728x90
반응형
시바견과 진돗개를 분류해보자!
+사람도 가끔 헷갈리는
+사람도 가끔 헷갈리는
모듈 로드
In [23]:import cv2
import matplotlib.pyplot as plt
import tensorflow as tf
import numpy as np
import os
import tqdm
import random
from sklearn.datasets import load_files
In [23]:
import cv2
import matplotlib.pyplot as plt
import tensorflow as tf
import numpy as np
import os
import tqdm
import random
from sklearn.datasets import load_files
데이터 로드
In [24]:X_dog = list()
for fileName in os.listdir('./shiba'):
if fileName.startswith('google'):
X_dog.append(fileName)
In [25]:len(X_dog)
Out[25]:232
In [26]:for fileName in os.listdir('./진돗개'):
X_dog.append(fileName)
In [27]:len(X_dog)
Out[27]:476
시바견 : 232 / 진돗개 : 244
In [28]:n_shiba = 232
n_jindo = 244
In [24]:
X_dog = list()
for fileName in os.listdir('./shiba'):
if fileName.startswith('google'):
X_dog.append(fileName)
In [25]:
len(X_dog)
Out[25]:
232
In [26]:
for fileName in os.listdir('./진돗개'):
X_dog.append(fileName)
In [27]:
len(X_dog)
Out[27]:
476
시바견 : 232 / 진돗개 : 244
In [28]:
n_shiba = 232
n_jindo = 244
Category 만들기
시바견이면 [1,0]
진돗개면 [0,1]
In [29]:shiba_labels = list(list([1,0] for _ in range(n_shiba)))
jindo_labels = list(list([0,1] for _ in range(n_jindo)))
In [30]:labels = shiba_labels + jindo_labels
In [31]:labels = np.array(labels).astype('int')
In [32]:len(labels),labels[:3]
Out[32]:(476, array([[1, 0],
[1, 0],
[1, 0]]))
시바견이면 [1,0]
진돗개면 [0,1]
In [29]:
shiba_labels = list(list([1,0] for _ in range(n_shiba)))
jindo_labels = list(list([0,1] for _ in range(n_jindo)))
In [30]:
labels = shiba_labels + jindo_labels
In [31]:
labels = np.array(labels).astype('int')
In [32]:
len(labels),labels[:3]
Out[32]:
(476, array([[1, 0],
[1, 0],
[1, 0]]))image_resize
시바견과 진돗개의 폴더 위치가 달라서 각가 resize 후 합쳐준다.
Color가 아닌 Gray-scale을 이용해서 분석하였다.
(stanford dog breed classification도 gray-scale이길래..)
In [33]:resize_dog = list()
for dog in X_dog[:n_shiba]:
img = cv2.imread('./shiba/' + dog,cv2.IMREAD_GRAYSCALE)
resize = cv2.resize(img,(224,224))
resize_dog.append(resize)
In [34]:len(resize_dog)
Out[34]:232
In [35]:resize_dog[0].shape
Out[35]:(224, 224)
In [36]:for i in X_dog[n_shiba:]:
img = cv2.imread('./진돗개/'+i,cv2.IMREAD_GRAYSCALE)
resize = cv2.resize(img,(224,224))
resize_dog.append(resize)
In [37]:len(resize_dog)
Out[37]:476
In [38]:# 잘 됐는지 임의로 하나 출력해보자.
plt.imshow(resize_dog[2],cmap=plt.cm.gray)
plt.show()
시바견과 진돗개의 폴더 위치가 달라서 각가 resize 후 합쳐준다.
Color가 아닌 Gray-scale을 이용해서 분석하였다.
(stanford dog breed classification도 gray-scale이길래..)
In [33]:
resize_dog = list()
for dog in X_dog[:n_shiba]:
img = cv2.imread('./shiba/' + dog,cv2.IMREAD_GRAYSCALE)
resize = cv2.resize(img,(224,224))
resize_dog.append(resize)
In [34]:
len(resize_dog)
Out[34]:
232
In [35]:
resize_dog[0].shape
Out[35]:
(224, 224)
In [36]:
for i in X_dog[n_shiba:]:
img = cv2.imread('./진돗개/'+i,cv2.IMREAD_GRAYSCALE)
resize = cv2.resize(img,(224,224))
resize_dog.append(resize)
In [37]:
len(resize_dog)
Out[37]:
476
In [38]:
