Softmax Returns 单位矩阵

Question

我对 softmax、y = tf.nn.softmax(tf.matmul(x, W) + b) 的输入是一个值矩阵

tf.matmul(x, W) + b =
[[  9.77206726e+02]
 [  5.72391296e+02]
 [  3.53560760e+02]
 [  4.75727379e-01]
 [  6.58911804e+02]]

但是当这个输入到 softmax 时，我得到：

tf.nn.softmax(tf.matmul(x, W) + b) =
[[ 1.]
 [ 1.]
 [ 1.]
 [ 1.]
 [ 1.]]

使我的训练输出成为 1 的数组，这意味着 none 的权重 W 或偏差 b 更新每批训练数据。这也导致我在一组随机测试数据

上的准确度为1

下面是我的代码：

x = tf.placeholder(tf.float32, [None, 2])

W = tf.Variable(tf.random_normal([2, 1]))

b = tf.Variable(tf.random_normal([1]))

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

## placeholder for cross-entropy
y_ = tf.placeholder(tf.float32, [None, 1])

## cross-entropy function
cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y), reduction_indices=[1]))

## backpropagation & gradienct descent
train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)

## initialize variables
init = tf.initialize_all_variables()

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

ITER_RANGE = 10
EVAL_BATCH_SIZE = ( len(training_outputs)/ITER_RANGE )
training_outputs = np.reshape(training_outputs, (300, 1))
## training
for i in range(ITER_RANGE):
  print 'iterator:'
  print i

  ## batch out training data
  BEGIN = ( i*EVAL_BATCH_SIZE )
  END = ( (i*EVAL_BATCH_SIZE) + EVAL_BATCH_SIZE )

  batch_ys = training_outputs[BEGIN:END]
  batch_xs = training_inputs[BEGIN:END]

  print 'batch_xs'
  print batch_xs

  print 'batch_ys'
  print batch_ys

  sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})

  # y = tf.nn.softmax(tf.matmul(x, W) + b)
  print 'y'
  print (sess.run(y, feed_dict={x: batch_xs, y_: batch_ys}))

  #print 'x'
  #print sess.run(x)

  print 'W'
  print sess.run(W)

  print 'b'
  print sess.run(b)

  print 'tf.matmul(x, W) + b'
  print sess.run(tf.matmul(x, W) + b, feed_dict={x: batch_xs, y_: batch_ys})

  print 'tf.nn.softmaxtf.matmul(x, W) + b)'
  print sess.run((tf.nn.softmax(tf.matmul(x, W) + b)), feed_dict={x: batch_xs, y_: batch_ys})

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

accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

test_outputs = np.random.rand(300, 1)

## the following prints 1
print(sess.run(accuracy, feed_dict={x: test_inputs, y_: test_outputs}))

Answer 1

根据Softmax的定义，它“'squashes'任意实数值的K维向量到(0, 1)范围内实数值的K维向量加起来等于 1"

如果只有 1 个输出值，那么 Softmax 输出的分类概率分布就是 1，而不是加起来为 1 的值。

Answer 2

您的 softmax 函数似乎已应用于输出向量中的每个不同值。尝试转置您的输出，即将 tf.nn.softmax(tf.matmul(x, W) + b)) 更改为 tf.nn.softmax(tf.transpose(tf.matmul(x, W) + b))).

Answer 3

看起来你只有两个类 {yes, no} 并且 tf.matmul(x, W) + b 代表了 {yes} 的概率。在这种情况下，您应该使用 tf.nn.sigmoid_cross_entropy_with_logits 而不是 softmax。类似于：

y_pred = tf.matmul(x, W) + b
loss = tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(y_pred, y_))
train_step = tf.train.GradientDescentOptimizer(0.5).minimize(loss)

Answer 4

交叉熵损失不完整。使用带有 logits 的交叉熵。

Softmax Returns 单位矩阵

Softmax Returns Identity Matrix

tensorflow

softmax