Fixed issue in K.logsumexp in tensorflow backend

[linxihui]

Argument reduction_indices is depreciated in tensorflow. It also causes shape inference issue for placeholder, example

x = K.placeholder((None, 9, 9))

y1 = tf.reduce_logsumexp(x, reduction_indices=2)
K.int_shape(y1)  # gives None

y2 = tf.reduce_logsumexp(x, axis=2)
K.int_shape(y2)  # gives (None, 9)

[fchollet]

Correct, thank you.

[gitfourteen]

What is the proper way to use K.logsumexp for now? Seems like there is no such things