# 잘 됐는지 임의로 하나 출력해보자.
plt.imshow(resize_dog[2],cmap=plt.cm.gray)
plt.show()
Train / Test 분할
그전에 먼저 데이터 Shuffle 해주기
In [44]:# data shuffle
np.random.seed(42)
tmp = [[x,y] for x, y in zip(resize_dog, labels)]
random.shuffle(tmp)
X_sample = [n[0] for n in tmp]
y_sample = [n[1] for n in tmp]
In [45]:# Train / Test Split
# 80 : 20
train_size = np.ceil(0.8 * len(resize_dog)).astype(int) # 381 / 나머지 95개 test 할당
X_train = X_sample[:train_size]
y_train = y_sample[:train_size]
X_test = X_sample[train_size:]
y_test = y_sample[train_size:]
In [46]:X_train = np.array(X_train)
y_train = np.array(y_train)
X_test = np.array(X_test)
y_test = np.array(y_test)
In [47]:len(X_train), len(y_train), len(X_test), len(y_test)
Out[47]:(381, 381, 95, 95)
In [48]:plt.imshow(X_test[1],cmap=plt.cm.gray)
print(y_test[1])
plt.show()
[1 0]
그전에 먼저 데이터 Shuffle 해주기
In [44]:
# data shuffle
np.random.seed(42)
tmp = [[x,y] for x, y in zip(resize_dog, labels)]
random.shuffle(tmp)
X_sample = [n[0] for n in tmp]
y_sample = [n[1] for n in tmp]
In [45]:
# Train / Test Split
# 80 : 20
train_size = np.ceil(0.8 * len(resize_dog)).astype(int) # 381 / 나머지 95개 test 할당
X_train = X_sample[:train_size]
y_train = y_sample[:train_size]
X_test = X_sample[train_size:]
y_test = y_sample[train_size:]
In [46]:
X_train = np.array(X_train)
y_train = np.array(y_train)
X_test = np.array(X_test)
y_test = np.array(y_test)
In [47]:
len(X_train), len(y_train), len(X_test), len(y_test)
Out[47]:
(381, 381, 95, 95)
In [48]:
plt.imshow(X_test[1],cmap=plt.cm.gray)
print(y_test[1])
plt.show()
[1 0]
Network 구축하기
by Keras
In [49]:IMG_SIZE = 224
In [50]:# (None,224,224,1) 형태로 reshape
X_train = X_train.reshape(X_train.shape[0],IMG_SIZE,IMG_SIZE,1)
X_test = X_test.reshape(X_test.shape[0],IMG_SIZE,IMG_SIZE,1)
In [51]:from keras import models
from keras.layers import Conv2D, MaxPooling2D,BatchNormalization,Dropout,Flatten,Dense
Using TensorFlow backend.
In [52]:IMG_SIZE = 224
def Network(name):
name.add(Conv2D(32, kernel_size = (3, 3), activation='relu', input_shape=(IMG_SIZE, IMG_SIZE, 1)))
name.add(MaxPooling2D(pool_size=(2,2)))
name.add(BatchNormalization())
name.add(Conv2D(64, kernel_size=(3,3), activation='relu'))
name.add(MaxPooling2D(pool_size=(2,2)))
name.add(BatchNormalization())
name.add(Conv2D(64, kernel_size=(3,3), activation='relu'))
name.add(MaxPooling2D(pool_size=(2,2)))
name.add(BatchNormalization())
name.add(Conv2D(96, kernel_size=(3,3), activation='relu'))
name.add(MaxPooling2D(pool_size=(2,2)))
name.add(BatchNormalization())
name.add(Conv2D(32, kernel_size=(3,3), activation='relu'))
name.add(MaxPooling2D(pool_size=(2,2)))
name.add(BatchNormalization())
name.add(Dropout(0.2))
name.add(Flatten())
name.add(Dense(128, activation='relu'))
name.add(Dropout(0.3))
name.add(Dense(2, activation = 'softmax'))
In [53]:model1 = models.Sequential()
model2 = models.Sequential()
Network(model1)
Network(model2)
WARNING:tensorflow:From /Users/charming/anaconda3/lib/python3.7/site-packages/tensorflow/python/framework/op_def_library.py:263: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version.
Instructions for updating:
Colocations handled automatically by placer.
WARNING:tensorflow:From /Users/charming/anaconda3/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py:3445: calling dropout (from tensorflow.python.ops.nn_ops) with keep_prob is deprecated and will be removed in a future version.
Instructions for updating:
Please use `rate` instead of `keep_prob`. Rate should be set to `rate = 1 - keep_prob`.
