[tensorflow2.x 기초 - 4] MNIST data를 활용해 optimization, loss_function, training 구현 기본 (keras)

 

 

 

 

 

Optimization & Training (Beginner)¶

 

• tf와 layers 패키지 불러오기

In [1]:

import tensorflow as tf
from tensorflow.keras import layers

from tensorflow.keras import datasets

 

학습 과정 돌아보기¶

 

Data -> (Model -> logit -> loss -> Optm : 반복) => result

 

Prepare MNIST Datset¶

In [2]:

(train_x, train_y), (test_x, test_y) = datasets.mnist.load_data()

 

Build Model¶

In [3]:

inputs = layers.Input((28, 28, 1))
net = layers.Conv2D(32, (3, 3), padding='SAME')(inputs)
net = layers.Activation('relu')(net)
net = layers.Conv2D(32, (3, 3), padding='SAME')(net)
net = layers.Activation('relu')(net)
net = layers.MaxPooling2D(pool_size=(2, 2))(net)
net = layers.Dropout(0.25)(net)

net = layers.Conv2D(64, (3, 3), padding='SAME')(net)
net = layers.Activation('relu')(net)
net = layers.Conv2D(64, (3, 3), padding='SAME')(net)
net = layers.Activation('relu')(net)
net = layers.MaxPooling2D(pool_size=(2, 2))(net)
net = layers.Dropout(0.25)(net)

net = layers.Flatten()(net)
net = layers.Dense(512)(net)
net = layers.Activation('relu')(net)
net = layers.Dropout(0.5)(net)
net = layers.Dense(10)(net)  # num_classes
net = layers.Activation('softmax')(net)

model = tf.keras.Model(inputs=inputs, outputs=net, name='Basic_CNN')

 

Optimization¶

모델을 학습하기 전 설정

 

• Loss Function
• Optimization
• Metrics

 

Loss Function¶

Loss Function 방법 확인

 

Categorical vs Binary¶

In [4]:

# binary
loss = 'binary_crossentropy'

# more than two categories
loss = 'categorical_crossentropy'

 

sparse_categorical_crossentropy vs categorical_crossentropy¶

In [5]:

# 사용할 loss function, one hot encoding이 필요하지 않음
loss_function = tf.keras.losses.sparse_categorical_crossentropy

In [6]:

tf.keras.losses.categorical_crossentropy

Out[6]:

<function tensorflow.python.keras.losses.categorical_crossentropy(y_true, y_pred, from_logits=False, label_smoothing=0)>

In [7]:

tf.keras.losses.binary_crossentropy

Out[7]:

<function tensorflow.python.keras.losses.binary_crossentropy(y_true, y_pred, from_logits=False, label_smoothing=0)>

 

Metrics¶

모델을 평가하는 방법

 

accuracy를 이름으로 넣는 방법

In [8]:

metrics = ['accuracy']

 

tf.keras.metrics.

In [26]:

tf.keras.metrics.Accuracy()
tf.keras.metrics.Precision()
tf.keras.metrics.Recall()

metrics = tf.keras.metrics.SparseCategoricalAccuracy()

 

Compile¶

Optimizer 적용

 

• 'sgd'
• 'rmsprop'
• 'adam'

In [27]:

optm = tf.keras.optimizers.Adam()

 

• tf.keras.optimizers.SGD()
• tf.keras.optimizers.RMSprop()
• tf.keras.optimizers.Adam()

In [28]:

model.compile(optimizer = optm,
              loss = loss_function,
              metrics = metrics)

In [ ]:

 

Prepare Dataset¶

학습에 사용할 데이터셋 준비

 

shape 확인

In [12]:

train_x.shape, train_y.shape

Out[12]:

((60000, 28, 28), (60000,))

In [13]:

test_x.shape, test_y.shape

Out[13]:

((10000, 28, 28), (10000,))

 

차원 수 늘리기

In [14]:

import numpy as np

In [15]:

np.expand_dims(train_x, -1).shape

Out[15]:

(60000, 28, 28, 1)

In [16]:

tf.expand_dims(train_x, -1).shape

Out[16]:

TensorShape([60000, 28, 28, 1])

In [17]:

train_x = train_x[..., tf.newaxis]
test_x = test_x[..., tf.newaxis]

In [18]:

#train_y = train_y[..., tf.newaxis]
#test_y = test_y[..., tf.newaxis]

 

차원 수 잘 늘었는지 확인

In [19]:

train_x.shape, test_x.shape

Out[19]:

((60000, 28, 28, 1), (10000, 28, 28, 1))

 

Rescaling

In [20]:

np.min(train_x), np.max(train_x)

Out[20]:

(0, 255)

In [21]:

train_x = train_x / 255.

In [22]:

test_x = test_x / 255

In [23]:

np.min(train_x), np.max(train_x)

Out[23]:

(0.0, 1.0)

 

Training¶

본격적으로 학습 들어가기

 

학습용 Hyperparameter 설정

• num_epochs : 반복횟수, 데이터를 한 번 보면 epoch - 1
• batch_size : 한 번에 들어가는 데이터의 수

In [24]:

num_epochs = 1
batch_size = 32

 

• model.fit

In [29]:

model.fit(train_x,train_y,
          batch_size = batch_size,
          shuffle = True,
          epochs = num_epochs)

 

1875/1875 [==============================] - 154s 82ms/step - loss: 0.1473 - sparse_categorical_accuracy: 0.9536

Out[29]:

<tensorflow.python.keras.callbacks.History at 0x1c63b289488>

 

Check History¶

학습 과정(History) 결과 확인

In [ ]: