機械学習/MacでTensorFlowを使う のバックアップの現在との差分(No.6)

• バックアップ一覧
• 差分 を表示
• ソース を表示
• バックアップ を表示
• 機械学習/MacでTensorFlowを使う へ行く。
  • 1 (2016-04-11 (月) 11:02:57)
  • 2 (2016-04-11 (月) 11:03:28)
  • 3 (2016-04-14 (木) 13:10:45)
  • 4 (2016-04-15 (金) 18:48:36)
  • 5 (2016-04-16 (土) 09:31:26)
  • 6 (2016-04-16 (土) 11:08:22)
  • 7 (2016-04-19 (火) 00:10:26)
  • 8 (2016-04-19 (火) 00:56:36)
  • 9 (2016-04-19 (火) 18:19:26)
  • 10 (2016-04-19 (火) 18:19:26)
  • 11 (2016-04-19 (火) 18:19:26)
  • 12 (2017-06-19 (月) 19:23:31)
  • 13 (2017-06-20 (火) 08:33:44)
  • 14 (2017-06-20 (火) 08:33:44)
  • 15 (2017-07-24 (月) 14:38:39)

• 追加された行はこの色です。
• 削除された行はこの色です。

#freeze
*はじめに [#ja3312a3]

ここでは、Googleが公開しているオープン・ソース・ソフトウェア (OSS) のTensorFlowをMac (OS X) にインストールして使います。

以前、TensorFlow 0.8用に書いたものを、TensorFlow 1.2用に修正しました。


*環境 [#kd6c49b9]

-OS X Yosemite 10.10.5
-macOS Sierra 10.12.5
-Python 3.5.1
-TensorFlow 0.7
-TensorFlow 1.2

*Python3のインストール（アップグレード） [#zf2975fa]

まず、Python3を最新版にアップグレードします。

*Python3のインストール（アップデート） [#zf2975fa]

まず、Python3を最新版にアップデートします。
下記のサイトから''Mac OS X 64-bit/32-bit installer''をダウンロードして、インストール。
-[[Python 3.5.1:https://www.python.org/downloads/release/python-351/]]
-https://www.python.org/downloads/


*TensorFlowのインストール [#seb185d6]

基本的にはここに書いてある通りなんですが、なぜかsixのeasy_installが先にできなかったので、順序を逆にしました。
-[[Download and Setup:https://www.tensorflow.org/versions/r0.7/get_started/os_setup.html#download-and-setup]] - TensorFlow
基本的にはここに書いてある通りです。
-[[Installing TensorFlow on Mac OS X:https://www.tensorflow.org/install/install_mac]] - TensorFlow

#geshi(sh){{
$ sudo pip3 install --upgrade https://storage.googleapis.com/tensorflow/mac/tensorflow-0.7.1-cp35-none-any.whl
$ sudo easy_install --upgrade six
$ sudo easy_install pip
$ sudo pip3 install --upgrade pip
}}


*Virtualenvのインストール [#ab5cc2c7]

#geshi(sh){{
$ sudo pip3 install --upgrade virtualenv
$ virtualenv --system-site-packages /Users/Shared/tools/tensorflow
$ virtualenv --system-site-packages -p python3 /Users/Shared/tools/tensorflow
$ source /Users/Shared/tools/tensorflow/bin/activate
(tensorflow)$ pip3 install --upgrade https://storage.googleapis.com/tensorflow/mac/tensorflow-0.7.1-cp35-none-any.whl
(tensorflow)$ sudo pip3 install --upgrade tensorflow
(tensorflow)$ deactivate
}}
私の研究室では、共有フォルダーのtoolsというフォルダー (/Users/Shared/tools/) に、みんなで使うツールをインストールしています。
/Users/Shared/tools/tensorflow を ~/tensorflow などTensorFlowをインストールしたいフォルダーに変更してください。



*TensorFlowを動かす [#te2b5012]

TesorFlowのサイトに掲載されているHello Worldを動かします。
-[[Test the TensorFlow installation:https://www.tensorflow.org/versions/r0.7/get_started/os_setup.html#test-the-tensorflow-installation]] - TensorFlow
TesorFlowのサイトに掲載されているサンプルを動かします。
-[[Getting Started With TensorFlow:https://www.tensorflow.org/get_started/get_started]] - TensorFlow

