기말고사 예상문제
빅데이터분석특강
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
import tensorflow.experimental.numpy as tnp
tnp.experimental_enable_numpy_behavior()
%load_ext tensorboard
import graphviz
def gv(s): return graphviz.Source('digraph G{ rankdir="LR"'+ s + ';}')
(1) tf.keras.datasets.fashion_mnist.load_data()을 이용하여 fashion_mnist 자료를 불러온 뒤 아래의 네트워크를 이용하여 적합하라.
- 평가지표로 accuracy를 이용할 것
- epoch은 10으로 설정할 것
- optimizer는 adam을 이용할 것
gv('''
splines=line
subgraph cluster_1{
style=filled;
color=lightgrey;
"x1"
"x2"
".."
"x784"
label = "Layer 0"
}
subgraph cluster_2{
style=filled;
color=lightgrey;
"x1" -> "node1"
"x2" -> "node1"
".." -> "node1"
"x784" -> "node1"
"x1" -> "node2"
"x2" -> "node2"
".." -> "node2"
"x784" -> "node2"
"x1" -> "..."
"x2" -> "..."
".." -> "..."
"x784" -> "..."
"x1" -> "node20"
"x2" -> "node20"
".." -> "node20"
"x784" -> "node20"
label = "Layer 1: relu"
}
subgraph cluster_3{
style=filled;
color=lightgrey;
"node1" -> "node1 "
"node2" -> "node1 "
"..." -> "node1 "
"node20" -> "node1 "
"node1" -> "node2 "
"node2" -> "node2 "
"..." -> "node2 "
"node20" -> "node2 "
"node1" -> "... "
"node2" -> "... "
"..." -> "... "
"node20" -> "... "
"node1" -> "node30 "
"node2" -> "node30 "
"..." -> "node30 "
"node20" -> "node30 "
label = "Layer 2: relu"
}
subgraph cluster_4{
style=filled;
color=lightgrey;
"node1 " -> "y10"
"node2 " -> "y10"
"... " -> "y10"
"node30 " -> "y10"
"node1 " -> "y1"
"node2 " -> "y1"
"... " -> "y1"
"node30 " -> "y1"
"node1 " -> "."
"node2 " -> "."
"... " -> "."
"node30 " -> "."
label = "Layer 3: softmax"
}
''')
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.fashion_mnist.load_data()
X = tf.constant(x_train.reshape(-1,28,28,1),dtype=tf.float64)
y = tf.keras.utils.to_categorical(y_train)
XX = tf.constant(x_test.reshape(-1,28,28,1),dtype=tf.float64)
yy = tf.keras.utils.to_categorical(y_test)
tf.random.set_seed(4305)
net = tf.keras.Sequential()
net.add(tf.keras.layers.Flatten())
net.add(tf.keras.layers.Dense(20,activation='relu'))
net.add(tf.keras.layers.Dense(30,activation='relu'))
net.add(tf.keras.layers.Dense(10,activation='softmax'))
net.compile(loss=tf.losses.categorical_crossentropy, optimizer='adam',metrics=['accuracy'])
net.fit(X,y,epochs=10)
(2) (1)에서 적합된 네트워크를 이용하여 test data의 accuracy를 구하라.
net.evaluate(XX,yy)
(3) train set에서 20%의 자료를 validation 으로 분리하여 50에폭동안 학습하라. 텐서보드를 이용하여 train accuracy와 validation accuracy를 시각화 하고 결과를 해석하라. 오버피팅이라고 볼 수 있는가?
tf.random.set_seed(4305)
net = tf.keras.Sequential()
net.add(tf.keras.layers.Flatten())
net.add(tf.keras.layers.Dense(20,activation='relu'))
net.add(tf.keras.layers.Dense(30,activation='relu'))
net.add(tf.keras.layers.Dense(10,activation='softmax'))
net.compile(loss=tf.losses.categorical_crossentropy, optimizer='adam',metrics=['accuracy'])
cb1 = tf.keras.callbacks.TensorBoard()
net.fit(X,y,epochs=50,batch_size=200,validation_split=0.2,callbacks=cb1,verbose=1)
%tensorboard --logdir logs --host 0.0.0.0
(4) (3)에서 적합된 네트워크를 이용하여 test data의 accuracy를 구하라. (2)의 결과와 비교하라.
net.evaluate(XX,yy)
(5) 조기종료기능을 이용하여 (3)의 네트워크를 다시 학습하라. 학습결과를 텐서보드를 이용하여 시각화 하라.
- patience=3 으로 설정할 것
tf.random.set_seed(4305)
net = tf.keras.Sequential()
net.add(tf.keras.layers.Flatten())
net.add(tf.keras.layers.Dense(20,activation='relu'))
net.add(tf.keras.layers.Dense(30,activation='relu'))
net.add(tf.keras.layers.Dense(10,activation='softmax'))
net.compile(loss=tf.losses.categorical_crossentropy, optimizer='adam',metrics=['accuracy'])
cb1 = tf.keras.callbacks.TensorBoard()
cb2 = tf.keras.callbacks.EarlyStopping(patience=3)
net.fit(X,y,epochs=50,batch_size=200,validation_split=0.2,callbacks=[cb1,cb2])
%tensorboard --logdir logs --host 0.0.0.0
(1) tf.keras.datasets.fashion_mnist.load_data()을 이용하여 fashion_mnist 자료를 불러온 뒤 아래의 네트워크를 이용하여 적합하라.
