티스토리 뷰
지난번에 Deep CNN을 이용하여 99.3%의 정확도가 나왔다.
그런데 소스코드가 너무 복잡하고 관리하기가 힘들기 때문에 파이썬의 class를 이용하면 효과적으로 관리할 수 있다.
In [2]:
import tensorflow as tf
import numpy as np
# import matplotlib.pyplot as plt
from tensorflow.examples.tutorials.mnist import input_data
tf.set_random_seed(777) # reproducibility
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
# Check out https://www.tensorflow.org/get_started/mnist/beginners for
# more information about the mnist dataset
# hyper parameters
learning_rate = 0.001
training_epochs = 15
batch_size = 100
Python Class¶
In [2]:
class Model:
# initialize에서 session을 넘겨주면 좋다.
def __init__(self, sess, name):
self.sess = sess
self.name = name
self._build_net()
# 네트워크를 build 한다.
def _build_net(self):
with tf.variable_scope(self.name):
# dropout (keep_prob) rate 0.7~0.5 on training, but should be 1
# for testing
self.keep_prob = tf.placeholder(tf.float32)
# input place holders
self.X = tf.placeholder(tf.float32, [None, 784])
# img 28x28x1 (black/white)
X_img = tf.reshape(self.X, [-1, 28, 28, 1])
self.Y = tf.placeholder(tf.float32, [None, 10])
# L1 ImgIn shape=(?, 28, 28, 1)
W1 = tf.Variable(tf.random_normal([3, 3, 1, 32], stddev=0.01))
# Conv -> (?, 28, 28, 32)
# Pool -> (?, 14, 14, 32)
L1 = tf.nn.conv2d(X_img, W1, strides=[1, 1, 1, 1], padding='SAME')
L1 = tf.nn.relu(L1)
L1 = tf.nn.max_pool(L1, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME')
L1 = tf.nn.dropout(L1, keep_prob=self.keep_prob)
'''
Tensor("Conv2D:0", shape=(?, 28, 28, 32), dtype=float32)
Tensor("Relu:0", shape=(?, 28, 28, 32), dtype=float32)
Tensor("MaxPool:0", shape=(?, 14, 14, 32), dtype=float32)
Tensor("dropout/mul:0", shape=(?, 14, 14, 32), dtype=float32)
'''
# L2 ImgIn shape=(?, 14, 14, 32)
W2 = tf.Variable(tf.random_normal([3, 3, 32, 64], stddev=0.01))
# Conv ->(?, 14, 14, 64)
# Pool ->(?, 7, 7, 64)
L2 = tf.nn.conv2d(L1, W2, strides=[1, 1, 1, 1], padding='SAME')
L2 = tf.nn.relu(L2)
L2 = tf.nn.max_pool(L2, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME')
L2 = tf.nn.dropout(L2, keep_prob=self.keep_prob)
'''
Tensor("Conv2D_1:0", shape=(?, 14, 14, 64), dtype=float32)
Tensor("Relu_1:0", shape=(?, 14, 14, 64), dtype=float32)
Tensor("MaxPool_1:0", shape=(?, 7, 7, 64), dtype=float32)
Tensor("dropout_1/mul:0", shape=(?, 7, 7, 64), dtype=float32)
'''
# L3 ImgIn shape=(?, 7, 7, 64)
W3 = tf.Variable(tf.random_normal([3, 3, 64, 128], stddev=0.01))
# Conv ->(?, 7, 7, 128)
# Pool ->(?, 4, 4, 128)
# Reshape ->(?, 4 * 4 * 128) # Flatten them for FC
L3 = tf.nn.conv2d(L2, W3, strides=[1, 1, 1, 1], padding='SAME')
L3 = tf.nn.relu(L3)
L3 = tf.nn.max_pool(L3, ksize=[1, 2, 2, 1], strides=[
1, 2, 2, 1], padding='SAME')
L3 = tf.nn.dropout(L3, keep_prob=self.keep_prob)
L3_flat = tf.reshape(L3, [-1, 128 * 4 * 4])
'''
Tensor("Conv2D_2:0", shape=(?, 7, 7, 128), dtype=float32)
Tensor("Relu_2:0", shape=(?, 7, 7, 128), dtype=float32)
Tensor("MaxPool_2:0", shape=(?, 4, 4, 128), dtype=float32)
Tensor("dropout_2/mul:0", shape=(?, 4, 4, 128), dtype=float32)
Tensor("Reshape_1:0", shape=(?, 2048), dtype=float32)
'''
# L4 FC 4x4x128 inputs -> 625 outputs
W4 = tf.get_variable("W4", shape=[128 * 4 * 4, 625],
initializer=tf.contrib.layers.xavier_initializer())
b4 = tf.Variable(tf.random_normal([625]))
L4 = tf.nn.relu(tf.matmul(L3_flat, W4) + b4)
L4 = tf.nn.dropout(L4, keep_prob=self.keep_prob)
'''
Tensor("Relu_3:0", shape=(?, 625), dtype=float32)
Tensor("dropout_3/mul:0", shape=(?, 625), dtype=float32)
'''
# L5 Final FC 625 inputs -> 10 outputs
W5 = tf.get_variable("W5", shape=[625, 10],
initializer=tf.contrib.layers.xavier_initializer())
b5 = tf.Variable(tf.random_normal([10]))
self.logits = tf.matmul(L4, W5) + b5
'''
Tensor("add_1:0", shape=(?, 10), dtype=float32)
'''
# define cost/loss & optimizer
self.cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(
logits=self.logits, labels=self.Y))
self.optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(self.cost)
correct_prediction = tf.equal(
tf.argmax(self.logits, 1), tf.argmax(self.Y, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# 예측하는 것, 정확도 구하는 것, 학습하는 것에 대한 함수를 만든다.
