TensorFlow #Neural Network ? 1 with TB：Name Scope

Neural Network ? 1 にTensorBoardのモニタ解析を追記しておく。
今回は、name_scope()を用いて、Nameスペースを有効に設定しいたことで、グラフ構造が整理されて解りやすくなった。
隠れ層200ニューロンにアップ。

import tensorflow as tf
import tensorflow.contrib.learn as learn
 
# MNISTデータの読み込み
dataset = learn.datasets.mnist.read_data_sets('MNIST-data', one_hot=True)
 
# MNISTデータと出力結果を収めるプレースホルダー設定　
image = tf.placeholder(tf.float32, [None, 784]) #28x28行列要素を収める
label = tf.placeholder(tf.float32, [None, 10]) #0〜9の数字ラベルを収める
 
# レイヤーのノード数設定
hidden_nodes = 200
output_nodes = 10

with tf.name_scope('interface') as scope:
    step_var = tf.Variable(0, trainable=False)
    
    # 重み係数を変数として定義
    w0 = tf.Variable(tf.random_normal([784, hidden_nodes]))
    w1 = tf.Variable(tf.random_normal([hidden_nodes, hidden_nodes]))
    w2 = tf.Variable(tf.random_normal([hidden_nodes, hidden_nodes]))
    w3 = tf.Variable(tf.random_normal([hidden_nodes, output_nodes]))
 
    # バイアス値を変数として定義
    b0 = tf.Variable(tf.random_normal([hidden_nodes]))
    b1 = tf.Variable(tf.random_normal([hidden_nodes]))
    b2 = tf.Variable(tf.random_normal([hidden_nodes]))
    b3 = tf.Variable(tf.random_normal([output_nodes]))

with tf.name_scope('layers') as scope:
    # レイヤー作成：活性化関数はRelu
    layer_1 = tf.add(tf.matmul(image, w0), b0)
    layer_1 = tf.nn.relu(layer_1)
    layer_2 = tf.add(tf.matmul(layer_1, w1), b1)
    layer_2 = tf.nn.relu(layer_2)
    layer_3 = tf.add(tf.matmul(layer_2, w2), b2)
    layer_3 = tf.nn.relu(layer_3)
    output_layer = tf.matmul(layer_3, w3) + b3
 
with tf.name_scope('loss_func') as scope:
    # 損失関数
    loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=output_layer, labels=label))
    summary_op = tf.summary.scalar('loss', loss)

with tf.name_scope('training') as scope:
    # 最適化関数
    learning_rate = 0.01
    num_epochs = 20
    batch_size = 100
    num_batches = int(dataset.train.num_examples/batch_size)
    optimizer = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=step_var)

summary = tf.summary.merge_all()

# セッション立ち上げ
with tf.Session() as sess:
    writer = tf.summary.FileWriter("logs", graph=sess.graph)
    sess.run(tf.global_variables_initializer())
 
    # エポックループ
    for epoch in range(num_epochs):
 
        # バッチループ
        for batch in range(num_batches):
            image_batch, label_batch = dataset.train.next_batch(batch_size)
            step = sess.run(step_var)
            #セッションを動かす：最適化関数にデータをフィード
            sess.run(optimizer, feed_dict={image: image_batch, label: label_batch})
            summary_op2 = sess.run(summary_op, feed_dict={image: image_batch, label: label_batch})  
            writer.add_summary(summary_op2, global_step=step)
            
    # 成功率算定
    prediction = tf.equal(tf.argmax(output_layer, 1), tf.argmax(label, 1))
    success = tf.reduce_mean(tf.cast(prediction, tf.float32))
    print('Correct %: ', 100 * sess.run(success,
        feed_dict={image: dataset.test.images,
            label: dataset.test.labels}))

import tensorflow as tf

import tensorflow.contrib.learn as learn

# MNISTデータの読み込み

dataset = learn.datasets.mnist.read_data_sets('MNIST-data', one_hot=True)

# MNISTデータと出力結果を収めるプレースホルダー設定　

image = tf.placeholder(tf.float32, [None, 784]) #28x28行列要素を収める

label = tf.placeholder(tf.float32, [None, 10]) #0〜9の数字ラベルを収める

# レイヤーのノード数設定

hidden_nodes = 200

output_nodes = 10

with tf.name_scope('interface') as scope:

step_var = tf.Variable(0, trainable=False)

# 重み係数を変数として定義

w0 = tf.Variable(tf.random_normal([784, hidden_nodes]))

w1 = tf.Variable(tf.random_normal([hidden_nodes, hidden_nodes]))

w2 = tf.Variable(tf.random_normal([hidden_nodes, hidden_nodes]))

w3 = tf.Variable(tf.random_normal([hidden_nodes, output_nodes]))

# バイアス値を変数として定義

b0 = tf.Variable(tf.random_normal([hidden_nodes]))

b1 = tf.Variable(tf.random_normal([hidden_nodes]))

b2 = tf.Variable(tf.random_normal([hidden_nodes]))

b3 = tf.Variable(tf.random_normal([output_nodes]))

with tf.name_scope('layers') as scope:

# レイヤー作成：活性化関数はRelu

layer_1 = tf.add(tf.matmul(image, w0), b0)

layer_1 = tf.nn.relu(layer_1)

layer_2 = tf.add(tf.matmul(layer_1, w1), b1)

layer_2 = tf.nn.relu(layer_2)

layer_3 = tf.add(tf.matmul(layer_2, w2), b2)

layer_3 = tf.nn.relu(layer_3)

output_layer = tf.matmul(layer_3, w3) + b3

with tf.name_scope('loss_func') as scope:

# 損失関数

loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=output_layer, labels=label))

summary_op = tf.summary.scalar('loss', loss)

with tf.name_scope('training') as scope:

# 最適化関数

learning_rate = 0.01

num_epochs = 20

batch_size = 100

num_batches = int(dataset.train.num_examples/batch_size)

optimizer = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=step_var)

summary = tf.summary.merge_all()

# セッション立ち上げ

with tf.Session() as sess:

writer = tf.summary.FileWriter("logs", graph=sess.graph)

sess.run(tf.global_variables_initializer())

# エポックループ

for epoch in range(num_epochs):

# バッチループ

for batch in range(num_batches):

image_batch, label_batch = dataset.train.next_batch(batch_size)

step = sess.run(step_var)

#セッションを動かす：最適化関数にデータをフィード

sess.run(optimizer, feed_dict={image: image_batch, label: label_batch})

summary_op2 = sess.run(summary_op, feed_dict={image: image_batch, label: label_batch})

writer.add_summary(summary_op2, global_step=step)

# 成功率算定

prediction = tf.equal(tf.argmax(output_layer, 1), tf.argmax(label, 1))

success = tf.reduce_mean(tf.cast(prediction, tf.float32))

print('Correct %: ', 100 * sess.run(success,

feed_dict={image: dataset.test.images,

label: dataset.test.labels}))

出力：

Correct %:  93.58999729156494

1	Correct %: 93.58999729156494

今回は、name_scope()を用いて、Nameスペースを有効に設定しいたことで、グラフ構造が整理されて解りやすくなった。

Science To Medicine

Just My Daily Study Note by ts.anesth.kpum

TensorFlow #Neural Network ? 1 with TB：Name Scope