)

    # Hidden layer with RELU activation
    layer_2 = tf.add(tf.matmul(layer_1, weights['h2']), biases['b2'])
    layer_2 = tf.nn.relu(layer_2)

    # Hidden layer with RELU activation
    layer_3 = tf.add(tf.matmul(layer_2, weights['h3']), biases['b3'])
    layer_3 = tf.nn.relu(layer_3)

    # Hidden layer with RELU activation
    layer_4 = tf.add(tf.matmul(layer_3, weights['h4']), biases['b4'])
    layer_4 = tf.nn.relu(layer_4)

    # Output layer with linear activation
    out_layer = tf.matmul(layer_4, weights['out']) + biases['out']
    return out_layer

# Store layers weight & bias
weights = {
    'h1': tf.Variable(tf.random_normal([n_input, n_hidden_1], 0, 0.1)),
    'h2': tf.Variable(tf.random_normal([n_hidden_1, n_hidden_2], 0, 0.1)),
    'h3': tf.Variable(tf.random_normal([n_hidden_2, n_hidden_3], 0, 0.1)),
    'h4': tf.Variable(tf.random_normal([n_hidden_3, n_hidden_4], 0, 0.1)),
    'out': tf.Variable(tf.random_normal([n_hidden_4, n_classes], 0, 0.1))
}
biases = {
    'b1': tf.Variable(tf.random_normal([n_hidden_1], 0, 0.1)),
    'b2': tf.Variable(tf.random_normal([n_hidden_2], 0, 0.1)),
    'b3': tf.Variable(tf.random_normal([n_hidden_3], 0, 0.1)),
    'b4': tf.Variable(tf.random_normal([n_hidden_4], 0, 0.1)),
    'out': tf.Variable(tf.random_normal([n_classes], 0, 0.1))
}

# Construct model
pred = multilayer_perceptron(x, weights, biases)

# Define loss and optimizer
cost = tf.reduce_mean(tf.square(pred-y))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)

# Launch the graph
with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())

    # Training cycle
    for epoch in range(training_epochs):
        avg_cost = 0.
        total_batch = int(total_len/batch_size)
        # Loop over all batches
        for i in range(total_batch-1):
            batch_x = X_train[i*batch_size:(i+1)*batch_size]
            batch_y = Y_train[i*batch_size:(i+1)*batch_size]
            # Run optimization op (backprop) and cost op (to get loss value)
            _, c, p = sess.run([optimizer, cost, pred], feed_dict={x: batch_x,
                                                          y: batch_y})
            # Compute average loss
            avg_cost += c / total_batch

        # sample prediction
        label_value = batch_y
        estimate = p
        err = label_value-estimate
        print ("num batch:", total_batch)

        # Display logs per epoch step
        if epoch % display_step == 0:
            print ("Epoch:", '%04d' % (epoch+1), "cost=", \
                "{:.9f}".format(avg_cost))
            print ("[*]----------------------------")
            for i in xrange(3):
                print ("label value:", label_value[i], \
                    "estimated value:", estimate[i])
            print ("[*]============================")

    print ("Optimization Finished!")

    # Test model
    correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
    # Calculate accuracy
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
    print ("Accuracy:", accuracy.eval({x: X_test, y: Y_test}))

Answer 1

简短答案 ：

移调您pred使用向量tf.transpose(pred)。

更长的答案 ：

问题在于pred（预测）和y（标签）的形状不同：一个是行向量，另一个是列向量。显然，当对它们应用按元素进行操作时，将得到一个矩阵，这不是您想要的。

解决方案是使用来转置预测向量，tf.transpose()以获得适当的向量，从而获得适当的损失函数。实际上，如果在示例中将批处理大小设置为1，则即使没有修复，它也会起作用，因为转置1x1向量是无操作的。

我将此修复程序应用于您的示例代码，并观察到以下行为。修复之前：

Epoch: 0245 cost= 84.743440580
[*]----------------------------
label value: 23 estimated value: [ 27.47437096]
label value: 50 estimated value: [ 24.71126747]
label value: 22 estimated value: [ 23.87785912]

并在同一时间点修复后：

Epoch: 0245 cost= 4.181439120
[*]----------------------------
label value: 23 estimated value: [ 21.64333534]
label value: 50 estimated value: [ 48.76105118]
label value: 22 estimated value: [ 24.27996063]

您会看到成本要低得多，并且实际上它正确地获知了值50。您必须对学习率进行一些微调，以提高您的学习成绩。