def predict(self, x_test, keep_prop=1.0):
return self.sess.run(self.logits, feed_dict={self.X: x_test, self.keep_prob: keep_prop})
def get_accuracy(self, x_test, y_test, keep_prop=1.0):
return self.sess.run(self.accuracy, feed_dict={self.X: x_test, self.Y: y_test, self.keep_prob: keep_prop})
def train(self, x_data, y_data, keep_prop=0.7):
return self.sess.run([self.cost, self.optimizer], feed_dict={
self.X: x_data, self.Y: y_data, self.keep_prob: keep_prop})
이렇게 클래스로 만들게 되면 편리한 것은 다음과 같이 session을 열고 클래스를 만들수 있다.
In [3]:
#initialize
sess = tf.Session()
m1 = Model(sess, "m1")
sess.run(tf.global_variables_initializer())
print('Learning Started')
# train my model
for epoch in range(training_epochs):
avg_cost = 0
total_batch = int(mnist.train.num_examples / batch_size)
for i in range(total_batch):
batch_xs, batch_ys = mnist.train.next_batch(batch_size) # xs와 ys를 받아왔다.
c, _ = m1.train(batch_xs, batch_ys) # 여기서 session run 할 필요 없이 m1의 도움함수 train을 바로 호출 해버린다.
avg_cost += c / total_batch
print('Epoch:', '%04d' % (epoch + 1), 'cost =', '{:.9f}'.format(avg_cost))
print('Learning Finished!')
# Test model and check accuracy
print('Accuracy:', m1.get_accuracy(mnist.test.images, mnist.test.labels))
tf.layers¶
이 패키지는 계속 추가되고 있는 중 인데, 우리가 많이 사용하는 conv2d라던지 pooling이라던지 dense같은 것들을 만들어 두 high leve api. 숫자같은 것이 나와 복잡했던 작업들을 단순하게 처리할 수 있다.