In [54]:model1.compile(optimizer='rmsprop',loss='categorical_crossentropy',metrics=['accuracy'])
In [55]:model2.compile(optimizer='Adam',loss='categorical_crossentropy',metrics=['accuracy'])
In [56]:X_train.shape
Out[56]:(381, 224, 224, 1)
by Keras
In [49]:
IMG_SIZE = 224
In [50]:
# (None,224,224,1) 형태로 reshape
X_train = X_train.reshape(X_train.shape[0],IMG_SIZE,IMG_SIZE,1)
X_test = X_test.reshape(X_test.shape[0],IMG_SIZE,IMG_SIZE,1)
In [51]:
from keras import models
from keras.layers import Conv2D, MaxPooling2D,BatchNormalization,Dropout,Flatten,Dense
Using TensorFlow backend.
In [52]:
IMG_SIZE = 224
def Network(name):
name.add(Conv2D(32, kernel_size = (3, 3), activation='relu', input_shape=(IMG_SIZE, IMG_SIZE, 1)))
name.add(MaxPooling2D(pool_size=(2,2)))
name.add(BatchNormalization())
name.add(Conv2D(64, kernel_size=(3,3), activation='relu'))
name.add(MaxPooling2D(pool_size=(2,2)))
name.add(BatchNormalization())
name.add(Conv2D(64, kernel_size=(3,3), activation='relu'))
name.add(MaxPooling2D(pool_size=(2,2)))
name.add(BatchNormalization())
name.add(Conv2D(96, kernel_size=(3,3), activation='relu'))
name.add(MaxPooling2D(pool_size=(2,2)))
name.add(BatchNormalization())
name.add(Conv2D(32, kernel_size=(3,3), activation='relu'))
name.add(MaxPooling2D(pool_size=(2,2)))
name.add(BatchNormalization())
name.add(Dropout(0.2))
name.add(Flatten())
name.add(Dense(128, activation='relu'))
name.add(Dropout(0.3))
name.add(Dense(2, activation = 'softmax'))
In [53]:
model1 = models.Sequential()
model2 = models.Sequential()
Network(model1)
Network(model2)
WARNING:tensorflow:From /Users/charming/anaconda3/lib/python3.7/site-packages/tensorflow/python/framework/op_def_library.py:263: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version. Instructions for updating: Colocations handled automatically by placer. WARNING:tensorflow:From /Users/charming/anaconda3/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py:3445: calling dropout (from tensorflow.python.ops.nn_ops) with keep_prob is deprecated and will be removed in a future version. Instructions for updating: Please use `rate` instead of `keep_prob`. Rate should be set to `rate = 1 - keep_prob`.
In [54]:
model1.compile(optimizer='rmsprop',loss='categorical_crossentropy',metrics=['accuracy'])
In [55]:
model2.compile(optimizer='Adam',loss='categorical_crossentropy',metrics=['accuracy'])
In [56]:
X_train.shape
Out[56]:
(381, 224, 224, 1)
RmsProp
In [237]:model1.fit(np.array(X_train),np.array(y_train),epochs=15,batch_size=20,verbose=1)
Epoch 1/15
381/381 [==============================] - 32s 83ms/step - loss: 1.0687 - acc: 0.5722
Epoch 2/15
381/381 [==============================] - 28s 72ms/step - loss: 0.7263 - acc: 0.6877
Epoch 3/15
381/381 [==============================] - 28s 73ms/step - loss: 0.6903 - acc: 0.7349
Epoch 4/15
381/381 [==============================] - 28s 73ms/step - loss: 0.5124 - acc: 0.7848
Epoch 5/15
381/381 [==============================] - 28s 74ms/step - loss: 0.4785 - acc: 0.7664
Epoch 6/15
381/381 [==============================] - 28s 73ms/step - loss: 0.3344 - acc: 0.8451
Epoch 7/15
381/381 [==============================] - 28s 72ms/step - loss: 0.3802 - acc: 0.8346
Epoch 8/15
381/381 [==============================] - 28s 73ms/step - loss: 0.3330 - acc: 0.8688
Epoch 9/15
381/381 [==============================] - 28s 73ms/step - loss: 0.2578 - acc: 0.8898
