DataScience/TensorFlow[ANN]
딥러닝 텐서플로우 10개로 분류된 패션이미지 분류 예시, softmax, np.argmax(axis=1), overfitting, callbacks
leopard4
2022. 12. 29. 11:07
A Computer Vision Example¶
10개로 분류된 패션 이미지를 분류하는 딥러닝
Start Coding¶
In [1]:
# 버전확인 // 버전이 자주바뀌어 함수문제가 일어날수도 있기때무네
import tensorflow as tf
print(tf.__version__)
2.9.2
Fashion MNIST 데이터는 tf.keras datasets API에 들어있다.
In [2]:
mnist = tf.keras.datasets.fashion_mnist # 텐서플로우 기본제공 데이터
In [3]:
(X_train, y_train),(X_test, y_test) = mnist.load_data() # 트레인먼저 가져오고 테스트를 가져온다 .
# 원래는 storage서버에서 jpg,png등으로 가져온것을 numpy로 변환하는 작업을 해야하지만 제공데이터는 기본으로 되어있다.
Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-labels-idx1-ubyte.gz 29515/29515 [==============================] - 0s 0us/step Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/train-images-idx3-ubyte.gz 26421880/26421880 [==============================] - 2s 0us/step Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-labels-idx1-ubyte.gz 5148/5148 [==============================] - 0s 0us/step Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/t10k-images-idx3-ubyte.gz 4422102/4422102 [==============================] - 0s 0us/step
트레이닝과 테스트셋 가져온다.
In [4]:
X_train.shape # 갯수가 맨앞으로간다
Out[4]:
(60000, 28, 28)
In [5]:
y_train # 첫번째 사진에대한 정답 == 9
# uint8 == 언사인드(양수0~255) 정수 8바이트
Out[5]:
array([9, 0, 0, ..., 3, 0, 5], dtype=uint8)
In [6]:
y_train.shape
Out[6]:
(60000,)
In [7]:
X_test.shape
Out[7]:
(10000, 28, 28)
In [8]:
y_test.shape
Out[8]:
(10000,)
이미지는 숫자로 되어있다. 0부터 255까지의 숫자로 되어있다.
In [9]:
X_train[0, :, :] # == X_train[0][:][:] (리스트의문법) 3차원데이터 액세스
Out[9]:
array([[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1,
0, 0, 13, 73, 0, 0, 1, 4, 0, 0, 0, 0, 1,
1, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3,
0, 36, 136, 127, 62, 54, 0, 0, 0, 1, 3, 4, 0,
0, 3],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 6,
0, 102, 204, 176, 134, 144, 123, 23, 0, 0, 0, 0, 12,
10, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 155, 236, 207, 178, 107, 156, 161, 109, 64, 23, 77, 130,
72, 15],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0,
69, 207, 223, 218, 216, 216, 163, 127, 121, 122, 146, 141, 88,
172, 66],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0,
200, 232, 232, 233, 229, 223, 223, 215, 213, 164, 127, 123, 196,
229, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
183, 225, 216, 223, 228, 235, 227, 224, 222, 224, 221, 223, 245,
173, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
193, 228, 218, 213, 198, 180, 212, 210, 211, 213, 223, 220, 243,
202, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 3, 0, 12,
219, 220, 212, 218, 192, 169, 227, 208, 218, 224, 212, 226, 197,
209, 52],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 6, 0, 99,
244, 222, 220, 218, 203, 198, 221, 215, 213, 222, 220, 245, 119,
167, 56],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 4, 0, 0, 55,
236, 228, 230, 228, 240, 232, 213, 218, 223, 234, 217, 217, 209,
92, 0],
[ 0, 0, 1, 4, 6, 7, 2, 0, 0, 0, 0, 0, 237,
226, 217, 223, 222, 219, 222, 221, 216, 223, 229, 215, 218, 255,
77, 0],
[ 0, 3, 0, 0, 0, 0, 0, 0, 0, 62, 145, 204, 228,
207, 213, 221, 218, 208, 211, 218, 224, 223, 219, 215, 224, 244,
159, 0],
[ 0, 0, 0, 0, 18, 44, 82, 107, 189, 228, 220, 222, 217,
226, 200, 205, 211, 230, 224, 234, 176, 188, 250, 248, 233, 238,
215, 0],
[ 0, 57, 187, 208, 224, 221, 224, 208, 204, 214, 208, 209, 200,
159, 245, 193, 206, 223, 255, 255, 221, 234, 221, 211, 220, 232,
246, 0],
[ 3, 202, 228, 224, 221, 211, 211, 214, 205, 205, 205, 220, 240,
80, 150, 255, 229, 221, 188, 154, 191, 210, 204, 209, 222, 228,
225, 0],
[ 98, 233, 198, 210, 222, 229, 229, 234, 249, 220, 194, 215, 217,
241, 65, 73, 106, 117, 168, 219, 221, 215, 217, 223, 223, 224,
229, 29],
[ 75, 204, 212, 204, 193, 205, 211, 225, 216, 185, 197, 206, 198,
213, 240, 195, 227, 245, 239, 223, 218, 212, 209, 222, 220, 221,
230, 67],
[ 48, 203, 183, 194, 213, 197, 185, 190, 194, 192, 202, 214, 219,
221, 220, 236, 225, 216, 199, 206, 186, 181, 177, 172, 181, 205,
206, 115],
[ 0, 122, 219, 193, 179, 171, 183, 196, 204, 210, 213, 207, 211,
210, 200, 196, 194, 191, 195, 191, 198, 192, 176, 156, 167, 177,
210, 92],
[ 0, 0, 74, 189, 212, 191, 175, 172, 175, 181, 185, 188, 189,
188, 193, 198, 204, 209, 210, 210, 211, 188, 188, 194, 192, 216,
170, 0],
[ 2, 0, 0, 0, 66, 200, 222, 237, 239, 242, 246, 243, 244,
221, 220, 193, 191, 179, 182, 182, 181, 176, 166, 168, 99, 58,
0, 0],
[ 0, 0, 0, 0, 0, 0, 0, 40, 61, 44, 72, 41, 35,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0]], dtype=uint8)
In [10]:
# 첫번째이미지만 가져온다.
X_train[0]
Out[10]:
array([[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1,
0, 0, 13, 73, 0, 0, 1, 4, 0, 0, 0, 0, 1,
1, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 3,
0, 36, 136, 127, 62, 54, 0, 0, 0, 1, 3, 4, 0,
0, 3],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 6,
0, 102, 204, 176, 134, 144, 123, 23, 0, 0, 0, 0, 12,
10, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 155, 236, 207, 178, 107, 156, 161, 109, 64, 23, 77, 130,
72, 15],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0,
69, 207, 223, 218, 216, 216, 163, 127, 121, 122, 146, 141, 88,
172, 66],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0,
200, 232, 232, 233, 229, 223, 223, 215, 213, 164, 127, 123, 196,
229, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
183, 225, 216, 223, 228, 235, 227, 224, 222, 224, 221, 223, 245,
173, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
193, 228, 218, 213, 198, 180, 212, 210, 211, 213, 223, 220, 243,
202, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 3, 0, 12,
219, 220, 212, 218, 192, 169, 227, 208, 218, 224, 212, 226, 197,
209, 52],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 6, 0, 99,
244, 222, 220, 218, 203, 198, 221, 215, 213, 222, 220, 245, 119,
167, 56],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 4, 0, 0, 55,
236, 228, 230, 228, 240, 232, 213, 218, 223, 234, 217, 217, 209,
92, 0],
[ 0, 0, 1, 4, 6, 7, 2, 0, 0, 0, 0, 0, 237,
226, 217, 223, 222, 219, 222, 221, 216, 223, 229, 215, 218, 255,
77, 0],
[ 0, 3, 0, 0, 0, 0, 0, 0, 0, 62, 145, 204, 228,
207, 213, 221, 218, 208, 211, 218, 224, 223, 219, 215, 224, 244,
159, 0],
[ 0, 0, 0, 0, 18, 44, 82, 107, 189, 228, 220, 222, 217,
226, 200, 205, 211, 230, 224, 234, 176, 188, 250, 248, 233, 238,
215, 0],
[ 0, 57, 187, 208, 224, 221, 224, 208, 204, 214, 208, 209, 200,
159, 245, 193, 206, 223, 255, 255, 221, 234, 221, 211, 220, 232,
246, 0],
[ 3, 202, 228, 224, 221, 211, 211, 214, 205, 205, 205, 220, 240,
80, 150, 255, 229, 221, 188, 154, 191, 210, 204, 209, 222, 228,
225, 0],
[ 98, 233, 198, 210, 222, 229, 229, 234, 249, 220, 194, 215, 217,
241, 65, 73, 106, 117, 168, 219, 221, 215, 217, 223, 223, 224,
229, 29],
[ 75, 204, 212, 204, 193, 205, 211, 225, 216, 185, 197, 206, 198,
213, 240, 195, 227, 245, 239, 223, 218, 212, 209, 222, 220, 221,
230, 67],
[ 48, 203, 183, 194, 213, 197, 185, 190, 194, 192, 202, 214, 219,
221, 220, 236, 225, 216, 199, 206, 186, 181, 177, 172, 181, 205,
206, 115],
[ 0, 122, 219, 193, 179, 171, 183, 196, 204, 210, 213, 207, 211,
210, 200, 196, 194, 191, 195, 191, 198, 192, 176, 156, 167, 177,
210, 92],
[ 0, 0, 74, 189, 212, 191, 175, 172, 175, 181, 185, 188, 189,
188, 193, 198, 204, 209, 210, 210, 211, 188, 188, 194, 192, 216,
170, 0],
[ 2, 0, 0, 0, 66, 200, 222, 237, 239, 242, 246, 243, 244,
221, 220, 193, 191, 179, 182, 182, 181, 176, 166, 168, 99, 58,
0, 0],
[ 0, 0, 0, 0, 0, 0, 0, 40, 61, 44, 72, 41, 35,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0]], dtype=uint8)
In [11]:
# 사람용으로 표시
import matplotlib.pyplot as plt
In [12]:
plt.imshow( X_train[0] )
plt.show()
In [13]:
plt.imshow( X_train[200], cmap='gray' )
plt.show()
In [14]:
# 정답지 확인 8==가방(사람이 레이블링한것)
y_train[200]
Out[14]:
8
학습이 잘 되도록 0과 1사이의 값으로 노말라이징(normalizing) 한다.
In [15]:
# minmaxscaler를 해도되지만 귀찮으니까 넘파이 255.0로 나눠버린다(딥러닝은 무조건 플롯 , 인트는안댐)
In [16]:
X_train = X_train / 255.0
In [17]:
X_test = X_test / 255.0
이미지의 가로 세로를 전부 일렬로 만드는 작업이 flatten 이다.
모델을 만들어 보자.
In [18]:
import tensorflow as tf
from tensorflow import keras
from keras.models import Sequential
from keras.layers import Dense
In [19]:
from keras.layers import Flatten
In [19]:
In [19]:
Sequential: SEQUENCE layers 로 모델링
Flatten: 2차원, 3차원을 1차원으로 만드는것
Dense: 뉴런을 레이어에 붙인다.
각 레이어는 activation function 이 필요하다.
Relu 액티베이션 함수 참고 https://docs.google.com/presentation/d/1DCOW7Lum2rnoPU7gxpznimrXV0DOJFc-YkkJt3jZC0U/edit#slide=id.ga8aac09587_0_2
Softmax 여러개의 값 중에서 가장 큰 값을 선택. [0.1, 0.1, 0.05, 0.1, 9.5, 0.1, 0.05, 0.05, 0.05], 여기서 가장 큰 값을 1로 만들고 나머지는 0으로 만들어준다. [0,0,0,0,1,0,0,0,0]
In [20]:
# 1장의 사진 == 2차원을 1차원으로 만드는것 (하나의 행으로 만드는것)
def build_model() :
model = Sequential()
model.add( Flatten() ) # 열과행의 수에 관계없이 알아서 변환해줌
model.add( Dense(128, 'relu'))
model.add( Dense(64, 'relu'))
model.add( Dense(10, 'softmax')) # 2개로 분류할때는 노드1개 엑티베이션=시그모이드, 3개이상일때는 레이블의 갯수만큼
model.compile('adam', 'sparse_categorical_crossentropy', ['accuracy']) # 즉, 여기선10개의 뉴런(노드) 10개도 0과1로 나오게할것,
return model # 엑티베이션=softmax (확률로 나와서 가장큰걸로 결정하는 함수 , 전체의 합은1)
In [21]:
y_train # 레이블인코딩상태
Out[21]:
array([9, 0, 0, ..., 3, 0, 5], dtype=uint8)
In [22]:
# 2개로 분류하는 문제에서의 로스 펑션은 binary_crossentropy 를 사용.
# 3개 이상으로 분류하는 문제에서는, 두가지 방법이 있는데,
# y의 값을 확인하여 선택하는 것이다.
# 첫번째 방법은, y값이 레이블 인코딩으로 되어있으면, sparse_categorical_crossentropy
# 두번째 방법은, y값이 원핫인코딩으로 되어있으면, categorical_crossentropy
optimizer and loss function 으로 컴파일 한 후에 model.fit 함수로 학습한다..
In [23]:
model = build_model()
In [24]:
model.fit(X_train, y_train, epochs = 5)
Epoch 1/5 1875/1875 [==============================] - 11s 4ms/step - loss: 0.4956 - accuracy: 0.8227 Epoch 2/5 1875/1875 [==============================] - 7s 4ms/step - loss: 0.3704 - accuracy: 0.8640 Epoch 3/5 1875/1875 [==============================] - 5s 3ms/step - loss: 0.3346 - accuracy: 0.8756 Epoch 4/5 1875/1875 [==============================] - 5s 3ms/step - loss: 0.3114 - accuracy: 0.8844 Epoch 5/5 1875/1875 [==============================] - 5s 3ms/step - loss: 0.2952 - accuracy: 0.8907
Out[24]:
<keras.callbacks.History at 0x7fe0c01501f0>
In [24]:
accuracy 가 약 90% 라는 뜻. 5 epochs 라 금방 끝났다.
학습에 사용하지 않는, 테스트 데이터를 통해서, 얼마나 정확도되 나오는지 확인해 본다.
In [25]:
model.evaluate(X_test, y_test)
313/313 [==============================] - 1s 2ms/step - loss: 0.3690 - accuracy: 0.8645
Out[25]:
[0.3690395951271057, 0.8644999861717224]
In [26]:
# ANN == 기본적인구조
In [27]:
y_pred = model.predict(X_test)
313/313 [==============================] - 1s 2ms/step
In [28]:
y_pred
Out[28]:
array([[6.2642157e-06, 1.2893028e-05, 8.6467444e-06, ..., 5.1825982e-01,
7.6142972e-04, 4.7485906e-01],
[9.8711294e-05, 2.6011189e-09, 9.9776161e-01, ..., 1.2294908e-12,
2.8359122e-07, 1.7571462e-10],
[2.3507123e-09, 1.0000000e+00, 1.6610641e-10, ..., 4.3737702e-20,
6.1341182e-10, 9.1725611e-16],
...,
[3.1925822e-04, 4.2129066e-08, 6.7577137e-05, ..., 6.1767085e-09,
9.9935621e-01, 5.1176413e-10],
[6.1121850e-06, 9.9997795e-01, 5.7848951e-07, ..., 4.9597555e-13,
2.4816867e-07, 7.4746270e-10],
[1.5104997e-04, 1.2393270e-04, 6.9673464e-04, ..., 2.1131155e-01,
9.3408413e-03, 5.8493303e-04]], dtype=float32)
In [29]:
# 결과가 맞는지 확인
plt.imshow(X_test[0])
plt.show()
In [30]:
y_pred[0] # 9가 아닌이유는, output을 10개로 만들엇기 때문에 10개로 나온것.
Out[30]:
array([6.2642157e-06, 1.2893028e-05, 8.6467444e-06, 3.0621634e-06,
1.1318511e-05, 6.0653663e-03, 1.2088867e-05, 5.1825982e-01,
7.6142972e-04, 4.7485906e-01], dtype=float32)
In [31]:
y_pred[0].sum()
Out[31]:
1.0
In [32]:
y_pred[0].argmax() # 제일큰값의 인덱스를 찾는함수
Out[32]:
7
In [33]:
9.9584609 * 10**-1 # 99.5퍼 확률로 9라고 예측한것
Out[33]:
0.9958460900000001
In [33]:
# 컨퓨전 매트릭스를 이용해서, 어떤것을 컴퓨터가 많이 헷갈려 하는지 파악한다.
In [34]:
from sklearn.metrics import confusion_matrix
In [36]:
y_test
Out[36]:
array([9, 2, 1, ..., 8, 1, 5], dtype=uint8)
In [40]:
y_pred # 이건 소프트맥스값이기 때문에 아래와같이 변환한다.
Out[40]:
array([[6.2642157e-06, 1.2893028e-05, 8.6467444e-06, ..., 5.1825982e-01,
7.6142972e-04, 4.7485906e-01],
[9.8711294e-05, 2.6011189e-09, 9.9776161e-01, ..., 1.2294908e-12,
2.8359122e-07, 1.7571462e-10],
[2.3507123e-09, 1.0000000e+00, 1.6610641e-10, ..., 4.3737702e-20,
6.1341182e-10, 9.1725611e-16],
...,
[3.1925822e-04, 4.2129066e-08, 6.7577137e-05, ..., 6.1767085e-09,
9.9935621e-01, 5.1176413e-10],
[6.1121850e-06, 9.9997795e-01, 5.7848951e-07, ..., 4.9597555e-13,
2.4816867e-07, 7.4746270e-10],
[1.5104997e-04, 1.2393270e-04, 6.9673464e-04, ..., 2.1131155e-01,
9.3408413e-03, 5.8493303e-04]], dtype=float32)
In [41]:
y_pred = y_pred.argmax(axis=1)
# y_pred.argmax(axis=1) 이것을 풀어보자면 우선 y_pred는 numpy이의 10열 10000개의 데이터이다 (shape을 해보면 알수있음)
# 왜냐하면 []로 이루어진[] ,즉 [[],[]] 의 형태로 이루어져잇기때문
# 여기서 argmax(axis=1) 이라는것은 말로하면 10열의 행별 최대값을 구하라는 말과도 같다.
In [43]:
cm = confusion_matrix(y_test, y_pred)
In [46]:
cm
Out[46]:
array([[782, 0, 31, 57, 5, 2, 113, 0, 10, 0],
[ 0, 960, 4, 24, 8, 0, 2, 0, 2, 0],
[ 11, 0, 816, 20, 128, 0, 25, 0, 0, 0],
[ 8, 5, 11, 913, 34, 0, 25, 0, 4, 0],
[ 0, 0, 94, 38, 849, 0, 18, 0, 1, 0],
[ 0, 0, 0, 1, 0, 909, 0, 47, 2, 41],
[ 93, 0, 133, 46, 158, 0, 551, 0, 19, 0],
[ 0, 0, 0, 0, 0, 8, 0, 983, 0, 9],
[ 4, 0, 3, 5, 3, 2, 3, 7, 973, 0],
[ 0, 0, 0, 0, 0, 1, 1, 89, 0, 909]])
In [44]:
from sklearn.metrics import accuracy_score
In [47]:
accuracy_score(y_test,y_pred) # model.evaluate 과 같다.(내부적으로 컨퓨전매트릭스를 구현한것)
Out[47]:
0.8645
In [48]:
# 그냥보면 불편하니까 차트로 표현
import seaborn as sb
In [50]:
sb.heatmap(cm, annot=True, fmt='.0f', cmap='RdPu')
plt.show()
In [ ]:
# 크리티컬한 데이터는 보완해야함.
Exploration Exercises¶
Exercise 1:¶
테스트 이미지를 예측한 후, 첫번째 이미지에 대해서 예측한 결과를 출력한다.
In [51]:
y_pred
Out[51]:
array([7, 2, 1, ..., 8, 1, 5])
In [52]:
y_pred[0]
Out[52]:
7
In [53]:
plt.imshow(X_test[0])
plt.show()
In [ ]:
예측한 결과가 리스트로 나온다. 리스트는 무엇을 나타내는가 ?¶
It's the probability that this item is each of the 10 classes
In [33]:
# ANN의 아웃풋의 노드(뉴런)수가 10개 이기 때문에,
# softmax 활성화 함수를 통해서, 0~1사이의 10개의 숫자가 나오게 되어있다.(확률)
In [33]:
첫번째 이미지의 실제값을 출력한다.