#geshi(sh){{
$ source /Users/Shared/tools/tensorflow/bin/activate
(tensorflow) $ python3
Python 3.5.1 (v3.5.1:37a07cee5969, Dec  5 2015, 21:12:44) 
[GCC 4.2.1 (Apple Inc. build 5666) (dot 3)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
}}
#geshi(python){{
>>> import tensorflow as tf
>>> hello = tf.constant('Hello, TensorFlow!')
>>> node1 = tf.constant(3.0, dtype=tf.float32)
>>> node2 = tf.constant(4.0) 
>>> sess = tf.Session()
>>> print(sess.run(hello))
b'Hello, TensorFlow!'
>>> a = tf.constant(10)
>>> b = tf.constant(32)
>>> print(sess.run(a + b))
42
2017-06-20 08:28:06.121472: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2017-06-20 08:28:06.121492: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2017-06-20 08:28:06.121498: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
2017-06-20 08:28:06.121502: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
>>> print(sess.run([node1, node2]))
[3.0, 4.0]
>>> node3 = tf.add(node1, node2)
>>> print(sess.run(node3))
7.0
>>> quit()
(tensorflow) $ deactivate
}}
tf.Seesion()の後に警告が出ているのは、CPUのオプション使うともっと速く計算できるようになるよっていうメッセージです。


*irisで線形回帰 [#ka2954e2]

まずは、おなじみのirisデータセットで線形回帰をやってみましょう。
irisデータセットは、4つの説明変数（花びらの長さ、幅、がく片の長さ、幅）とカテゴリー (setosa, versicolor, virginica) からなる分類問題用のデータセットですが、4つの説明変数を3つの説明変数と1つの目標変数にして、回帰問題用のデータセットにします。
*TensorFlowでカテゴリー分析をやってみる [#ka2954e2]

UCI Machine Learning Repositoryからファイルをダウンロードします。
TensorFlowの最初のチュートリアルは手書き数字認識のMNISTですが、irisデータセットでカテゴリー分析（分類問題）をやってみます。


**データの準備 [#p76609a7]
[[Irisデータセット:http://archive.ics.uci.edu/ml/datasets/Iris]]は、4つの説明変数（花びらの長さ、幅、がく片の長さ、幅）とカテゴリー (setosa, versicolor, virginica) からなるデータセットです。

UCI Machine Learning Repositoryからファイル [[iris.data:http://archive.ics.uci.edu/ml/machine-learning-databases/iris/iris.data]] をダウンロードします。
中身はこんな感じです。
#geshi(txt){{
$ head -5 iris.data
5.1,3.5,1.4,0.2,Iris-setosa
4.9,3.0,1.4,0.2,Iris-setosa
4.7,3.2,1.3,0.2,Iris-setosa
4.6,3.1,1.5,0.2,Iris-setosa
5.0,3.6,1.4,0.2,Iris-setosa
}}

ソースコードはこんな感じです。
3クラスの分類なので、カテゴリーのラベルを3次元の0/1ベクトルに変換します。
つまり、setosaは 1,0,0、versicolorは 0,1,0、virginicaは 0,0,1 とします。

以下のページを参考にしました。
-[[TensorFlowで回帰をやってみる:http://qiita.com/syoamakase/items/db883d7ebad7a2220233]] - Qiita
Irisデータセットには、それぞれ、50個ずつ150個のデータが並んでいるので、次のように変換します。
#geshi(sh){{
$ head -50 iris.data | awk -F, '{printf("%s,%s,%s,%s,1,0,0\n",$1,$2,$3,$4);}' > iris.csv
$ head -100 iris.data | tail -50 | awk -F, '{printf("%s,%s,%s,%s,0,1,0\n",$1,$2,$3,$4);}' >> iris.csv
$ head -150 iris.data | tail -50 | awk -F, '{printf("%s,%s,%s,%s,0,0,1\n",$1,$2,$3,$4);}' >> iris.csv
}}

これで、データの準備は完了です。
このデータをdataフォルダーにおいておきます。

**カテゴリー分析用ソースコード [#xeaf4449]

TensorFlowのTutorialsについてる最初の手書き文字データセットMNIST用のサンプルをほとんどそのまま使ったソースコードはこんな感じです。
#ref(./tfc_iris.py)
#geshi(python){{
# Copyright 2016-2017 Tohgoroh Matsui All Rights Reserved.
#
# Changes:
# 2017-06-19 Changed for TensorFlow 1.2
#
#
# This includes software developed by Google, Inc.
# licensed under the Apache License, Version 2.0 (the 'License');
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an 'AS IS' BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import tensorflow as tf
import numpy as np

# 損失関数
def loss(o, y):
  with tf.name_scope('loss') as scope:
    loss = tf.reduce_mean(tf.square(o - y))
  return loss

# パラメーター
flags = tf.app.flags
FLAGS = flags.FLAGS
flags.DEFINE_integer('samples',       120,    'Number of training samples.')
flags.DEFINE_integer('max_steps',     1001,   'Number of steps to run trainer.')
flags.DEFINE_float(  'learning_rate', 0.001,  'Initial learning rate.')
flags.DEFINE_float(  'dropout',       0.9,    'Keep probability for training dropout.')
flags.DEFINE_string( 'data_dir',      'data', 'Directory for storing data.')
flags.DEFINE_string( 'summaries_dir', 'log',  'Summaries directory.')

# CSVファイルの読み込み
iris = np.genfromtxt('iris.data', delimiter=",")

# 学習データ (train) とテストデータ (test) に分割
N = 100
x_train, x_test = np.vsplit(iris[:,0:3].astype(np.float32), [N])
y_train, y_test = np.vsplit(iris[:,3:4].astype(np.float32), [N])
def train():
  # CSVファイルの読み込み
  data = np.genfromtxt(FLAGS.data_dir + '/iris.csv', delimiter=",")

  # データをシャッフル
  np.random.shuffle(data)
  
  # 訓練データ (train) とテストデータ (test) に分割
  x_train, x_test = np.vsplit(data[:,0:4].astype(np.float32), [FLAGS.samples])
  y_train, y_test = np.vsplit(data[:,4:7].astype(np.float32), [FLAGS.samples])

# 入力層
n0 = 3
x = tf.placeholder("float", shape=[None, n0])

# 中間層1（活性化関数はReLU）
n1 = 64
with tf.name_scope('l1') as scope:
  w1 = tf.Variable(tf.truncated_normal([n0, n1], stddev=0.1), name="w1")
  b1 = tf.Variable(tf.constant(1.0, shape=[n1]), name="b1")
  o1 = tf.nn.relu(tf.matmul(x, w1) + b1)
  # セッション
  sess = tf.InteractiveSession()

# 中間層2（活性化関数はReLU）
n2 = 64
with tf.name_scope('l2') as scope:
  w2 = tf.Variable(tf.truncated_normal([n1, n2], stddev=0.1), name="w2")
  b2 = tf.Variable(tf.constant(1.0, shape=[n2]), name="b2")
  o2 = tf.nn.relu(tf.matmul(o1, w2) + b2)
  # 入力
  with tf.name_scope('input'):
    x  = tf.placeholder(tf.float32, shape=[None, 4])
  
  # 出力
  with tf.name_scope('output'):
    y_ = tf.placeholder(tf.float32, shape=[None, 3])
    
  # ドロップアウトのキープ率
  with tf.name_scope('dropout'):
    keep_prob = tf.placeholder(tf.float32)
    tf.summary.scalar('dropout_keep_probability', keep_prob)
  
  # 重み（係数）
  def weight_variable(shape):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial)
  
  # バイアス（定数項）
  def bias_variable(shape):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial)
    
  # for TensorBoard
  def variable_summaries(var, name):
    with tf.name_scope('summaries'):
      mean = tf.reduce_mean(var)
      tf.summary.scalar('mean/' + name, mean)
      with tf.name_scope('stddev'):
        stddev = tf.sqrt(tf.reduce_sum(tf.square(var - mean)))
      tf.summary.scalar('sttdev/' + name, stddev)
      tf.summary.scalar('max/' + name, tf.reduce_max(var))
      tf.summary.scalar('min/' + name, tf.reduce_min(var))
      tf.summary.histogram(name, var)

# 出力層（線形結合）
n3 = 1
with tf.name_scope('l3') as scope:
  w3 = tf.Variable(tf.truncated_normal([n2, n3], stddev=0.1), name="w3")
  b3 = tf.Variable(tf.constant(1.0, shape=[n3]), name="b3")
  o3 = tf.matmul(o2, w3) + b3

  # ニューラル・ネットワークの層
  def nn_layer(input_tensor, input_dim, output_dim, layer_name, act=tf.nn.relu):
    with tf.name_scope(layer_name):
      with tf.name_scope('weights'):
        weights = weight_variable([input_dim, output_dim])
        variable_summaries(weights, layer_name + '/weights')
      with tf.name_scope('biases'):
        biases = bias_variable([output_dim])
        variable_summaries(biases, layer_name + '/biases')
      with tf.name_scope('Wx_plus_b'):
        preactivate = tf.matmul(input_tensor, weights) + biases
        tf.summary.histogram(layer_name + '/pre_activations', preactivate)
      activations = act(preactivate, name='activation')
      tf.summary.histogram(layer_name + '/acctivations', activations)
      return activations
      

loss_op  = loss(o3, y_train)
train_op = tf.train.AdagradOptimizer(0.01).minimize(loss_op)
min = float("inf")
  # 入力層
  hidden1 = nn_layer(x, 4, 64, 'layer1')
  dropped1 = tf.nn.dropout(hidden1, keep_prob)
  
  # 中間層
  hidden2 = nn_layer(dropped1, 64, 64, 'layer2')
  dropped2 = tf.nn.dropout(hidden2, keep_prob)
  
  # 出力層
  y = nn_layer(dropped2, 64, 3, 'layer3', act=tf.nn.softmax)


# 初期化
init_op = tf.initialize_all_variables()
  # クロス・エントロピー
  with tf.name_scope('cross_entropy'):
    #diff = y_ * tf.log(y)
    #with tf.name_scope('total'):
    #  cross_entropy = -tf.reduce_mean(diff)
    cross_entropy = -tf.reduce_sum(y_*tf.log(tf.clip_by_value(y,1e-10,1.0)))
    tf.summary.scalar('cross entropy', cross_entropy)

  # 学習
  with tf.name_scope('train'):
    train_step = tf.train.AdamOptimizer(FLAGS.learning_rate).minimize(cross_entropy)

# セッション
with tf.Session() as sess:
  sess.run(init_op)
  for i in range(100001):
    loss = sess.run(loss_op, feed_dict={x:x_train})
    sess.run(train_op, feed_dict={x:x_train})
    if loss < min:
      min  = loss
      best = sess.run(o3, feed_dict={x:x_test})
  # 精度
  with tf.name_scope('accuracy'):
    with tf.name_scope('correct_prediction'):
      correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
    with tf.name_scope('accuracy'):
      accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
    tf.summary.scalar('accuracy', accuracy)


  # for TensorBoard
  merged = tf.summary.merge_all()
  train_writer = tf.summary.FileWriter(FLAGS.summaries_dir + '/train', sess.graph)
  test_writer  = tf.summary.FileWriter(FLAGS.summaries_dir + '/test')
  tf.global_variables_initializer().run()


  # データとドロップアウト・キープ率の切り替え
  def feed_dict(train):
    if train:
      xs, ys = x_train, y_train
      k  = FLAGS.dropout
    else:
      xs, ys = x_test, y_test
      k  = 1.0
    return {x: xs, y_: ys, keep_prob: k}


  # 学習ルーチン
  for i in range(FLAGS.max_steps):
    # 学習
    sess.run(train_step, feed_dict=feed_dict(True))
    # 訓練データに対する評価
    summary, acc, cp = sess.run([merged, accuracy, cross_entropy], feed_dict=feed_dict(True))
    train_writer.add_summary(summary, i)
    # テスト・データに対する評価
    summary, acc, cp = sess.run([merged, accuracy, cross_entropy], feed_dict=feed_dict(False))
    test_writer.add_summary(summary, i)
    if i == 0 or i % np.power(10, np.floor(np.log10(i))) == 0:
      r = np.corrcoef(best.flatten(), y_test.flatten())[0][1]
      print('{}: Loss = {}, R = {}'.format(i, min, r))
      print('Accuracy and Cross-Entropy at step %s: %s, %s' % (i, acc, cp))


def main(_):
  if tf.gfile.Exists(FLAGS.summaries_dir):
    tf.gfile.DeleteRecursively(FLAGS.summaries_dir)
  tf.gfile.MakeDirs(FLAGS.summaries_dir)
  train()

if __name__ == '__main__':
  tf.app.run()
}}

実行すると、次のようになります。
mnist_with_summaries.pyからいくつか変更点があります。
-iris.csvからデータを読み込む
-データをシャッフルして120個を訓練データに、残りの30個をテスト・データにする
-入力は4次元、出力は3次元
-中間層（活性化関数はReLU）を追加
-入力層は4x64ユニット、中間層は64x64ユニット、出力層は64x3ユニット
-クロス・エントロピーの定義を変更（元の定義だとNaNになってしまうことがあるため）
-10回中9回学習して1回テストを行うのではなく、毎回学習とテストを行う
-標準出力に精度とクロス・エントロピーを出力


**実行 [#f0d2b637]
#geshi(sh){{
(tensorflow) $ python3 tfRegression.py 
0: Loss = 0.9924622178077698, R = -0.31376540185534374
1: Loss = 0.33512526750564575, R = -0.2809933838930298
2: Loss = 0.3103932738304138, R = -0.2605662262997624
3: Loss = 0.29998070001602173, R = -0.19988817587575336
4: Loss = 0.29054442048072815, R = -0.10261393047830965
5: Loss = 0.2815440893173218, R = -0.0030378746322473078
6: Loss = 0.2728765904903412, R = 0.06743248160241787
7: Loss = 0.26452258229255676, R = 0.11218694086822772
8: Loss = 0.25640469789505005, R = 0.14083619953594895
9: Loss = 0.24850055575370789, R = 0.15955179537740188
10: Loss = 0.2407849133014679, R = 0.17325851369183962
20: Loss = 0.17123618721961975, R = 0.2119044088136728
30: Loss = 0.11361732333898544, R = 0.21242390033328104
40: Loss = 0.07034626603126526, R = 0.21351552138337554
50: Loss = 0.04281529411673546, R = 0.2112161148516869
60: Loss = 0.027972474694252014, R = 0.21323103613228045
70: Loss = 0.02100631780922413, R = 0.2133350814777594
80: Loss = 0.01804741658270359, R = 0.21420981216605972
90: Loss = 0.016845906153321266, R = 0.2153937757021539
100: Loss = 0.01631508581340313, R = 0.21719647880323853
200: Loss = 0.01521082129329443, R = 0.2440402497636383
300: Loss = 0.014608824625611305, R = 0.2645834921419263
400: Loss = 0.014166736975312233, R = 0.28157404933081204
500: Loss = 0.013829934410750866, R = 0.2952680005290892
600: Loss = 0.013562848791480064, R = 0.30675864601619857
700: Loss = 0.013374903239309788, R = 0.3164113669058739
800: Loss = 0.013220920227468014, R = 0.3264013737506945
900: Loss = 0.013091418892145157, R = 0.3328174293176383
1000: Loss = 0.012949838303029537, R = 0.33802548623013406
2000: Loss = 0.012516467832028866, R = 0.3696120433568608
3000: Loss = 0.012403683736920357, R = 0.3809822832215991
4000: Loss = 0.012310308404266834, R = 0.38904775116540224
5000: Loss = 0.012245077639818192, R = 0.39475342114404327
6000: Loss = 0.0121763302013278, R = 0.39984150610626895
7000: Loss = 0.012104801833629608, R = 0.40494473659532704
8000: Loss = 0.012028640136122704, R = 0.4100277288577437
9000: Loss = 0.011952037923038006, R = 0.41544311557772695
10000: Loss = 0.011868388392031193, R = 0.42219982940390616
20000: Loss = 0.011160679161548615, R = 0.46866687115166755
30000: Loss = 0.0105402497574687, R = 0.4942357990758039
40000: Loss = 0.010395915247499943, R = 0.5026822424224162
50000: Loss = 0.01029052771627903, R = 0.49972020218055674
60000: Loss = 0.010193731635808945, R = 0.48942378308652223
70000: Loss = 0.010111675597727299, R = 0.4797740509869646
80000: Loss = 0.009985744953155518, R = 0.4579861414250584
90000: Loss = 0.009862487204372883, R = 0.43707398862896835
100000: Loss = 0.009671863168478012, R = 0.4074325332085266
(tensorflow) $ python3 tfc_iris.py
2017-06-19 18:32:23.449363: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2017-06-19 18:32:23.449387: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2017-06-19 18:32:23.449393: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
2017-06-19 18:32:23.449398: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
Accuracy and Cross-Entropy at step 0: 0.366667, 31.8714
Accuracy and Cross-Entropy at step 1: 0.3, 31.5792
Accuracy and Cross-Entropy at step 2: 0.3, 31.289
Accuracy and Cross-Entropy at step 3: 0.3, 31.0007
Accuracy and Cross-Entropy at step 4: 0.3, 30.7225
Accuracy and Cross-Entropy at step 5: 0.333333, 30.4467
Accuracy and Cross-Entropy at step 6: 0.4, 30.1522
Accuracy and Cross-Entropy at step 7: 0.5, 29.8428
Accuracy and Cross-Entropy at step 8: 0.566667, 29.5071
Accuracy and Cross-Entropy at step 9: 0.633333, 29.1437
Accuracy and Cross-Entropy at step 10: 0.666667, 28.7539
Accuracy and Cross-Entropy at step 20: 0.666667, 24.4503
Accuracy and Cross-Entropy at step 30: 0.666667, 19.725
Accuracy and Cross-Entropy at step 40: 0.666667, 15.7528
Accuracy and Cross-Entropy at step 50: 0.866667, 13.1522
Accuracy and Cross-Entropy at step 60: 0.9, 11.3919
Accuracy and Cross-Entropy at step 70: 0.933333, 9.89667
Accuracy and Cross-Entropy at step 80: 0.933333, 8.50913
Accuracy and Cross-Entropy at step 90: 0.933333, 7.18226
Accuracy and Cross-Entropy at step 100: 0.933333, 6.18636
Accuracy and Cross-Entropy at step 200: 0.933333, 4.82435
Accuracy and Cross-Entropy at step 300: 0.933333, 4.7089
Accuracy and Cross-Entropy at step 400: 0.933333, 5.20262
Accuracy and Cross-Entropy at step 500: 0.933333, 6.28935
Accuracy and Cross-Entropy at step 600: 0.933333, 6.46973
Accuracy and Cross-Entropy at step 700: 0.933333, 6.57728
Accuracy and Cross-Entropy at step 800: 0.933333, 7.21051
Accuracy and Cross-Entropy at step 900: 0.933333, 7.32343
Accuracy and Cross-Entropy at step 1000: 0.933333, 7.07634
}}
最初に出ている警告は、CPUのオプションを使って高速に計算できることを教えてくれています。
このオプションを使うには、TensorFlowをソースコードからビルドしてインストールする必要があります。


**TensorBoardによるログの確認 [#nd118cd7]
TensorBoardでログを確認します。

次のようにしてTensorBoardを起動すると、6006番ポートでHTTPサーバーが起動します。
#geshi(sh){{
(tensorflow) $ tensorboard --logdir=log
}}

Webブラウザーで http://localhost:6006 にアクセスします。
#ref(./tfc_iris.png,nolink,50%)



*TensorFlowで回帰分析をやってみる [#d6ec9fd8]

次に、Housingデータセットで回帰分析をやってみます。


**データの準備 [#l2dbe849]
[[Housingデータセット:http://archive.ics.uci.edu/ml/datasets/Housing]]は、ボストン近郊の住宅の価格を、地区の犯罪発生率、宅地率など13個の説明変数を用いて予測する問題です。

UCI Machine Learning Repositoryから、 [[housing.data:http://archive.ics.uci.edu/ml/machine-learning-databases/housing/housing.data]] をダウンロードします。
中身はこんな感じです。
#geshi(sh){{
$ head -5 housing.data
 0.00632  18.00   2.310  0  0.5380  6.5750  65.20  4.0900   1  296.0  15.30 396.90   4.98  24.00
 0.02731   0.00   7.070  0  0.4690  6.4210  78.90  4.9671   2  242.0  17.80 396.90   9.14  21.60
 0.02729   0.00   7.070  0  0.4690  7.1850  61.10  4.9671   2  242.0  17.80 392.83   4.03  34.70
 0.03237   0.00   2.180  0  0.4580  6.9980  45.80  6.0622   3  222.0  18.70 394.63   2.94  33.40
 0.06905   0.00   2.180  0  0.4580  7.1470  54.20  6.0622   3  222.0  18.70 396.90   5.33  36.20
}}

これをコンマ区切りのCSVファイルに変換します。
#geshi(sh){{
$ sed "s/^ //" < housing.data | sed "s/  */,/g" > housing.csv
}}