Epoch 10/15
381/381 [==============================] - 28s 73ms/step - loss: 0.2201 - acc: 0.9055
Epoch 11/15
381/381 [==============================] - 27s 72ms/step - loss: 0.2843 - acc: 0.8740
Epoch 12/15
381/381 [==============================] - 28s 73ms/step - loss: 0.1711 - acc: 0.9344
Epoch 13/15
381/381 [==============================] - 27s 72ms/step - loss: 0.1568 - acc: 0.9423
Epoch 14/15
381/381 [==============================] - 29s 75ms/step - loss: 0.1747 - acc: 0.9344
Epoch 15/15
381/381 [==============================] - 28s 73ms/step - loss: 0.1116 - acc: 0.9554
Out[237]:<keras.callbacks.History at 0x14ff520f0>
In [239]:# optimizer : RmsProp / epochs : 15 / batch_size = 20
# network는 동일
loss, acc = model1.evaluate(X_test,y_test,verbose=0)
print("Accuracy : %0.2f" % (acc*100))
Accuracy : 67.37
In [251]:model2.fit(np.array(X_train),np.array(y_train),epochs=15,batch_size=20,verbose=1)
Epoch 1/15
381/381 [==============================] - 27s 72ms/step - loss: 0.5673 - acc: 0.7743
Epoch 2/15
381/381 [==============================] - 28s 72ms/step - loss: 0.3224 - acc: 0.8845
Epoch 3/15
381/381 [==============================] - 27s 71ms/step - loss: 0.3590 - acc: 0.8504
Epoch 4/15
381/381 [==============================] - 28s 74ms/step - loss: 0.3302 - acc: 0.8661
Epoch 5/15
381/381 [==============================] - 35s 91ms/step - loss: 0.2623 - acc: 0.8740
Epoch 6/15
381/381 [==============================] - 28s 73ms/step - loss: 0.1526 - acc: 0.9423
Epoch 7/15
381/381 [==============================] - 28s 75ms/step - loss: 0.2321 - acc: 0.9213
Epoch 8/15
381/381 [==============================] - 28s 73ms/step - loss: 0.2306 - acc: 0.9055
Epoch 9/15
381/381 [==============================] - 28s 73ms/step - loss: 0.2819 - acc: 0.8766
Epoch 10/15
381/381 [==============================] - 27s 72ms/step - loss: 0.1660 - acc: 0.9265
Epoch 11/15
381/381 [==============================] - 28s 73ms/step - loss: 0.0973 - acc: 0.9606
Epoch 12/15
381/381 [==============================] - 28s 73ms/step - loss: 0.1545 - acc: 0.9370
Epoch 13/15
381/381 [==============================] - 27s 72ms/step - loss: 0.1528 - acc: 0.9396
Epoch 14/15
381/381 [==============================] - 28s 73ms/step - loss: 0.1079 - acc: 0.9738
Epoch 15/15
381/381 [==============================] - 28s 72ms/step - loss: 0.0835 - acc: 0.9738
Out[251]:<keras.callbacks.History at 0x14dd13358>
In [252]:# optimizer : Adam / Dropout = 0.3 추가 / epochs : 15 / batch_size = 20
# network는 동일
loss, acc = model2.evaluate(X_test,y_test,verbose=0)
print("Accuracy : %0.2f" % (acc*100))
Accuracy : 69.47
Trainin Acc는 계속 높게 나오지만 Test Acc는 Epoch, batch_size등등에 따라 계속 변화하며 낮은 수치를 보인다.
Training Acc는 높고 Test Acc가 낮다는 것은 이는 Overfitting을 의미한다.
이를 해결하기 위해 다른 방법들을 적용해 볼 것이다.
In [237]:
model1.fit(np.array(X_train),np.array(y_train),epochs=15,batch_size=20,verbose=1)