In [4]:
from PIL import Image
Image.open('tflayer.png')
Out[4]:
In [5]:
Image.open('tflayer2.png')
Out[5]:
In [6]:
class Model:
def __init__(self, sess, name):
self.sess = sess
self.name = name
self._build_net()
def _build_net(self):
with tf.variable_scope(self.name):
# dropout (keep_prob) rate 0.7~0.5 on training, but should be 1
# for testing
self.training = tf.placeholder(tf.bool)
# input place holders
self.X = tf.placeholder(tf.float32, [None, 784])
# img 28x28x1 (black/white), Input Layer
X_img = tf.reshape(self.X, [-1, 28, 28, 1])
self.Y = tf.placeholder(tf.float32, [None, 10])
# Convolutional Layer #1
conv1 = tf.layers.conv2d(inputs=X_img, filters=32, kernel_size=[3, 3],
padding="SAME", activation=tf.nn.relu)
# Pooling Layer #1
pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=[2, 2],
padding="SAME", strides=2)
dropout1 = tf.layers.dropout(inputs=pool1,
rate=0.3, training=self.training)
# Convolutional Layer #2 and Pooling Layer #2
conv2 = tf.layers.conv2d(inputs=dropout1, filters=64, kernel_size=[3, 3],
padding="SAME", activation=tf.nn.relu)
pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2],
padding="SAME", strides=2)
dropout2 = tf.layers.dropout(inputs=pool2,
rate=0.3, training=self.training)
# Convolutional Layer #2 and Pooling Layer #2
conv3 = tf.layers.conv2d(inputs=dropout2, filters=128, kernel_size=[3, 3],
padding="same", activation=tf.nn.relu)
pool3 = tf.layers.max_pooling2d(inputs=conv3, pool_size=[2, 2],
padding="same", strides=2)
dropout3 = tf.layers.dropout(inputs=pool3,
rate=0.3, training=self.training)
# Dense Layer with Relu
flat = tf.reshape(dropout3, [-1, 128 * 4 * 4])
dense4 = tf.layers.dense(inputs=flat,
units=625, activation=tf.nn.relu)
dropout4 = tf.layers.dropout(inputs=dense4,
rate=0.5, training=self.training)
# Logits (no activation) Layer: L5 Final FC 625 inputs -> 10 outputs
self.logits = tf.layers.dense(inputs=dropout4, units=10)
# define cost/loss & optimizer
self.cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(
logits=self.logits, labels=self.Y))
self.optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(self.cost)
correct_prediction = tf.equal(
tf.argmax(self.logits, 1), tf.argmax(self.Y, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
def predict(self, x_test, training=False):
return self.sess.run(self.logits,
feed_dict={self.X: x_test, self.training: training})
def get_accuracy(self, x_test, y_test, training=False):
return self.sess.run(self.accuracy,
feed_dict={self.X: x_test,
self.Y: y_test, self.training: training})
def train(self, x_data, y_data, training=True):
return self.sess.run([self.cost, self.optimizer], feed_dict={
self.X: x_data, self.Y: y_data, self.training: training})
# initialize
sess = tf.Session()
m1 = Model(sess, "m1")
sess.run(tf.global_variables_initializer())
print('Learning Started!')
# train my model
for epoch in range(training_epochs):
avg_cost = 0
total_batch = int(mnist.train.num_examples / batch_size)
for i in range(total_batch):
batch_xs, batch_ys = mnist.train.next_batch(batch_size)
c, _ = m1.train(batch_xs, batch_ys)
avg_cost += c / total_batch
print('Epoch:', '%04d' % (epoch + 1), 'cost =', '{:.9f}'.format(avg_cost))
print('Learning Finished!')
# Test model and check accuracy
print('Accuracy:', m1.get_accuracy(mnist.test.images, mnist.test.labels))
Ensemble¶
앙살블은 여러개를 조합하여 조화롭게 만들어낸다는 의미를 갖고 있다.
Clssification 모델에 대해 앙상블을 적용한다는 것은, 여러개의 독립된 모델을 training 시키고 우리에게 testing 할 새로운 데이터가 들어왔다고 하면 각각의 모델이 예측을 하여 예측한 결과를 조합한다. 이 최종적인 결론이 굉장히 좋은 결과를 가져온다.
In [7]:
Image.open('ensemble.png')
Out[7]:
각각의 모델별로 예측을 시킨다음 회별로 합친다. 이 중에서 가장 점수가 높은것으로 예측하겠다 하는 것이다.
In [8]:
Image.open('ensemble2.png')
Out[8]:
우리가 하고 있는 test에 이 앙상블이 어떻게 적용되는지 확인해 보자.
In [3]:
