The easiest way is to change the keep_prob parameter using a placeholder_with_default:
prob = tf.placeholder_with_default(1.0, shape=())
layer = tf.nn.dropout(layer, prob)
in this way when you train you can set the parameter like this:
sess.run(train_step, feed_dict={prob: 0.5})
and when you evaluate the default value of 1.0 is used.