Epoch 1/15 381/381 [==============================] - 32s 83ms/step - loss: 1.0687 - acc: 0.5722 Epoch 2/15 381/381 [==============================] - 28s 72ms/step - loss: 0.7263 - acc: 0.6877 Epoch 3/15 381/381 [==============================] - 28s 73ms/step - loss: 0.6903 - acc: 0.7349 Epoch 4/15 381/381 [==============================] - 28s 73ms/step - loss: 0.5124 - acc: 0.7848 Epoch 5/15 381/381 [==============================] - 28s 74ms/step - loss: 0.4785 - acc: 0.7664 Epoch 6/15 381/381 [==============================] - 28s 73ms/step - loss: 0.3344 - acc: 0.8451 Epoch 7/15 381/381 [==============================] - 28s 72ms/step - loss: 0.3802 - acc: 0.8346 Epoch 8/15 381/381 [==============================] - 28s 73ms/step - loss: 0.3330 - acc: 0.8688 Epoch 9/15 381/381 [==============================] - 28s 73ms/step - loss: 0.2578 - acc: 0.8898 Epoch 10/15 381/381 [==============================] - 28s 73ms/step - loss: 0.2201 - acc: 0.9055 Epoch 11/15 381/381 [==============================] - 27s 72ms/step - loss: 0.2843 - acc: 0.8740 Epoch 12/15 381/381 [==============================] - 28s 73ms/step - loss: 0.1711 - acc: 0.9344 Epoch 13/15 381/381 [==============================] - 27s 72ms/step - loss: 0.1568 - acc: 0.9423 Epoch 14/15 381/381 [==============================] - 29s 75ms/step - loss: 0.1747 - acc: 0.9344 Epoch 15/15 381/381 [==============================] - 28s 73ms/step - loss: 0.1116 - acc: 0.9554
Out[237]:
<keras.callbacks.History at 0x14ff520f0>
In [239]:
# optimizer : RmsProp / epochs : 15 / batch_size = 20
# network는 동일
loss, acc = model1.evaluate(X_test,y_test,verbose=0)
print("Accuracy : %0.2f" % (acc*100))
Accuracy : 67.37
In [251]:
model2.fit(np.array(X_train),np.array(y_train),epochs=15,batch_size=20,verbose=1)
Epoch 1/15 381/381 [==============================] - 27s 72ms/step - loss: 0.5673 - acc: 0.7743 Epoch 2/15 381/381 [==============================] - 28s 72ms/step - loss: 0.3224 - acc: 0.8845 Epoch 3/15 381/381 [==============================] - 27s 71ms/step - loss: 0.3590 - acc: 0.8504 Epoch 4/15 381/381 [==============================] - 28s 74ms/step - loss: 0.3302 - acc: 0.8661 Epoch 5/15 381/381 [==============================] - 35s 91ms/step - loss: 0.2623 - acc: 0.8740 Epoch 6/15 381/381 [==============================] - 28s 73ms/step - loss: 0.1526 - acc: 0.9423 Epoch 7/15 381/381 [==============================] - 28s 75ms/step - loss: 0.2321 - acc: 0.9213 Epoch 8/15 381/381 [==============================] - 28s 73ms/step - loss: 0.2306 - acc: 0.9055 Epoch 9/15 381/381 [==============================] - 28s 73ms/step - loss: 0.2819 - acc: 0.8766 Epoch 10/15 381/381 [==============================] - 27s 72ms/step - loss: 0.1660 - acc: 0.9265 Epoch 11/15 381/381 [==============================] - 28s 73ms/step - loss: 0.0973 - acc: 0.9606 Epoch 12/15 381/381 [==============================] - 28s 73ms/step - loss: 0.1545 - acc: 0.9370 Epoch 13/15 381/381 [==============================] - 27s 72ms/step - loss: 0.1528 - acc: 0.9396 Epoch 14/15 381/381 [==============================] - 28s 73ms/step - loss: 0.1079 - acc: 0.9738 Epoch 15/15 381/381 [==============================] - 28s 72ms/step - loss: 0.0835 - acc: 0.9738
Out[251]:
<keras.callbacks.History at 0x14dd13358>
In [252]:
# optimizer : Adam / Dropout = 0.3 추가 / epochs : 15 / batch_size = 20
# network는 동일
loss, acc = model2.evaluate(X_test,y_test,verbose=0)
print("Accuracy : %0.2f" % (acc*100))
Accuracy : 69.47
Trainin Acc는 계속 높게 나오지만 Test Acc는 Epoch, batch_size등등에 따라 계속 변화하며 낮은 수치를 보인다.
Training Acc는 높고 Test Acc가 낮다는 것은 이는 Overfitting을 의미한다.
이를 해결하기 위해 다른 방법들을 적용해 볼 것이다.
1. Validation 이용
dataset 자체가 작기 때문에 Validation을 시도하기는 어렵다.
dataset 자체가 작기 때문에 Validation을 시도하기는 어렵다.
2. Optimizer 바꾸기
기존에는 RmsProp를 이용했는데 Adam으로 바꿨을 때,67% -> 69%로 조금 낮아졌다.
하지만 초기 파라미터가 랜덤하기 때문에 정확도가 무작위성으로 바뀐다.
