4. Basic CNN Implementation(v1)

 

 

 

 

 

Basic Convolutional Neural Network Implementation¶

In [1]:

import tensorflow as tf
import numpy as np

In [2]:

#1. hyperparameter
Epochs = 10

In [17]:

#2. build model #VGG16과 비슷한 모델 : pooling을 하기 전 동일한 conv layer을 반복해서 사용하는 것이 특징입니다. 
class ConvNet(tf.keras.Model):
    def __init__(self):
        super(ConvNet, self).__init__()
        self.sequence = list()
        self.sequence.append(tf.keras.layers.Conv2D(16, (3,3), padding = 'same', activation = 'relu'))#28x28x16
        self.sequence.append(tf.keras.layers.Conv2D(16, (3, 3), padding = 'same', activation = 'relu'))#28x28x16
        self.sequence.append(tf.keras.layers.MaxPool2D((2, 2))) #14x14x16
        
        self.sequence.append(tf.keras.layers.Conv2D(32, (3,3), padding = 'same', activation = 'relu'))#14x14x32
        self.sequence.append(tf.keras.layers.Conv2D(32, (3, 3), padding = 'same', activation = 'relu'))#14x14x32
        self.sequence.append(tf.keras.layers.MaxPool2D((2, 2))) #7x7x32
        
        self.sequence.append(tf.keras.layers.Conv2D(64, (3,3), padding = 'same', activation = 'relu'))#7x7x64
        self.sequence.append(tf.keras.layers.Conv2D(64, (3, 3), padding = 'same', activation = 'relu'))#7x7x64
        
        self.sequence.append(tf.keras.layers.Flatten()) #7x7x64 = 1568 크기의 feature vector가 완성됩니다. 
        
        self.sequence.append(tf.keras.layers.Dense(2048, activation = 'relu')) #이 feature vector을 10개로 classification하기 위해서 Dense layer 사용
        self.sequence.append(tf.keras.layers.Dense(10, activation = 'softmax'))
        
        #이렇게 사용할 모든 layer을 선언해 주었습니다. 
        
    def call(self, x, training = False, mask = None): #이렇게 선언된 layer들을 call 에서 모두 연결시켜 주어야 합니다. 
        #list에 각 layer들을 넣어주었기 때문에 각 sequence에 있는 layer을 순회하며 돌면서 연결해줘서 최종 출력까지 갈 수 있습니다. 
        for layer in self.sequence:
                x = layer(x)
        return x
        

In [35]:

#3. Training function
@tf.function
def train_step(model, images, labels, loss_object, optimizer, train_loss, train_accuracy):
    with tf.GradientTape() as tape: #tf.GradientTape을 사용하여 그래디언트를 계산합니다.
        predictions = model(images)
        loss = loss_object(labels, predictions)
    gradients = tape.gradient(loss, model.trainable_variables)
    
    optimizer.apply_gradients(zip(gradients, model.trainable_variables))
    train_loss(loss)
    train_accuracy(labels, predictions)
    

In [36]:

#4. Test function
@tf.function
def test_step(model, images, labels, loss_object, test_loss, test_accuracy):
    predictions = model(images)
    
    t_loss = loss_object(labels, predictions)
    test_loss(t_loss)
    test_accuracy(labels, predictions)

In [37]:

#5. Prepare Dataset : keras에서 제공하는 mnist dataset 사용

#call Data
mnist = tf.keras.datasets.mnist #이렇게 keras에서 mnist데이터를 가져옵니다. 
(X_train, y_train), (X_test, y_test) = mnist.load_data() #이 데이터의 구성이 train/test로 나뉘어져 있다. 

#Normalization
X_train, X_test = X_train/255.0, X_test/255.0 #0~255로 표현이 되어 있으니 0~1로 바꿔줄 수 있다. 

#(num_sample, height, weight)-> (num_sample, height, weight, channel in)추가
X_train = X_train[..., tf.newaxis].astype(np.float32)#casting도 잊지말고 해줍니다. 
X_test = X_test[..., tf.newaxis].astype(np.float32)

#dataset 만둘어주기
train_ds = tf.data.Dataset.from_tensor_slices((X_train, y_train)).shuffle(10000).batch(32) #tf.data.Dataset.from_tensor_slices() 이 함수를 이용하면 numpy나 tensor로부터 dataset을 구축할 수 있습니다. 
test_ds = tf.data.Dataset.from_tensor_slices((X_test, y_test)).batch(32) #test는 shuffle이 필요없습니다. 

In [38]:

#6. Define the training environment

#create model
model = ConvNet()

#Define loss and optimizer
loss_object = tf.keras.losses.SparseCategoricalCrossentropy()
optimizer = tf.keras.optimizers.Adam()

#Define performance metrics
train_loss = tf.keras.metrics.Mean(name = 'train_loss')
train_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name = 'train_accuracy')

test_loss = tf.keras.metrics.Mean(name = 'test_loss')
test_accuracy = tf.keras.metrics.SparseCategoricalAccuracy(name = 'test_accuracy')

In [ ]:

# 7. Training

for epoch in range(Epochs):
    for images, labels in train_ds:
        train_step(model, images, labels, loss_object, optimizer, train_loss, train_accuracy)
        
    for test_images, test_labels in test_ds:
        test_step(model, test_images, test_labels, loss_object, test_loss, test_accuracy)
        
    template = 'Epoch {}, loss: {}, Accuracy: {}, Test Loss: {}, Test Accuracy: {}'
    print(template.format(epoch+1, 
                         train_loss.result(), #result() — 누적된 측정 결과를 얻습니다.
                         train_accuracy.result()*100, 
                         test_loss.result(), 
                         test_accuracy.result()*100))

 

Epoch 1, loss: 0.0779663473367691, Accuracy: 97.57250213623047, Test Loss: 0.03051144815981388, Test Accuracy: 98.97999572753906
Epoch 2, loss: 0.06343494355678558, Accuracy: 98.03111267089844, Test Loss: 0.033057499676942825, Test Accuracy: 98.97000122070312

In [ ]:

참고 : https://www.tensorflow.org/guide/migrate?hl=ko