TensorBoard Sample

tensorboard-sample1.py

import numpy as np
import tensorflow as tf

# セッションを生成
sess = tf.InteractiveSession()

# トレーニングするパラメータを定義
A = tf.Variable(tf.zeros(shape=[3, 3]), dtype=tf.float32, name="A")
b = tf.Variable(tf.zeros(shape=[3, 1]), dtype=tf.float32, name="b")
sess.run(tf.global_variables_initializer())

# Define placeholder for training input
# 学習用入力データのプレースホルダの定義
x = tf.placeholder(tf.float32, shape=[3, 1], name="x")
# 学習用出力データのプレースホルダの定義
y = tf.placeholder(tf.float32, shape=[3, 1], name="y")

# 機械学習のモデルを定義
y_ = tf.matmul(A, x) + b

# 誤差を定義
with tf.name_scope('loss'):
    loss = tf.reduce_mean(tf.square(y_ - y))

# トレーニングアルゴリズムを定義
with tf.name_scope('train'):
    learning_rate = 1.0e-3
    train = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss)

# TensorBoardで追跡する変数を定義
with tf.name_scope('summary'):
    tf.summary.scalar('A', tf.reshape(tf.matrix_determinant(A), []))
    tf.summary.scalar('b', tf.reshape(tf.norm(b), []))
    tf.summary.scalar('loss', loss)
    merged = tf.summary.merge_all()
    writer = tf.summary.FileWriter('/tmp/tensorboard-sample', sess.graph)

# トレーニングデータを生成
samples_count = 100
x_input = np.ndarray(shape=[samples_count, 3, 1])
y_input = np.ndarray(shape=[samples_count, 3, 1])
for i in range(samples_count):
    x_input[i][0] = np.random.rand()
    x_input[i][1] = np.random.rand()
    x_input[i][2] = np.random.rand()

    for j in range(3):
        y_input[i][j] = x_input[i][j] * 2

# バッチ処理開始
summary = None
total_epochs = 1000
for _ in range(total_epochs):
    for i in range(samples_count):
        __, summary = sess.run([train, merged], feed_dict={x: x_input[i], y: y_input[i]})

    # 進捗10%毎に途中経過をコンソール出力、TensorBoardに監視対象変数の値を記録
    if _ % (total_epochs/10) == 0:
        curr_A, curr_b, curr_loss = sess.run([A, b, loss], feed_dict={x: x_input[_ % samples_count], y: y_input[_ % samples_count]})
        print('-----------------------------')
        print('y')
        print(y_input[_ % samples_count])
        print('A')
        print(curr_A)
        print('x')
        print(x_input[_ % samples_count])
        print('b')
        print(curr_b)
        print('loss')
        print(curr_loss)
        writer.add_summary(summary, _)

# 最終結果の確認
print('==========================')
# x = (1,2,3) を入力すると (2,4,6)に近いベクトルが出力されるはずだが、実際にどうなるか確認。
x_test = np.array([1,2,3]).reshape([3,1])
y_test = np.array([2,4,6]).reshape([3,1])
estimated_y, final_loss = sess.run([y_, loss], feed_dict={x:x_test, y: y_test})
print('Input x')
print(x_test)
print('Computed y')
print(estimated_y)
print('loss')
print(final_loss)