- 이때 n1=6, n2=16, n3=120 으로 설정한다, 드랍아웃비율은 20%로 설정한다.
net.summary()를 출력하여 설계결과를 확인하라.
tf.random.set_seed(4305)
net1 = tf.keras.Sequential()
net1.add(tf.keras.layers.Conv2D(6,(4,4),activation='relu'))
net1.add(tf.keras.layers.MaxPool2D())
net1.add(tf.keras.layers.Conv2D(16,(4,4),activation='relu'))
net1.add(tf.keras.layers.MaxPool2D())
net1.add(tf.keras.layers.Flatten())
net1.add(tf.keras.layers.Dense(120,activation='relu'))
net1.add(tf.keras.layers.Dense(10,activation='softmax'))
net1.add(tf.keras.layers.Dropout(0.2))
net1.compile(optimizer='adam', loss=tf.losses.categorical_crossentropy,metrics='accuracy')
net1.fit(X,y,epochs=5,batch_size=200)
net1.evaluate(XX,yy)
net1.summary()
c1, m1, c2, m2, flttn, dns1, dns2, dropout = net1.layers
print(X.shape)
print(c1(X).shape)
print(m1(c1(X)).shape)
print(c2(m1(c1(X))).shape)
print(m2(c2(m1(c1(X)))).shape)
print(flttn(m2(c2(m1(c1(X))))).shape)
print(dns1(flttn(m2(c2(m1(c1(X)))))).shape)
print(dns2(dns1(flttn(m2(c2(m1(c1(X))))))).shape)
print(dropout(dns2(dns1(flttn(m2(c2(m1(c1(X)))))))).shape)
(2) n1=(6,64,128), n2=(16,256)에 대하여 test set의 loss가 최소화되는 조합을 찾아라. 결과를 텐서보드로 시각화하는 코드를 작성하라.
- epoc은 3회로 한정한다.
- validation_split은 0.2로 설정한다.
from tensorboard.plugins.hparams import api as hp
!rm -rf logs
for u in [6,64,128]:
for d in [16,256]:
logdir = 'logs/hpguebin_{}_{}'.format(u,d)
with tf.summary.create_file_writer(logdir).as_default():
tf.random.set_seed(4305)
net1 = tf.keras.Sequential()
net1.add(tf.keras.layers.Conv2D(6,(4,4),activation='relu'))
net1.add(tf.keras.layers.MaxPool2D())
net1.add(tf.keras.layers.Conv2D(16,(4,4),activation='relu'))
net1.add(tf.keras.layers.MaxPool2D())
net1.add(tf.keras.layers.Flatten())
net1.add(tf.keras.layers.Dense(120,activation='relu'))
net1.add(tf.keras.layers.Dense(10,activation='softmax'))
net1.add(tf.keras.layers.Dropout(0.2))
net1.compile(optimizer='adam', loss=tf.losses.categorical_crossentropy,metrics='accuracy')
cb3 = hp.KerasCallback(logdir, {'n1':u, 'n2':d})
net1.fit(X,y,epochs=3,batch_size=200,callbacks=cb3,validation_split=0.2)
_rslt=net.evaluate(XX,yy)
tf.summary.scalar('test set loss', _rslt[0], step=1)
%tensorboard --logdir logs --host 0.0.0.0
tf.keras.datasets.cifar10.load_data()을 이용하여 CIFAR10을 불러온 뒤 적당한 네트워크를 사용하여 적합하라.
- 결과를 텐서보드로 시각화할 필요는 없다.
- 자유롭게 모형을 설계하여 적합하라.
- test set의 accuracy가 70%이상인 경우만 정답으로 인정한다.
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.cifar10.load_data()
x_train.shape
X = tf.constant(x_train.reshape(-1,32,32,3),dtype=tf.float64)
y = tf.keras.utils.to_categorical(y_train)
XX = tf.constant(x_test.reshape(-1,32,32,3),dtype=tf.float64)
yy = tf.keras.utils.to_categorical(y_test)
print(X.shape)
print(y.shape)
print(XX.shape)
print(yy.shape)
net2 = tf.keras.Sequential()
net2.add(tf.keras.layers.Conv2D(512,(2,2),activation='relu'))
net2.add(tf.keras.layers.Conv2D(512,(2,2),activation='relu'))
net2.add(tf.keras.layers.Dropout(0.5))
net2.add(tf.keras.layers.MaxPool2D())
net2.add(tf.keras.layers.Conv2D(512,(2,2),activation='relu'))
net2.add(tf.keras.layers.Conv2D(512,(2,2),activation='relu'))
net2.add(tf.keras.layers.Dropout(0.5))
net2.add(tf.keras.layers.MaxPool2D())
net2.add(tf.keras.layers.Flatten())
net2.add(tf.keras.layers.Dense(10,activation='softmax'))
net2.compile(optimizer='adam', loss=tf.losses.categorical_crossentropy,metrics='accuracy')
net2.fit(X,y,epochs=5,validation_split=0.2)
net2.fit(X,y,epochs=10,validation_split=0.2)
net2.evaluate(XX,yy)
(1) (1,128,128,3)의 shape을 가진 텐서가 tf.keras.layers.Conv2D(5,(2,2))으로 만들어진 커널을 통과할시 나오는 shape은?
tf.random.set_seed(43052)
cnv = tf.keras.layers.Conv2D(5,(2,2))
XXX = tnp.array([1]*1*128*128*3,dtype=tf.float64).reshape(1,128,128,3)
cnv(XXX)
답 : (1, 127, 127, 5)
(2) (1,24,24,16)의 shape을 가진 텐서가 tf.keras.layers.Flatten()을 통과할때 나오는 텐서의 shape은?
24*24*16
답 : (1, 9216)
(3)
(4)
(5)