[tensorflow] CNN model構築

tensorflowでMNISTデータセットを使った数字認識のためのCNNモデル構築

参考 : https://github.com/tensorflow/models/tree/master/official/mnist

python

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data

class Model:

  def __init__(self, data_format):
    if data_format == 'channels_first':
      # NCHW : [入力数, チャンネル数, 高さ, 幅]
      # GPU向け、sess.run()などでsliceしたデータが見れない
      # lenaだったら[-1, 3, 512, 512], -1は入力数を意味
      self._input_shape = [-1, 1, 28, 28]
    else:
      # NHWC : [入力数, 高さ, 幅, チャンネル数]
      # CPU向け、sess.run()などでsliceしたデータが見れる
      # lenaだったら[-1, 512, 512, 3], -1は入力数を意味
      assert data_format == 'channels_last'
      self._input_shape = [-1, 28, 28, 1]

    # Neural Networkの構成要素の定義
    # hidden layer 1 : Conv2D (output : 32次元, conv窓大きさ : 5ピクセル)
    # conv mode : same (zero padding), validだとpaddingしない
    # relu : activation関数で負の入力を0にする
    self.conv1 = tf.layers.Conv2D(
        32, 5, padding='same', data_format=data_format, activation=tf.nn.relu)
    self.conv2 = tf.layers.Conv2D(
        64, 5, padding='same', data_format=data_format, activation=tf.nn.relu)
    # output 1024
    self.fc1 = tf.layers.Dense(1024, activation=tf.nn.relu)
    # output 10 (labels)
    self.fc2 = tf.layers.Dense(10)
    # 0.4の確率で入力を0にする
    self.dropout = tf.layers.Dropout(0.4)
    # (2, 2)ブロックでscanしてmaxを取る
    self.max_pool2d = tf.layers.MaxPooling2D(
        (2, 2), (2, 2), padding='same', data_format=data_format)

  def __call__(self, inputs, training):
    # inputs : (5000, 784)
    y = tf.reshape(inputs, self._input_shape)
    # (5000, 28, 28, 1)
    y = self.conv1(y)
    # (5000, 28, 28, 32)
    # if mode -> 'valid' : (5000, 24, 24, 32)
    y = self.max_pool2d(y)
    # (5000, 14, 14, 32)
    y = self.conv2(y)
    # (5000, 14, 14, 64)
    y = self.max_pool2d(y)
    # (5000, 7, 7, 64)
    y = tf.layers.flatten(y)
    # (5000, 3136)
    y = self.fc1(y)
    # (5000, 1024)
    # testingだとdropoutしない
    y = self.dropout(y, training=training)
    # (5000, 1024) : 0の数が0.4倍増える
    return self.fc2(y)
    # (5000, 10)

mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)

# (5000, 784)
val_images = mnist.validation.images
val_labels = mnist.validation.labels

model = Model('channels_last')
logits = model(val_images, training=True)

途中でデータを見たいとき

python

with tf.Session() as sess:
   sess.run(tf.global_variables_initializer())
   tmp = sess.run(y)
   print(tmp)
   cv2.imshow('', tmp)
   cv2.waitKey(0)