In [54]:
X_test[0]
Out[54]:
array([[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0.01176471,
0.00392157, 0. , 0. , 0.02745098, 0. ,
0.14509804, 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0.00392157, 0.00784314,
0. , 0.10588235, 0.32941176, 0.04313725, 0. ,
0. , 0. , 0. , 0. , 0. ,
0.46666667, 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0.00392157, 0. ,
0. , 0.34509804, 0.56078431, 0.43137255, 0. ,
0. , 0. , 0. , 0.08627451, 0.36470588,
0.41568627, 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0.01568627, 0. ,
0.20784314, 0.50588235, 0.47058824, 0.57647059, 0.68627451,
0.61568627, 0.65098039, 0.52941176, 0.60392157, 0.65882353,
0.54901961, 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0.00784314, 0. , 0.04313725,
0.5372549 , 0.50980392, 0.50196078, 0.62745098, 0.69019608,
0.62352941, 0.65490196, 0.69803922, 0.58431373, 0.59215686,
0.56470588, 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0.00392157, 0. , 0.00784314, 0.00392157,
0. , 0.01176471, 0. , 0. , 0.45098039,
0.44705882, 0.41568627, 0.5372549 , 0.65882353, 0.6 ,
0.61176471, 0.64705882, 0.65490196, 0.56078431, 0.61568627,
0.61960784, 0.04313725, 0. ],
[0. , 0. , 0. , 0. , 0.00392157,
0. , 0. , 0. , 0. , 0. ,
0.01176471, 0. , 0. , 0.34901961, 0.54509804,
0.35294118, 0.36862745, 0.6 , 0.58431373, 0.51372549,
0.59215686, 0.6627451 , 0.6745098 , 0.56078431, 0.62352941,
0.6627451 , 0.18823529, 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0.00784314, 0.01568627, 0.00392157, 0. ,
0. , 0. , 0.38431373, 0.53333333, 0.43137255,
0.42745098, 0.43137255, 0.63529412, 0.52941176, 0.56470588,
0.58431373, 0.62352941, 0.65490196, 0.56470588, 0.61960784,
0.6627451 , 0.46666667, 0. ],
[0. , 0. , 0.00784314, 0.00784314, 0.00392157,
0.00784314, 0. , 0. , 0. , 0. ,
0.10196078, 0.42352941, 0.45882353, 0.38823529, 0.43529412,
0.45882353, 0.53333333, 0.61176471, 0.5254902 , 0.60392157,
0.60392157, 0.61176471, 0.62745098, 0.55294118, 0.57647059,
0.61176471, 0.69803922, 0. ],
[0.01176471, 0. , 0. , 0. , 0. ,
0. , 0. , 0.08235294, 0.20784314, 0.36078431,
0.45882353, 0.43529412, 0.40392157, 0.45098039, 0.50588235,
0.5254902 , 0.56078431, 0.60392157, 0.64705882, 0.66666667,
0.60392157, 0.59215686, 0.60392157, 0.56078431, 0.54117647,
0.58823529, 0.64705882, 0.16862745],
[0. , 0. , 0.09019608, 0.21176471, 0.25490196,
0.29803922, 0.33333333, 0.4627451 , 0.50196078, 0.48235294,
0.43529412, 0.44313725, 0.4627451 , 0.49803922, 0.49019608,
0.54509804, 0.52156863, 0.53333333, 0.62745098, 0.54901961,
0.60784314, 0.63137255, 0.56470588, 0.60784314, 0.6745098 ,
0.63137255, 0.74117647, 0.24313725],
[0. , 0.26666667, 0.36862745, 0.35294118, 0.43529412,
0.44705882, 0.43529412, 0.44705882, 0.45098039, 0.49803922,
0.52941176, 0.53333333, 0.56078431, 0.49411765, 0.49803922,
0.59215686, 0.60392157, 0.56078431, 0.58039216, 0.49019608,
0.63529412, 0.63529412, 0.56470588, 0.54117647, 0.6 ,
0.63529412, 0.76862745, 0.22745098],
[0.2745098 , 0.6627451 , 0.50588235, 0.40784314, 0.38431373,
0.39215686, 0.36862745, 0.38039216, 0.38431373, 0.4 ,
0.42352941, 0.41568627, 0.46666667, 0.47058824, 0.50588235,
0.58431373, 0.61176471, 0.65490196, 0.74509804, 0.74509804,
0.76862745, 0.77647059, 0.77647059, 0.73333333, 0.77254902,
0.74117647, 0.72156863, 0.14117647],
[0.0627451 , 0.49411765, 0.67058824, 0.7372549 , 0.7372549 ,
0.72156863, 0.67058824, 0.6 , 0.52941176, 0.47058824,
0.49411765, 0.49803922, 0.57254902, 0.7254902 , 0.76470588,
0.81960784, 0.81568627, 1. , 0.81960784, 0.69411765,
0.96078431, 0.98823529, 0.98431373, 0.98431373, 0.96862745,
0.8627451 , 0.80784314, 0.19215686],
[0. , 0. , 0. , 0.04705882, 0.2627451 ,
0.41568627, 0.64313725, 0.7254902 , 0.78039216, 0.82352941,
0.82745098, 0.82352941, 0.81568627, 0.74509804, 0.58823529,
0.32156863, 0.03137255, 0. , 0. , 0. ,
0.69803922, 0.81568627, 0.7372549 , 0.68627451, 0.63529412,
0.61960784, 0.59215686, 0.04313725],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. ],
[0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. , 0. , 0. ,
0. , 0. , 0. ]])
In [33]:
In [33]:
In [33]:
이 예측결과가 앵클부츠인지는, 예측결과로 나온 리스트를 보고 어떻게 알 수 있는가?¶
10번째 값이 가장 크니까, 레이블은 0부터라, 9 이다.
In [33]:
# numpy 의 argmax() 함수를 이용하면, 가장 큰값이 들어있는 곳의 인덱스를 얻을 수 있다.
In [33]:
Exercise 2:¶
히든레이어를 1024개의 뉴런으로 바꿔서 실행해보자
In [55]:
def build_model() :
model = Sequential()
model.add( Flatten() )
model.add( Dense(1024, 'relu'))
model.add( Dense(64, 'relu'))
model.add( Dense(10, 'softmax'))
model.compile('adam', 'sparse_categorical_crossentropy', ['accuracy'])
return model
In [56]:
model = build_model()
In [57]:
model.fit(X_train, y_train, epochs = 5)
Epoch 1/5 1875/1875 [==============================] - 5s 2ms/step - loss: 0.4743 - accuracy: 0.8290 Epoch 2/5 1875/1875 [==============================] - 4s 2ms/step - loss: 0.3594 - accuracy: 0.8678 Epoch 3/5 1875/1875 [==============================] - 4s 2ms/step - loss: 0.3236 - accuracy: 0.8801 Epoch 4/5 1875/1875 [==============================] - 4s 2ms/step - loss: 0.2985 - accuracy: 0.8900 Epoch 5/5 1875/1875 [==============================] - 4s 2ms/step - loss: 0.2796 - accuracy: 0.8953
Out[57]:
<keras.callbacks.History at 0x7fe0286fb250>
In [58]:
model.evaluate(X_test,y_test) # 평가 : 위보다 2%정도 성능향상
313/313 [==============================] - 1s 2ms/step - loss: 0.3315 - accuracy: 0.8837
Out[58]:
[0.3315395712852478, 0.8837000131607056]
In [ ]:
In [ ]:
Question 1. 1024 개로 뉴런을 늘리면 어떤 일이 일어나는가?¶
In [33]:
# 위보다 2%정도 성능향상
In [33]:
In [33]:
Exercise 3:¶
Flatten() layer를 지우면 어떻게 되는가?
In [61]:
def build_model() :
model = Sequential()
model.add( Dense(128, 'relu'))
model.add( Dense(64, 'relu'))
model.add( Dense(10, 'softmax'))
model.compile('adam', 'sparse_categorical_crossentropy', ['accuracy'])
return model
In [62]:
model = build_model()
In [63]:
model.fit(X_train,y_train, epochs=5 ) # 2차원이 들어왓기 때문에 에러발생. 말이안된다는것
Epoch 1/5
--------------------------------------------------------------------------- ValueError Traceback (most recent call last) <ipython-input-63-f49c2b7e688a> in <module> ----> 1 model.fit(X_train,y_train, epochs=5 ) /usr/local/lib/python3.8/dist-packages/keras/utils/traceback_utils.py in error_handler(*args, **kwargs) 65 except Exception as e: # pylint: disable=broad-except 66 filtered_tb = _process_traceback_frames(e.__traceback__) ---> 67 raise e.with_traceback(filtered_tb) from None 68 finally: 69 del filtered_tb /usr/local/lib/python3.8/dist-packages/keras/engine/training.py in tf__train_function(iterator) 13 try: 14 do_return = True ---> 15 retval_ = ag__.converted_call(ag__.ld(step_function), (ag__.ld(self), ag__.ld(iterator)), None, fscope) 16 except: 17 do_return = False ValueError: in user code: File "/usr/local/lib/python3.8/dist-packages/keras/engine/training.py", line 1051, in train_function * return step_function(self, iterator) File "/usr/local/lib/python3.8/dist-packages/keras/engine/training.py", line 1040, in step_function ** outputs = model.distribute_strategy.run(run_step, args=(data,)) File "/usr/local/lib/python3.8/dist-packages/keras/engine/training.py", line 1030, in run_step ** outputs = model.train_step(data) File "/usr/local/lib/python3.8/dist-packages/keras/engine/training.py", line 890, in train_step loss = self.compute_loss(x, y, y_pred, sample_weight) File "/usr/local/lib/python3.8/dist-packages/keras/engine/training.py", line 948, in compute_loss return self.compiled_loss( File "/usr/local/lib/python3.8/dist-packages/keras/engine/compile_utils.py", line 201, in __call__ loss_value = loss_obj(y_t, y_p, sample_weight=sw) File "/usr/local/lib/python3.8/dist-packages/keras/losses.py", line 139, in __call__ losses = call_fn(y_true, y_pred) File "/usr/local/lib/python3.8/dist-packages/keras/losses.py", line 243, in call ** return ag_fn(y_true, y_pred, **self._fn_kwargs) File "/usr/local/lib/python3.8/dist-packages/keras/losses.py", line 1860, in sparse_categorical_crossentropy return backend.sparse_categorical_crossentropy( File "/usr/local/lib/python3.8/dist-packages/keras/backend.py", line 5238, in sparse_categorical_crossentropy res = tf.nn.sparse_softmax_cross_entropy_with_logits( ValueError: `labels.shape` must equal `logits.shape` except for the last dimension. Received: labels.shape=(32,) and logits.shape=(896, 10)
Exercise 4:¶
마지막 레이어를 10개가 아니라, 5개로 바꾸면 어떻게 되는가?