これで、データの準備は完了です。 このデータをdataフォルダーに置いておきます。


**回帰分析用ソースコード [#a6c1ec07]

カテゴリー分析用のソースコードを一部変更して作成した回帰分析用のソースコードはこんな感じです。
#ref(./tfr_housing.py)
#geshi(python){{
# Copyright 2016-2017 Tohgoroh Matsui All Rights Reserved.
#
# Changes:
# 2017-06-19 Changed for TensorFlow 1.2
#
#
# This includes software developed by Google, Inc.
# licensed under the Apache License, Version 2.0 (the 'License');
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an 'AS IS' BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import tensorflow as tf
import numpy as np


# パラメーター
flags = tf.app.flags
FLAGS = flags.FLAGS
flags.DEFINE_integer('samples',       400,    'Number of training samples.')
flags.DEFINE_integer('max_steps',     1001,   'Number of steps to run trainer.')
flags.DEFINE_float(  'learning_rate', 0.001,  'Initial learning rate.')
flags.DEFINE_float(  'dropout',       0.9,    'Keep probability for training dropout.')
flags.DEFINE_string( 'data_dir',      'data', 'Directory for storing data.')
flags.DEFINE_string( 'summaries_dir', 'log',  'Summaries directory.')


def train():
  # CSVファイルの読み込み
  data = np.genfromtxt(FLAGS.data_dir + '/housing.csv', delimiter=",")

  # データをシャッフル
  np.random.shuffle(data)
  
  # 学習データ (train) とテストデータ (test) に分割
  x_train, x_test = np.vsplit(data[:, 0:13].astype(np.float32), [FLAGS.samples])
  y_train, y_test = np.vsplit(data[:,13:14].astype(np.float32), [FLAGS.samples])


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

  # 入力
  with tf.name_scope('input'):
    x  = tf.placeholder(tf.float32, shape=[None, 13])
  
  # 出力
  with tf.name_scope('output'):
    y_ = tf.placeholder(tf.float32, shape=[None, 1])
    
  # ドロップアウトのキープ率
  with tf.name_scope('dropout'):
    keep_prob = tf.placeholder("float")
    tf.summary.scalar('dropout_keep_probability', keep_prob)
  
  # 重み（係数）
  def weight_variable(shape):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial)
  
  # バイアス（定数項）
  def bias_variable(shape):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial)
    
  # for TensorBoard
  def variable_summaries(var, name):
    with tf.name_scope('summaries'):
      mean = tf.reduce_mean(var)
      tf.summary.scalar('mean/' + name, mean)
      with tf.name_scope('stddev'):
        stddev = tf.sqrt(tf.reduce_sum(tf.square(var - mean)))
      tf.summary.scalar('sttdev/' + name, stddev)
      tf.summary.scalar('max/' + name, tf.reduce_max(var))
      tf.summary.scalar('min/' + name, tf.reduce_min(var))
      tf.summary.histogram(name, var)


  # ニューラル・ネットワークの層
  def nn_layer(input_tensor, input_dim, output_dim, layer_name, relu=True):
    with tf.name_scope(layer_name):
      with tf.name_scope('weights'):
        weights = weight_variable([input_dim, output_dim])
        variable_summaries(weights, layer_name + '/weights')
      with tf.name_scope('biases'):
        biases = bias_variable([output_dim])
        variable_summaries(biases, layer_name + '/biases')
      with tf.name_scope('Wx_plus_b'):
        preactivate = tf.matmul(input_tensor, weights) + biases
        tf.summary.histogram(layer_name + '/pre_activations', preactivate)
      if relu:
        activations = tf.nn.relu(preactivate, name='activation')
      else:
        activations = preactivate
      tf.summary.histogram(layer_name + '/acctivations', activations)
      return activations
      

  # 入力層
  hidden1 = nn_layer(x, 13, 64, 'layer1')
  dropped1 = tf.nn.dropout(hidden1, keep_prob)
  
  # 中間層
  hidden2 = nn_layer(dropped1, 64, 64, 'layer2')
  dropped2 = tf.nn.dropout(hidden2, keep_prob)
  
  # 出力層
  y = nn_layer(dropped2, 64, 1, 'layer3', relu=False)


  # 損失
  with tf.name_scope('loss'):
    loss = tf.reduce_mean(tf.square(y_ - y))
    tf.summary.scalar('loss', loss)

  # 学習
  with tf.name_scope('train'):
    train_step = tf.train.AdagradOptimizer(FLAGS.learning_rate).minimize(loss)


  # for TensorBoard
  merged = tf.summary.merge_all()
  train_writer = tf.summary.FileWriter(FLAGS.summaries_dir + '/train', sess.graph)
  test_writer  = tf.summary.FileWriter(FLAGS.summaries_dir + '/test')
  tf.global_variables_initializer().run()