In [305]:model2.fit(np.array(X_train),np.array(y_train),epochs=15,batch_size=20,verbose=1,validation_split=0.1)
Train on 342 samples, validate on 39 samples
Epoch 1/15
342/342 [==============================] - 36s 106ms/step - loss: 1.0194 - acc: 0.6023 - val_loss: 1.1626 - val_acc: 0.5897
Epoch 2/15
342/342 [==============================] - 35s 104ms/step - loss: 0.6595 - acc: 0.6959 - val_loss: 1.2722 - val_acc: 0.4872
Epoch 3/15
342/342 [==============================] - 38s 111ms/step - loss: 0.5092 - acc: 0.7661 - val_loss: 0.8841 - val_acc: 0.7179
Epoch 4/15
342/342 [==============================] - 35s 103ms/step - loss: 0.3492 - acc: 0.8304 - val_loss: 0.9681 - val_acc: 0.6154
Epoch 5/15
342/342 [==============================] - 35s 102ms/step - loss: 0.3792 - acc: 0.8538 - val_loss: 1.2840 - val_acc: 0.5128
Epoch 6/15
342/342 [==============================] - 38s 112ms/step - loss: 0.2903 - acc: 0.8772 - val_loss: 0.8751 - val_acc: 0.6154
Epoch 7/15
342/342 [==============================] - 35s 101ms/step - loss: 0.2455 - acc: 0.9123 - val_loss: 0.7071 - val_acc: 0.7436
Epoch 8/15
342/342 [==============================] - 32s 94ms/step - loss: 0.3172 - acc: 0.8684 - val_loss: 1.0034 - val_acc: 0.5641
Epoch 9/15
342/342 [==============================] - 32s 93ms/step - loss: 0.1891 - acc: 0.9386 - val_loss: 1.6376 - val_acc: 0.4872
Epoch 10/15
342/342 [==============================] - 35s 102ms/step - loss: 0.1186 - acc: 0.9737 - val_loss: 1.5700 - val_acc: 0.5128
Epoch 11/15
342/342 [==============================] - 35s 103ms/step - loss: 0.1189 - acc: 0.9503 - val_loss: 0.9379 - val_acc: 0.6154
Epoch 12/15
342/342 [==============================] - 39s 115ms/step - loss: 0.1121 - acc: 0.9591 - val_loss: 1.4268 - val_acc: 0.6154
Epoch 13/15
342/342 [==============================] - 37s 107ms/step - loss: 0.0871 - acc: 0.9678 - val_loss: 3.4008 - val_acc: 0.4615
Epoch 14/15
342/342 [==============================] - 35s 102ms/step - loss: 0.0789 - acc: 0.9678 - val_loss: 1.0088 - val_acc: 0.5897
Epoch 15/15
342/342 [==============================] - 35s 102ms/step - loss: 0.0918 - acc: 0.9708 - val_loss: 1.4803 - val_acc: 0.6154
Out[305]:<keras.callbacks.History at 0x14ab858d0>
In [306]:# optimizer : Adam / Dropout = 0.3 추가 / epochs : 15 / batch_size = 20
# Validation 추가
loss, acc = model2.evaluate(X_test,y_test,verbose=0)
print("Accuracy : %0.2f" % (acc*100))
Accuracy : 70.53
기존에는 RmsProp를 이용했는데 Adam으로 바꿨을 때,67% -> 69%로 조금 낮아졌다.
하지만 초기 파라미터가 랜덤하기 때문에 정확도가 무작위성으로 바뀐다.
In [305]:
model2.fit(np.array(X_train),np.array(y_train),epochs=15,batch_size=20,verbose=1,validation_split=0.1)
Train on 342 samples, validate on 39 samples Epoch 1/15 342/342 [==============================] - 36s 106ms/step - loss: 1.0194 - acc: 0.6023 - val_loss: 1.1626 - val_acc: 0.5897 Epoch 2/15 342/342 [==============================] - 35s 104ms/step - loss: 0.6595 - acc: 0.6959 - val_loss: 1.2722 - val_acc: 0.4872 Epoch 3/15 342/342 [==============================] - 38s 111ms/step - loss: 0.5092 - acc: 0.7661 - val_loss: 0.8841 - val_acc: 0.7179 Epoch 4/15 342/342 [==============================] - 35s 103ms/step - loss: 0.3492 - acc: 0.8304 - val_loss: 0.9681 - val_acc: 0.6154 Epoch 5/15 342/342 [==============================] - 35s 102ms/step - loss: 0.3792 - acc: 0.8538 - val_loss: 1.2840 - val_acc: 0.5128 Epoch 6/15 342/342 [==============================] - 38s 112ms/step - loss: 0.2903 - acc: 0.8772 - val_loss: 0.8751 - val_acc: 0.6154 Epoch 7/15 342/342 [==============================] - 35s 101ms/step - loss: 0.2455 - acc: 0.9123 - val_loss: 0.7071 - val_acc: 0.7436 Epoch 8/15 342/342 [==============================] - 32s 94ms/step - loss: 0.3172 - acc: 0.8684 - val_loss: 1.0034 - val_acc: 0.5641 Epoch 9/15 342/342 [==============================] - 32s 93ms/step - loss: 0.1891 - acc: 0.9386 - val_loss: 1.6376 - val_acc: 0.4872 Epoch 10/15 342/342 [==============================] - 35s 102ms/step - loss: 0.1186 - acc: 0.9737 - val_loss: 1.5700 - val_acc: 0.5128 Epoch 11/15 342/342 [==============================] - 35s 103ms/step - loss: 0.1189 - acc: 0.9503 - val_loss: 0.9379 - val_acc: 0.6154 Epoch 12/15 342/342 [==============================] - 39s 115ms/step - loss: 0.1121 - acc: 0.9591 - val_loss: 1.4268 - val_acc: 0.6154 Epoch 13/15 342/342 [==============================] - 37s 107ms/step - loss: 0.0871 - acc: 0.9678 - val_loss: 3.4008 - val_acc: 0.4615 Epoch 14/15 342/342 [==============================] - 35s 102ms/step - loss: 0.0789 - acc: 0.9678 - val_loss: 1.0088 - val_acc: 0.5897 Epoch 15/15 342/342 [==============================] - 35s 102ms/step - loss: 0.0918 - acc: 0.9708 - val_loss: 1.4803 - val_acc: 0.6154
Out[305]:
<keras.callbacks.History at 0x14ab858d0>
In [306]:
# optimizer : Adam / Dropout = 0.3 추가 / epochs : 15 / batch_size = 20
# Validation 추가
loss, acc = model2.evaluate(X_test,y_test,verbose=0)
print("Accuracy : %0.2f" % (acc*100))
Accuracy : 70.53
Epoch 늘리기
In [307]:model2.fit(np.array(X_train),np.array(y_train),epochs=30,batch_size=20,verbose=1,validation_split=0.1)
Train on 342 samples, validate on 39 samples
Epoch 1/30
342/342 [==============================] - 31s 92ms/step - loss: 0.1395 - acc: 0.9591 - val_loss: 1.2031 - val_acc: 0.6154
Epoch 2/30
342/342 [==============================] - 31s 90ms/step - loss: 0.0693 - acc: 0.9708 - val_loss: 1.0359 - val_acc: 0.6923
Epoch 3/30
342/342 [==============================] - 30s 88ms/step - loss: 0.0739 - acc: 0.9737 - val_loss: 4.0959 - val_acc: 0.4615
Epoch 4/30
342/342 [==============================] - 29s 84ms/step - loss: 0.0737 - acc: 0.9766 - val_loss: 1.6908 - val_acc: 0.5385
Epoch 5/30
342/342 [==============================] - 31s 91ms/step - loss: 0.0370 - acc: 0.9883 - val_loss: 0.7877 - val_acc: 0.6410
Epoch 6/30
342/342 [==============================] - 31s 91ms/step - loss: 0.0274 - acc: 0.9912 - val_loss: 1.0398 - val_acc: 0.6154
Epoch 7/30
342/342 [==============================] - 30s 87ms/step - loss: 0.0151 - acc: 1.0000 - val_loss: 1.1282 - val_acc: 0.7179
Epoch 8/30
342/342 [==============================] - 30s 88ms/step - loss: 0.0208 - acc: 0.9942 - val_loss: 1.4929 - val_acc: 0.6923
Epoch 9/30
342/342 [==============================] - 29s 85ms/step - loss: 0.1117 - acc: 0.9474 - val_loss: 1.7814 - val_acc: 0.5641
Epoch 10/30
342/342 [==============================] - 31s 90ms/step - loss: 0.0651 - acc: 0.9766 - val_loss: 2.3173 - val_acc: 0.4872
Epoch 11/30
342/342 [==============================] - 30s 89ms/step - loss: 0.0502 - acc: 0.9854 - val_loss: 1.9904 - val_acc: 0.5897
Epoch 12/30
342/342 [==============================] - 30s 86ms/step - loss: 0.0431 - acc: 0.9854 - val_loss: 2.1962 - val_acc: 0.5385
Epoch 13/30
342/342 [==============================] - 27s 80ms/step - loss: 0.0483 - acc: 0.9825 - val_loss: 1.5308 - val_acc: 0.6923
Epoch 14/30
342/342 [==============================] - 29s 83ms/step - loss: 0.0164 - acc: 0.9971 - val_loss: 1.3372 - val_acc: 0.6923
Epoch 15/30
342/342 [==============================] - 29s 86ms/step - loss: 0.0318 - acc: 0.9942 - val_loss: 1.5257 - val_acc: 0.6923
Epoch 16/30
342/342 [==============================] - 30s 87ms/step - loss: 0.0186 - acc: 0.9942 - val_loss: 0.8881 - val_acc: 0.7179
Epoch 17/30
342/342 [==============================] - 31s 90ms/step - loss: 0.0467 - acc: 0.9854 - val_loss: 4.5571 - val_acc: 0.5128
Epoch 18/30