In [68]:
def build_model() :
model = Sequential()
model.add( Flatten() )
model.add( Dense(128, 'relu'))
model.add( Dense(64, 'relu'))
model.add( Dense(5, 'softmax'))
model.compile('adam', 'sparse_categorical_crossentropy', ['accuracy'])
return model
In [69]:
model =build_model()
In [70]:
model.fit(X_train,y_train, epochs=5 ) # 5~9 의 값이 나온다면 y_test의 값과 비교를 할수없기 때문에 에러발생
Epoch 1/5 1875/1875 [==============================] - 4s 2ms/step - loss: nan - accuracy: 0.1001 Epoch 2/5 1875/1875 [==============================] - 4s 2ms/step - loss: nan - accuracy: 0.1000 Epoch 3/5 1875/1875 [==============================] - 5s 3ms/step - loss: nan - accuracy: 0.1000 Epoch 4/5 1875/1875 [==============================] - 4s 2ms/step - loss: nan - accuracy: 0.1000 Epoch 5/5 1875/1875 [==============================] - 4s 2ms/step - loss: nan - accuracy: 0.1000
Out[70]:
<keras.callbacks.History at 0x7fe049cb7610>
In [33]:
In [33]:
In [33]:
In [33]:
In [33]:
In [71]:
def build_model() :
model = Sequential()
model.add( Flatten() )
model.add( Dense(128, 'relu'))
model.add( Dense(64, 'relu'))
model.add( Dense(10, 'softmax'))
model.compile('adam', 'sparse_categorical_crossentropy', ['accuracy'])
return model
In [72]:
model = build_model()
In [74]:
epoch_history = model.fit(X_train,y_train, epochs=30, validation_split = 0.2) # 일부분만 학습시키고 중간시험 (80%기 때문에 48000개는학습 12000개는 시험)
Epoch 1/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2174 - accuracy: 0.9190 - val_loss: 0.2308 - val_accuracy: 0.9146 Epoch 2/30 1500/1500 [==============================] - 5s 3ms/step - loss: 0.2106 - accuracy: 0.9203 - val_loss: 0.2606 - val_accuracy: 0.8978 Epoch 3/30 1500/1500 [==============================] - 7s 5ms/step - loss: 0.2041 - accuracy: 0.9221 - val_loss: 0.2371 - val_accuracy: 0.9107 Epoch 4/30 1500/1500 [==============================] - 6s 4ms/step - loss: 0.1987 - accuracy: 0.9244 - val_loss: 0.2331 - val_accuracy: 0.9142 Epoch 5/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1894 - accuracy: 0.9280 - val_loss: 0.2561 - val_accuracy: 0.9073 Epoch 6/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1835 - accuracy: 0.9307 - val_loss: 0.2428 - val_accuracy: 0.9109 Epoch 7/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1829 - accuracy: 0.9315 - val_loss: 0.2556 - val_accuracy: 0.9059 Epoch 8/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1752 - accuracy: 0.9335 - val_loss: 0.2353 - val_accuracy: 0.9159 Epoch 9/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1698 - accuracy: 0.9353 - val_loss: 0.2632 - val_accuracy: 0.9082 Epoch 10/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1643 - accuracy: 0.9384 - val_loss: 0.2499 - val_accuracy: 0.9122 Epoch 11/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1606 - accuracy: 0.9391 - val_loss: 0.2523 - val_accuracy: 0.9128 Epoch 12/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1565 - accuracy: 0.9412 - val_loss: 0.2662 - val_accuracy: 0.9099 Epoch 13/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1531 - accuracy: 0.9408 - val_loss: 0.2803 - val_accuracy: 0.9087 Epoch 14/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1478 - accuracy: 0.9439 - val_loss: 0.2714 - val_accuracy: 0.9064 Epoch 15/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1467 - accuracy: 0.9447 - val_loss: 0.2904 - val_accuracy: 0.9057 Epoch 16/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1400 - accuracy: 0.9467 - val_loss: 0.2858 - val_accuracy: 0.9071 Epoch 17/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1369 - accuracy: 0.9485 - val_loss: 0.2971 - val_accuracy: 0.9045 Epoch 18/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1343 - accuracy: 0.9496 - val_loss: 0.3031 - val_accuracy: 0.9081 Epoch 19/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1299 - accuracy: 0.9511 - val_loss: 0.3267 - val_accuracy: 0.9048 Epoch 20/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1286 - accuracy: 0.9514 - val_loss: 