Deep Learning [Day 6]

Published On July 08, 2020

Exercise

문제

keras에서 cifar10 불러오기
이미지데이터를 0과 1사이로 보정
레이블데이터를 one-hot encoding으로 변경

문제 1) CNN으로 구성(138, 139)
1번째 conv2d레이어 3by3 필터, 90개
2번째 conv2d레이어 3by3 필터, 90개
3번재 레이어 softmax의 output레이어

문제 2) MLP로 구성(이미지 데이터를 1차원으로 변형)
multi-layer perception으로 구성
1번째 히든레이어 unit 개수 2000개
2번째 히든레이어 unit 개수 3000개
3번째 히든레이어 unit 개수 2000개
output레이어 unit 개수 10개

힌트) MNIST 경우 mnist.train.next_batch() 등의 함수로 데이터를 가져왔으나 cifar10은 이를 지원하지 않음. 따라서 139페이지처럼 구현

Tensorflow

CNN으로 구성

from keras.utils import np_utils
from keras.datasets import cifar10
import tensorflow as tf
import time

(train_images, train_labels), (test_images, test_labels) = cifar10.load_data()
train_images = train_images.reshape(train_images.shape[0], 32, 32, 3).astype('float32')/255.0
test_images = test_images.reshape(test_images.shape[0], 32, 32, 3).astype('float32')/255.0
train_labels = np_utils.to_categorical(train_labels)
test_labels = np_utils.to_categorical(test_labels)

training_epochs = 15
batch_size = 100

X = tf.placeholder(tf.float32, [None, 32, 32, 3])
Y = tf.placeholder(tf.float32, [None, 10])
X_img = tf.reshape(X, [-1, 32, 32, 3])

# Convolution Layer 1
W1 = tf.Variable(tf.random_normal([3, 3, 3, 90], stddev=0.01))
CL1 = tf.nn.conv2d(X_img, W1, strides=[1, 1, 1, 1], padding='SAME')
CL1 = tf.nn.relu(CL1)

# Convolution Layer 2
W2 = tf.Variable(tf.random_normal([3, 3, 90, 90], stddev=0.01))
CL2 = tf.nn.conv2d(CL1, W2, strides=[1, 1, 1, 1], padding='SAME')
CL2 = tf.nn.relu(CL2)

# Fully Connected Layer
L_flat = tf.reshape(CL2, [-1, 32*32*90])
W3 = tf.Variable(tf.random_normal([32*32*90, 10], stddev=0.01))
b3 = tf.Variable(tf.random_normal([10]))

# Model, Cost, Train
model_LC = tf.matmul(L_flat, W3) + b3
model = tf.nn.softmax(model_LC)
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=model_LC, labels=Y))
train = tf.train.AdamOptimizer(0.01).minimize(cost)

# Accuracy
accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(model, 1), tf.argmax(Y, 1)), tf.float32))

# Session
with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())

  # Training
  t1 = time.time()
  for epoch in range(training_epochs):
    total_batch = int(train_images.shape[0] / batch_size)
    for i in range(total_batch):
      batch_train_images = train_images[i*batch_size : (i+1)*batch_size]
      batch_train_labels = train_labels[i*batch_size : (i+1)*batch_size]
      c, _ = sess.run([cost, train], feed_dict={X: batch_train_images, Y: batch_train_labels})

      if i % 10 == 0:
        print('epoch: ', epoch, ', batch number: ', i)

  t2 = time.time()
  
  # Testing
  print('Training Time (Seconds): ', t2 - t1)

  total_batch = int(test_images.shape[0] / batch_size)
  for i in range(total_batch):
    batch_test_images = test_images[i*batch_size : (i+1)*batch_size]
    batch_test_labels = test_labels[i*batch_size : (i+1)*batch_size]

  print('Accuracy: ', sess.run(accurcay, feed_dict={X: batch_test_images, Y: batch_test_labels}))

ANN으로 구성

from keras.utils import np_utils
import tensorflow as tf
from keras.datasets import cifar10
from keras.models import Sequential
from keras.layers import Dense, Flatten

# cifar10 데이터 가져오기
(train_images, train_labels), (test_images, test_labels) = cifar10.load_data()
train_images = train_images.reshape(train_images.shape[0], 32, 32, 3).astype('float32')/255.0
test_images = test_images.reshape(test_images.shape[0], 32, 32, 3).astype('float32')/255.0
train_labels = np_utils.to_categorical(train_labels)
test_labels = np_utils.to_categorical(test_labels)
print(train_images.shape, train_labels.shape, test_images.shape, test_labels.shape)

training_epochs = 5
batch_size = 100

nH1 = 2000
nH2 = 3000
nH3 = 2000

X = tf.placeholder(tf.float32, [None, 32, 32, 3])
Y = tf.placeholder(tf.float32, [None, 10])
X_img = tf.reshape(X, [-1, 32*32*3])

def mlp_LC(img):
  HL1 = tf.nn.relu(tf.add(tf.matmul(img, W['HL1']), b['HL1']))
  HL2 = tf.nn.relu(tf.add(tf.matmul(HL1, W['HL2']), b['HL2']))
  HL3 = tf.nn.relu(tf.add(tf.matmul(HL2, W['HL3']), b['HL3']))
  Out = tf.matmul(HL3, W['Out']) + b['Out']
  return Out

W = {
    'HL1': tf.Variable(tf.random_normal([32*32*3, nH1])),
    'HL2': tf.Variable(tf.random_normal([nH1, nH2])),
    'HL3': tf.Variable(tf.random_normal([nH2, nH3])),
    'Out': tf.Variable(tf.random_normal([nH3, 10]))
}

b = {
    'HL1': tf.Variable(tf.random_normal([nH1])),
    'HL2': tf.Variable(tf.random_normal([nH2])),
    'HL3': tf.Variable(tf.random_normal([nH3])),
    'Out': tf.Variable(tf.random_normal([10]))
}

# Model, Cost, Train
model_LC = mlp_LC(X_img)
model = tf.nn.softmax(model_LC)
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=model_LC, labels=Y))
train = tf.train.AdamOptimizer(0.01).minimize(cost)

# Accuracy
accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(model, 1), tf.argmax(Y, 1)), tf.float32))

# Session
with tf.Session() as sess:
  sess.run(tf.global_variables_initializer())

  # Training
  for epoch in range(training_epochs):
    total_batch = int(train_images.shape[0] / batch_size)

    # batch_size가 너무 크면 메모리가 잡아먹어, 시간이 오래 걸리는 문제가 있지만,
    # 반대로, batch_size가 너무 작으면 학습이 잘 이루어지지 않아, overfitting 문제가 있을 수 있다.
    for i in range(total_batch):
      batch_train_images = train_images[i*batch_size : (i+1)*batch_size]
      batch_train_labels = train_labels[i*batch_size : (i+1)*batch_size]
      c, _ = sess.run([cost, train], feed_dict={X: batch_train_images, Y: batch_train_labels})

      if i % 100 == 0:
        print('epoch: ', epoch, ', batch number: ', i)

  # Testing
  total_batch = int(test_images.shape[0] / batch_size)
  for i in range(total_batch):
    batch_test_images = test_images[i*batch_size : (i+1)*batch_size]
    batch_test_labels = test_labels[i*batch_size : (i+1)*batch_size]

    print('Accuracy: ', sess.run(accuracy, feed_dict={X: batch_test_images, Y: batch_test_labels}))

Keras