Tensorboard使用說明

1. Graph（生成神經網絡的資料流圖）

   writer = tf.summary.Filewriter(“/tmp/mnist-demo/2”)

   writer.add_graph(sess.graph)

2. Summary（記錄網絡訓練過程中各種變量的變化情況）

   – tf.summay.scalar

   – tf.summay.image

   – tf.summay.audio

   – tf.summay.histogram

   – tf.summay.tensor(under development)

   用法：

   tf.summay.scalar(‘cross_entropy’, xent)

   tf.summay.image(‘accuracy’, accuracy)

   tf.summay.audio(‘input’, x_image, 3)

   tf.summay.histogram(”weights”, w)

   tf.summay.histogram(“bias”, b)

   合并：

   merged_summary = tf.summary.merge_all()

   程式示例：

merged_summary = tf.summary_merge_all()
    writer = tf.summary.FileWriter(logdir) 
    writer.add_graph(sess.graph)
    for i in range():
        batch = mnist.train.next_batch()
        if i %  == :
            s = sess.run(merged_summary, feed_dict={x: batch[], y:batch[]})
            writer.add_summay(s, i)
        sess.run(train_step, feed_dict={x: batch[], y: batch[]})
           

1. Hyperparameters search（設定多組超參數進行訓練，對比不同超參數訓練結果，用以挑選最好的超參數組合）

#try a few learning rate
    for learning_rate in [, , ]:
        #Try a model with fewer layers
        for ues_two_fc in [True, False]:
            for use_two_conv in [True, False]:
                #Construct a hyperparameters string for each one(example: "lr=1e-3, fc=2, conv=2")
                haparam_str = make_hparam_string(learning_rate, use_two_fc, use_two_conv)
                writer = tf.summary.FileWriter("/tmp/mnist_tutorial/" + hparam_str)
           

1. Embeding Visualizer（将高維資料投影至三維空間，友善直覺的觀察資料的分布情況，如可以友善的檢視網絡進行預測時誤比對樣本的具體情況）