0.3583 - val_accuracy: 0.8987 Epoch 21/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1249 - accuracy: 0.9519 - val_loss: 0.3172 - val_accuracy: 0.9068 Epoch 22/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1202 - accuracy: 0.9543 - val_loss: 0.3193 - val_accuracy: 0.9093 Epoch 23/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1209 - accuracy: 0.9535 - val_loss: 0.3467 - val_accuracy: 0.9032 Epoch 24/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1162 - accuracy: 0.9556 - val_loss: 0.3555 - val_accuracy: 0.9007 Epoch 25/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1174 - accuracy: 0.9558 - val_loss: 0.3517 - val_accuracy: 0.9042 Epoch 26/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1138 - accuracy: 0.9575 - val_loss: 0.3477 - val_accuracy: 0.9020 Epoch 27/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1118 - accuracy: 0.9570 - val_loss: 0.3639 - val_accuracy: 0.9029 Epoch 28/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1072 - accuracy: 0.9590 - val_loss: 0.3413 - val_accuracy: 0.9029 Epoch 29/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1049 - accuracy: 0.9596 - val_loss: 0.3607 - val_accuracy: 0.9077 Epoch 30/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.1057 - accuracy: 0.9604 - val_loss: 0.3640 - val_accuracy: 0.9019
In [78]:
epoch_history.history.keys()
Out[78]:
dict_keys(['loss', 'accuracy', 'val_loss', 'val_accuracy'])
In [79]:
plt.plot(epoch_history.history['loss'] ) # 학습로스
plt.plot(epoch_history.history['val_loss'] ) # 시험로스
plt.legend( ['Train Loss', 'Vaildation Loss']) # 그냥보면 헷갈리니까 레전드로본다.
plt.show()
In [ ]:
# 밸리데이션 로스가 가장낮을때가 적당한 에포크이다.
# 위의 그래프를 보면 밸리대이션 로스가 점점 커진다
# 오버피팅되었기 때문 : 에포크가 지나쳐서 잘못맞추는것
In [80]:
plt.plot(epoch_history.history['accuracy'] ) # 학습로스
plt.plot(epoch_history.history['val_accuracy'] ) # 시험로스
plt.legend( ['Train accuracy', 'Vaildation accuracy']) # 그냥보면 헷갈리니까 레전드로본다.
plt.show()
In [ ]:
# 정확도가 오히려 떨어지고 있다 (이것도 오버피팅의 증거)
In [ ]:
30 에포크를 했을때의 결과를 보면 어떠한 문제가 있다. 이러한 문제를 무엇이라고 하는가???
In [33]:
# 오버 피팅 이라고 한다. Over Fitting
In [33]:
# 반대로 에포크가 너무 적으면 언더피팅 이라고 한다. Under Fitting
Exercise 7:¶
accuracy 나 loss 가, 우리가 원하는 특정 값이 되면 자동으로 학습을 멈추가 하고 싶다. 즉, 특정값에 도달하면 학습을 멈추게 할 수 있는 콜백기능을 제공한다. callbacks
In [33]:
# 04 강의에서 본것은, early stopping : 특정횟수동안 성능개선이 없으면 멈춘다였고.
In [33]:
# val_accuracy 가 88% 가 넘으면 멈추도록 하고싶다. (오버피팅,언더피팅을 안하려고)
In [81]:
class myCallback(tf.keras.callbacks.Callback) :
def on_epoch_end(self, epoch, logs= {} ) :
if logs[ 'val_accuracy'] > 0.88 :
print('\n내가 정한 정확도에 도달했으니, 학습을 멈춘다.')
self.model.stop_training = True
In [82]:
my_cb = myCallback()
In [83]:
def build_model() :
model = Sequential()
model.add( Flatten() )
model.add( Dense(128, 'relu'))
model.add( Dense(64, 'relu'))
model.add( Dense(10, 'softmax'))
model.compile('adam', 'sparse_categorical_crossentropy', ['accuracy'])
return model
In [84]:
model = build_model()
In [85]:
epoch_history = model.fit(X_train,y_train, epochs=30, validation_split = 0.2, callbacks = [my_cb] )
Epoch 1/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.5127 - accuracy: 0.8176 - val_loss: 0.3990 - val_accuracy: 0.8532 Epoch 2/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.3767 - accuracy: 0.8616 - val_loss: 0.3491 - val_accuracy: 0.8740 Epoch 3/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.3370 - accuracy: 0.8749 - val_loss: 0.3744 - val_accuracy: 0.8704 Epoch 4/30 1500/1500 [==============================] - 4s 3ms/step - loss: 0.3143 - accuracy: 0.8835 - val_loss: 0.3543 - val_accuracy: 0.8730 Epoch 5/30 1478/1500 [============================>.] - ETA: 0s - loss: 0.2961 - accuracy: 0.8898 내가 정한 정확도에 도달했으니, 학습을 멈춘다. 1500/1500 [==============================] - 4s 3ms/step - loss: 0.2956 - accuracy: 0.8900 - val_loss: 0.3260 - val_accuracy: 0.8833
In [86]:
model.evaluate(X_test, y_test)
313/313 [==============================] - 1s 2ms/step - loss: 0.3549 - accuracy: 0.8764
Out[86]:
[0.3549431264400482, 0.8763999938964844]