  # データとドロップアウト・キープ率の切り替え
  def feed_dict(train):
    if train:
      xs, ys = x_train, y_train
      k  = FLAGS.dropout
    else:
      xs, ys = x_test, y_test
      k  = 1.0
    return {x: xs, y_: ys, keep_prob: k}


  # 学習ルーチン
  for i in range(FLAGS.max_steps):
    # 学習
    sess.run(train_step, feed_dict=feed_dict(True))
    # 訓練データに対する評価
    summary, train_loss = sess.run([merged, loss], feed_dict=feed_dict(True))
    train_writer.add_summary(summary, i)
    # テスト・データに対する評価
    summary, test_loss, est = sess.run([merged, loss, y], feed_dict=feed_dict(False))
    test_writer.add_summary(summary, i)
    if i == 0 or i % np.power(10, np.floor(np.log10(i))) == 0:
      print('Train loss and test loss at step %s: %s, %s' % (i, train_loss, test_loss))


def main(_):
  if tf.gfile.Exists(FLAGS.summaries_dir):
    tf.gfile.DeleteRecursively(FLAGS.summaries_dir)
  tf.gfile.MakeDirs(FLAGS.summaries_dir)
  train()

if __name__ == '__main__':
  tf.app.run()
}}

カテゴリー分析用のソースコードからの変更点は以下のようなものです。
-housing.csvからデータを読み込む
-訓練データを400個、残りをテスト・データに
-入力ユニット13個、出力ユニット1個
-出力層の活性化関数はなし（線形和をそのまま出力）
-目的関数は平均二乗誤差
-最適化アルゴリズムはAdagradOptimizer


**実行 [#sf335df5]
#geshi(sh){{
(tensorflow) $ python3 tfr_housing.py
2017-06-19 19:18:56.446456: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2017-06-19 19:18:56.446483: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2017-06-19 19:18:56.446489: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
2017-06-19 19:18:56.446494: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
Train loss and test loss at step 0: 224.93, 130.847
Train loss and test loss at step 1: 183.852, 85.3385
Train loss and test loss at step 2: 168.604, 68.0674
Train loss and test loss at step 3: 148.723, 60.0659
Train loss and test loss at step 4: 133.088, 56.0457
Train loss and test loss at step 5: 136.931, 53.1182
Train loss and test loss at step 6: 130.554, 51.7777
Train loss and test loss at step 7: 133.33, 50.4604
Train loss and test loss at step 8: 124.398, 49.5372
Train loss and test loss at step 9: 153.345, 48.7853
Train loss and test loss at step 10: 126.658, 48.1853
Train loss and test loss at step 20: 131.693, 44.5592
Train loss and test loss at step 30: 115.671, 43.4062
Train loss and test loss at step 40: 112.891, 43.2623
Train loss and test loss at step 50: 113.978, 43.1277
Train loss and test loss at step 60: 114.35, 43.1381
Train loss and test loss at step 70: 96.5845, 43.1183
Train loss and test loss at step 80: 102.218, 43.1819
Train loss and test loss at step 90: 105.793, 43.2823
Train loss and test loss at step 100: 105.103, 43.2404
Train loss and test loss at step 200: 96.8645, 42.1962
Train loss and test loss at step 300: 81.3204, 40.8889
Train loss and test loss at step 400: 75.9416, 40.0916
Train loss and test loss at step 500: 83.1751, 39.7313
Train loss and test loss at step 600: 77.3653, 39.0446
Train loss and test loss at step 700: 71.6672, 38.6893
Train loss and test loss at step 800: 74.5351, 38.3223
Train loss and test loss at step 900: 76.4663, 37.7926
Train loss and test loss at step 1000: 66.521, 37.3258
}}

**TensorBoardによるログの確認 [#hd935499]

カテゴリー分析と同じです。
#geshi(sh){{
(tensorflow) $ tensorboard --logdir=log
}}
#ref(./tfr_housing.png,nolink,50%)