如何处理张量流中0-1范围外的输入？

Question

在http://www.tensorflow.org/get_started中提供的示例中,如果我将输入乘以2

x_data = np.float32(np.random.rand(2, 100))*2

我得到无意义的输出,而我期望获得相同的解决方案.

0 [[ -67.06586456 -109.13352203]] [-7.67297792]
20 [[ nan  nan]] [ nan]
40 [[ nan  nan]] [ nan]
60 [[ nan  nan]] [ nan]
80 [[ nan  nan]] [ nan]
100 [[ nan  nan]] [ nan]
120 [[ nan  nan]] [ nan]
140 [[ nan  nan]] [ nan]
160 [[ nan  nan]] [ nan]
180 [[ nan  nan]] [ nan]
200 [[ nan  nan]] [ nan]

tensorflow如何处理不在0-1范围内的输入？

编辑:使用AdagradOptimizer没有问题的作品.

Answer 1

问题是该示例使用了非常积极的学习率:

optimizer = tf.train.GradientDescentOptimizer(0.5)

这使得学习更快,但如果您稍微改变问题则停止工作.学习率0.01更典型:

optimizer = tf.train.GradientDescentOptimizer(0.01)

现在您的修改工作正常.:)