为什么TensorFlow会在训练后预测所有0或全部1？

Question

所以我的问题是我正在运行TensorFlow教程中的初学者级别代码，并根据我的需要对其进行了修改，但是当我创建print sess.run(accuracy, feed_dict={x: x_test, y_: y_test})时，它常常打印出1.0，现在它已经＃39 ; s总是猜测0并打印出约93％的准确度。当我使用tf.argmin(y,1), tf.argmin(y_,1)时，它会猜测所有1并且产生约7％的准确率。将两者相加，它等于100％。我不知道tf.argmin猜测1和tf.argmax猜测0＆＃。显然代码有问题。请看一下，让我知道我可以做些什么来解决这个问题。我认为在培训期间代码出错了，但我错了。

import tensorflow as tf
import numpy as np
from numpy import genfromtxt

data = genfromtxt('cs-training.csv',delimiter=',')  # Training data
test_data = genfromtxt('cs-test.csv',delimiter=',')  # Test data

x_train = []
for i in data:
    x_train.append(i[1:])
x_train = np.array(x_train)

y_train = []
for i in data:
    if i[0] == 0:
        y_train.append([1., i[0]])
    else:
        y_train.append([0., i[0]])
y_train = np.array(y_train)

where_are_NaNs = isnan(x_train)
x_train[where_are_NaNs] = 0

x_test = []
for i in test_data:
    x_test.append(i[1:])
x_test = np.array(x_test)

y_test = []
for i in test_data:
    if i[0] == 0:
        y_test.append([1., i[0]])
    else:
        y_test.append([0., i[0]])
y_test = np.array(y_test)

where_are_NaNs = isnan(x_test)
x_test[where_are_NaNs] = 0

x = tf.placeholder("float", [None, 10])
W = tf.Variable(tf.zeros([10,2]))
b = tf.Variable(tf.zeros([2]))

y = tf.nn.softmax(tf.matmul(x,W) + b)

y_ = tf.placeholder("float", [None,2])

cross_entropy = -tf.reduce_sum(y_*tf.log(y))

train_step = tf.train.GradientDescentOptimizer(0.01).minimize(cross_entropy)

init = tf.initialize_all_variables()

sess = tf.Session()
sess.run(init)

print "...Training..."

g = 0
for i in range(len(x_train)):

    sess.run(train_step, feed_dict={x: [x_train[g]], y_: [y_train[g]]})

    g += 1

此时，如果我将其设为print [x_train[g]]和print [y_train[g]]，则结果如下所示。

[array([  7.66126609e-01,   4.50000000e+01,   2.00000000e+00,
     8.02982129e-01,   9.12000000e+03,   1.30000000e+01,
     0.00000000e+00,   6.00000000e+00,   0.00000000e+00,
     2.00000000e+00])]

[array([ 0.,  1.])]

好的，让我们继续吧。

correct_prediction = tf.equal(tf.argmax(y,1), tf.argmax(y_,1))

accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))

print sess.run(accuracy, feed_dict={x: x_test, y_: y_test})
0.929209

此百分比不会改变。无论我为2个等级（1或0）创建的单人投掷，它都会猜测所有零。

在这里查看数据 -

print x_train[:10]

[[  7.66126609e-01   4.50000000e+01   2.00000000e+00   8.02982129e-01
9.12000000e+03   1.30000000e+01   0.00000000e+00   6.00000000e+00
0.00000000e+00   2.00000000e+00]
 [  9.57151019e-01   4.00000000e+01   0.00000000e+00   1.21876201e-01
2.60000000e+03   4.00000000e+00   0.00000000e+00   0.00000000e+00
0.00000000e+00   1.00000000e+00]
 [  6.58180140e-01   3.80000000e+01   1.00000000e+00   8.51133750e-02
3.04200000e+03   2.00000000e+00   1.00000000e+00   0.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  2.33809776e-01   3.00000000e+01   0.00000000e+00   3.60496820e-02
3.30000000e+03   5.00000000e+00   0.00000000e+00   0.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  9.07239400e-01   4.90000000e+01   1.00000000e+00   2.49256950e-02
6.35880000e+04   7.00000000e+00   0.00000000e+00   1.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  2.13178682e-01   7.40000000e+01   0.00000000e+00   3.75606969e-01
3.50000000e+03   3.00000000e+00   0.00000000e+00   1.00000000e+00
0.00000000e+00   1.00000000e+00]
 [  3.05682465e-01   5.70000000e+01   0.00000000e+00   5.71000000e+03
0.00000000e+00   8.00000000e+00   0.00000000e+00   3.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  7.54463648e-01   3.90000000e+01   0.00000000e+00   2.09940017e-01
3.50000000e+03   8.00000000e+00   0.00000000e+00   0.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  1.16950644e-01   2.70000000e+01   0.00000000e+00   4.60000000e+01
0.00000000e+00   2.00000000e+00   0.00000000e+00   0.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  1.89169052e-01   5.70000000e+01   0.00000000e+00   6.06290901e-01
2.36840000e+04   9.00000000e+00   0.00000000e+00   4.00000000e+00
0.00000000e+00   2.00000000e+00]]

print y_train[:10]

[[ 0.  1.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]]

print x_test[:20]

