the dropout probability should be different between train and inference

[zhengjxu]

During the training, the keep_prob = 0.5, during inference, the keep_prob should be 1. I update the code by adding the keep_prob as tf.placeholder.