# 우선 독립된 모델을 여러개 만들어야 한다. 그러기 위해서 class를 활용하면 모델을 쉽게 만들 수 있다.
class Model:
def __init__(self, sess, name):
self.sess = sess
self.name = name
self._build_net()
def _build_net(self):
with tf.variable_scope(self.name):
# dropout (keep_prob) rate 0.7~0.5 on training, but should be 1
# for testing
self.training = tf.placeholder(tf.bool)
# input place holders
self.X = tf.placeholder(tf.float32, [None, 784])
# img 28x28x1 (black/white), Input Layer
X_img = tf.reshape(self.X, [-1, 28, 28, 1])
self.Y = tf.placeholder(tf.float32, [None, 10])
# Convolutional Layer #1
conv1 = tf.layers.conv2d(inputs=X_img, filters=32, kernel_size=[3, 3],
padding="SAME", activation=tf.nn.relu)
# Pooling Layer #1
pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=[2, 2],
padding="SAME", strides=2)
dropout1 = tf.layers.dropout(inputs=pool1,
rate=0.3, training=self.training)
# Convolutional Layer #2 and Pooling Layer #2
conv2 = tf.layers.conv2d(inputs=dropout1, filters=64, kernel_size=[3, 3],
padding="SAME", activation=tf.nn.relu)
pool2 = tf.layers.max_pooling2d(inputs=conv2, pool_size=[2, 2],
padding="SAME", strides=2)
dropout2 = tf.layers.dropout(inputs=pool2,
rate=0.3, training=self.training)
# Convolutional Layer #3 and Pooling Layer #3
conv3 = tf.layers.conv2d(inputs=dropout2, filters=128, kernel_size=[3, 3],
padding="SAME", activation=tf.nn.relu)
pool3 = tf.layers.max_pooling2d(inputs=conv3, pool_size=[2, 2],
padding="SAME", strides=2)
dropout3 = tf.layers.dropout(inputs=pool3,
rate=0.3, training=self.training)
# Dense Layer with Relu
flat = tf.reshape(dropout3, [-1, 128 * 4 * 4])
dense4 = tf.layers.dense(inputs=flat,
units=625, activation=tf.nn.relu)
dropout4 = tf.layers.dropout(inputs=dense4,
rate=0.5, training=self.training)
# Logits (no activation) Layer: L5 Final FC 625 inputs -> 10 outputs
self.logits = tf.layers.dense(inputs=dropout4, units=10)
# define cost/loss & optimizer
self.cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(
logits=self.logits, labels=self.Y))
self.optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(self.cost)
correct_prediction = tf.equal(
tf.argmax(self.logits, 1), tf.argmax(self.Y, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
def predict(self, x_test, training=False):
return self.sess.run(self.logits,
feed_dict={self.X: x_test, self.training: training})
def get_accuracy(self, x_test, y_test, training=False):
return self.sess.run(self.accuracy,
feed_dict={self.X: x_test,
self.Y: y_test, self.training: training})
def train(self, x_data, y_data, training=True):
return self.sess.run([self.cost, self.optimizer], feed_dict={
self.X: x_data, self.Y: y_data, self.training: training})
# initialize
sess = tf.Session()
#모델을 담아낼 리스트를 만들고 몇 개의 모델을 할지 정한다.(여기는 2개)
#그 다음 for루프를 돌며 7개의 모델을 만들어 낸다. 클래스를 소개하면 클래스 인스턴스가 생겨나고 리스트에 추가되는 구조이다.
models = []
num_models = 2
for m in range(num_models):
models.append(Model(sess, "model" + str(m)))
sess.run(tf.global_variables_initializer())
print('Learning Started!')
# train my model
# 에폭마다 필요한 배치 사이즈(xs, ys)를 불러와서 학습을 시킨다.
for epoch in range(training_epochs):
avg_cost_list = np.zeros(len(models))
total_batch = int(mnist.train.num_examples / batch_size)
for i in range(total_batch):
batch_xs, batch_ys = mnist.train.next_batch(batch_size)
# train each model. 기존에는 모델 하나만 학습시켰던 반면 앙상블 에서는 각각의 모델에 대해 학습을 시킨다.
for m_idx, m in enumerate(models):
c, _ = m.train(batch_xs, batch_ys)
avg_cost_list[m_idx] += c / total_batch # 각각의 모델별로 cost를 구해준다.
print('Epoch:', '%04d' % (epoch + 1), 'cost =', avg_cost_list)
print('Learning Finished!')
# Test model and check accuracy
test_size = len(mnist.test.labels)
predictions = np.zeros([test_size, 10])
for m_idx, m in enumerate(models):
print(m_idx, 'Accuracy:', m.get_accuracy(
mnist.test.images, mnist.test.labels))
p = m.predict(mnist.test.images) # 각각의 모델에 대해서 예측
predictions += p # 예측 한 것들의 합을 구한다.
ensemble_correct_prediction = tf.equal(
tf.argmax(predictions, 1), tf.argmax(mnist.test.labels, 1))
ensemble_accuracy = tf.reduce_mean(
tf.cast(ensemble_correct_prediction, tf.float32))
print('Ensemble accuracy:', sess.run(ensemble_accuracy))
'beginner > 파이썬 딥러닝 기초' 카테고리의 다른 글
| RNN in TensorFlow (0) | 2019.05.14 |
|---|---|
| NN의 꽃 RNN 이야기 (0) | 2019.05.13 |
| TensorFlow로 MNIST 99% 구현하기 (0) | 2019.05.13 |
| TensorFlow CNN의 기본 (0) | 2019.05.12 |
| ConvNet의 활용 (0) | 2019.05.12 |