[[  4.83539240e-02   4.40000000e+01   0.00000000e+00   3.02297622e-01
7.48500000e+03   1.10000000e+01   0.00000000e+00   1.00000000e+00
0.00000000e+00   2.00000000e+00]
 [  9.10224439e-01   4.20000000e+01   5.00000000e+00   1.72900000e+03
0.00000000e+00   5.00000000e+00   2.00000000e+00   0.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  2.92682927e-01   5.80000000e+01   0.00000000e+00   3.66480079e-01
3.03600000e+03   7.00000000e+00   0.00000000e+00   1.00000000e+00
0.00000000e+00   1.00000000e+00]
 [  3.11547538e-01   3.30000000e+01   1.00000000e+00   3.55431993e-01
4.67500000e+03   1.10000000e+01   0.00000000e+00   1.00000000e+00
0.00000000e+00   1.00000000e+00]
 [  0.00000000e+00   7.20000000e+01   0.00000000e+00   2.16630600e-03
6.00000000e+03   9.00000000e+00   0.00000000e+00   0.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  2.79217052e-01   4.50000000e+01   1.00000000e+00   4.89921122e-01
6.84500000e+03   8.00000000e+00   0.00000000e+00   2.00000000e+00
0.00000000e+00   2.00000000e+00]
 [  0.00000000e+00   7.80000000e+01   0.00000000e+00   0.00000000e+00
0.00000000e+00   1.00000000e+00   0.00000000e+00   0.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  9.10363487e-01   2.80000000e+01   0.00000000e+00   4.99451497e-01
6.38000000e+03   8.00000000e+00   0.00000000e+00   2.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  6.36595797e-01   4.40000000e+01   0.00000000e+00   7.85457163e-01
4.16600000e+03   6.00000000e+00   0.00000000e+00   1.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  1.41549211e-01   2.60000000e+01   0.00000000e+00   2.68407434e-01
4.25000000e+03   4.00000000e+00   0.00000000e+00   0.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  4.14101100e-03   7.80000000e+01   0.00000000e+00   2.26362500e-03
5.74200000e+03   7.00000000e+00   0.00000000e+00   0.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  9.99999900e-01   6.00000000e+01   0.00000000e+00   1.20000000e+02
0.00000000e+00   2.00000000e+00   0.00000000e+00   0.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  6.28525944e-01   4.70000000e+01   0.00000000e+00   1.13100000e+03
0.00000000e+00   5.00000000e+00   0.00000000e+00   0.00000000e+00
0.00000000e+00   2.00000000e+00]
 [  4.02283095e-01   6.00000000e+01   0.00000000e+00   3.79442065e-01
8.63800000e+03   1.00000000e+01   0.00000000e+00   1.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  5.70997900e-03   8.10000000e+01   0.00000000e+00   2.17382000e-04
2.30000000e+04   4.00000000e+00   0.00000000e+00   0.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  4.71171849e-01   5.10000000e+01   0.00000000e+00   1.53700000e+03
0.00000000e+00   1.40000000e+01   0.00000000e+00   0.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  1.42395210e-02   8.20000000e+01   0.00000000e+00   7.40466500e-03
2.70000000e+03   1.00000000e+01   0.00000000e+00   0.00000000e+00
0.00000000e+00   0.00000000e+00]
 [  4.67455800e-02   3.70000000e+01   0.00000000e+00   1.48010090e-02
9.12000000e+03   8.00000000e+00   0.00000000e+00   0.00000000e+00
0.00000000e+00   4.00000000e+00]
 [  9.99999900e-01   4.70000000e+01   0.00000000e+00   3.54604127e-01
1.10000000e+04   1.10000000e+01   0.00000000e+00   2.00000000e+00
0.00000000e+00   3.00000000e+00]
 [  8.96417860e-02   2.70000000e+01   0.00000000e+00   8.14664000e-03
5.40000000e+03   6.00000000e+00   0.00000000e+00   0.00000000e+00
0.00000000e+00   0.00000000e+00]]

print y_test[:20]
[[ 1.  0.]
 [ 0.  1.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 0.  1.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]
 [ 1.  0.]]

Answer 1

tl; dr：上面发布的示例代码计算交叉熵的方式在数值上并不健全。请改用tf.nn.cross_entropy_with_logits。

（响应问题的v1，已经发生变化）：我担心您的培训实际上没有完成或工作，基于x_train数据中的nan s你表现出来了我建议先修复一下 - 并确定他们出现并修复该错误的原因，并查看您的测试集中是否还有nan。可能也有助于显示x_test和y_test。

最后，我认为y_处理与x相关的方式存在错误。代码编写为y_是一个热门矩阵，但是当您显示y_train[:10]时，它只有10个元素，而不是10*num_classes个类别。我怀疑那里有一个bug。当你在轴1上进行argmax时，你总是会得到一个充满零的向量（因为那个轴上只有一个元素，所以它当然是最大元素）。将其与估计值产生始终为零的输出错误相结合，并且您始终会生成一个正确的＆＃34;回答。 ：）

修订版的更新 在更改后的版本中，如果您运行它并在每次执行结束时打印出W，请将代码更改为：

 _, w_out, b_out = sess.run([train_step, W, b], feed_dict={x: [x_train[g]], y_: [y_train[g]]})

你会发现W充满了nan s。要对此进行调试，您可以盯着您的代码看看是否存在您可以发现的数学问题，或者您可以通过管道检测它们以查看它们出现的位置。我们来试试吧。首先，cross_entropy是什么？ （将cross_entropy添加到run语句中的事物列表中并将其打印出来

Cross entropy:  inf

大！所以为什么？嗯，一个答案就是：

y = [0, 1]
tf.log(y) = [-inf, 0]

这是y的有效可能输出，但是你的交叉熵计算不稳定。您可以手动添加一些epsilons以避免角落情况，或使用tf.nn.softmax_cross_entropy_with_logits为您执行此操作。我推荐后者：

yprime = tf.matmul(x,W)+b
y = tf.nn.softmax(yprime)
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(yprime, y_)

我不保证您的模型能够正常运行，但这可以解决您当前的NaN问题。