在TensorFlow培训期间打印丢失

Question

我正在看TensorFlow" MNIST for ML初学者 "教程,我希望在每个训练步骤后打印出训练损失.

我的训练循环目前看起来像这样:

for i in range(100):
    batch_xs, batch_ys = mnist.train.next_batch(100)
    sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})

现在,train_step定义为:

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

cross_entropy我要打印的损失在哪里:

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

打印这种方法的一种方法是cross_entropy在训练循环中明确计算:

for i in range(100):
    batch_xs, batch_ys = mnist.train.next_batch(100)
    cross_entropy = -tf.reduce_sum(y_ * tf.log(y))
    print 'loss = ' + str(cross_entropy)
    sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})

我现在有两个问题:

1. 鉴于cross_entropy已经计算过sess.run(train_step, ...),计算两次似乎效率低,需要两倍于所有训练数据的前向传递次数.有没有办法访问cross_entropy计算期间的值sess.run(train_step, ...)？

2. 我怎么打印tf.Variable？使用str(cross_entropy)给我一个错误......

谢谢!

Answer 1

您可以cross_entropy通过将其添加到参数列表来获取值sess.run(...).例如,您的for-loop可以重写如下:

for i in range(100):
    batch_xs, batch_ys = mnist.train.next_batch(100)
    cross_entropy = -tf.reduce_sum(y_ * tf.log(y))
    _, loss_val = sess.run([train_step, cross_entropy],
                           feed_dict={x: batch_xs, y_: batch_ys})
    print 'loss = ' + loss_val

可以使用相同的方法来打印变量的当前值.比方说,除了值之外cross_entropy,你想打印一个tf.Variable被调用的值W,你可以做以下事情:

for i in range(100):
    batch_xs, batch_ys = mnist.train.next_batch(100)
    cross_entropy = -tf.reduce_sum(y_ * tf.log(y))
    _, loss_val, W_val = sess.run([train_step, cross_entropy, W],
                                  feed_dict={x: batch_xs, y_: batch_ys})
    print 'loss = %s' % loss_val
    print 'W = %s' % W_val