342/342 [==============================] - 31s 90ms/step - loss: 0.0257 - acc: 0.9854 - val_loss: 1.7651 - val_acc: 0.5385
Epoch 19/30
342/342 [==============================] - 34s 98ms/step - loss: 0.0359 - acc: 0.9854 - val_loss: 1.6619 - val_acc: 0.6667
Epoch 20/30
342/342 [==============================] - 33s 95ms/step - loss: 0.0175 - acc: 0.9971 - val_loss: 1.4698 - val_acc: 0.6923
Epoch 21/30
342/342 [==============================] - 33s 95ms/step - loss: 0.0121 - acc: 0.9971 - val_loss: 1.3313 - val_acc: 0.7436
Epoch 22/30
342/342 [==============================] - 32s 94ms/step - loss: 0.0120 - acc: 1.0000 - val_loss: 2.5492 - val_acc: 0.5897
Epoch 23/30
342/342 [==============================] - 32s 95ms/step - loss: 0.0180 - acc: 0.9971 - val_loss: 2.7641 - val_acc: 0.5385
Epoch 24/30
342/342 [==============================] - 32s 94ms/step - loss: 0.0128 - acc: 0.9942 - val_loss: 2.9518 - val_acc: 0.5641
Epoch 25/30
342/342 [==============================] - 31s 91ms/step - loss: 0.0433 - acc: 0.9883 - val_loss: 2.7912 - val_acc: 0.5641
Epoch 26/30
342/342 [==============================] - 31s 92ms/step - loss: 0.0321 - acc: 0.9854 - val_loss: 1.7816 - val_acc: 0.6154
Epoch 27/30
342/342 [==============================] - 31s 89ms/step - loss: 0.0247 - acc: 0.9912 - val_loss: 1.0928 - val_acc: 0.6923
Epoch 28/30
342/342 [==============================] - 27s 79ms/step - loss: 0.0099 - acc: 1.0000 - val_loss: 1.3549 - val_acc: 0.7179
Epoch 29/30
342/342 [==============================] - 26s 75ms/step - loss: 0.0143 - acc: 0.9971 - val_loss: 1.0779 - val_acc: 0.7949
Epoch 30/30
342/342 [==============================] - 26s 76ms/step - loss: 0.0058 - acc: 1.0000 - val_loss: 1.1930 - val_acc: 0.7436
Out[307]:<keras.callbacks.History at 0x14e0b8fd0>
In [308]:# optimizer : Adam / Dropout = 0.3 추가 / epochs : 30 / batch_size = 20
loss, acc = model2.evaluate(X_test,y_test,verbose=0)
print("Accuracy : %0.2f" % (acc*100))
Accuracy : 73.68
Epoch를 늘렸더니 성능이 조금 나아지긴 한다. Epoch를 100이상 주고싶지만 지금의 환경에서는 더 성능이 나올 것 같지는 않아서 Data Augmentation을 거치고 다시 학습해볼 것이다. 데이터 수가 많아지면 Validation에도 데이터를 더 부여해서 학습하는 과정을 가질 것이다.
In [307]:
model2.fit(np.array(X_train),np.array(y_train),epochs=30,batch_size=20,verbose=1,validation_split=0.1)
Train on 342 samples, validate on 39 samples Epoch 1/30 342/342 [==============================] - 31s 92ms/step - loss: 0.1395 - acc: 0.9591 - val_loss: 1.2031 - val_acc: 0.6154 Epoch 2/30 342/342 [==============================] - 31s 90ms/step - loss: 0.0693 - acc: 0.9708 - val_loss: 1.0359 - val_acc: 0.6923 Epoch 3/30 342/342 [==============================] - 30s 88ms/step - loss: 0.0739 - acc: 0.9737 - val_loss: 4.0959 - val_acc: 0.4615 Epoch 4/30 342/342 [==============================] - 29s 84ms/step - loss: 0.0737 - acc: 0.9766 - val_loss: 1.6908 - val_acc: 0.5385 Epoch 5/30 342/342 [==============================] - 31s 91ms/step - loss: 0.0370 - acc: 0.9883 - val_loss: 0.7877 - val_acc: 0.6410 Epoch 6/30 342/342 [==============================] - 31s 91ms/step - loss: 0.0274 - acc: 0.9912 - val_loss: 1.0398 - val_acc: 0.6154 Epoch 7/30 342/342 [==============================] - 30s 87ms/step - loss: 0.0151 - acc: 1.0000 - val_loss: 1.1282 - val_acc: 0.7179 Epoch 8/30 342/342 [==============================] - 30s 88ms/step - loss: 0.0208 - acc: 0.9942 - val_loss: 1.4929 - val_acc: 0.6923 Epoch 9/30 342/342 [==============================] - 29s 85ms/step - loss: 0.1117 - acc: 0.9474 - val_loss: 1.7814 - val_acc: 0.5641 Epoch 10/30 342/342 [==============================] - 31s 90ms/step - loss: 0.0651 - acc: 0.9766 - val_loss: 2.3173 - val_acc: 0.4872 Epoch 11/30 342/342 [==============================] - 30s 89ms/step - loss: 0.0502 - acc: 0.9854 - val_loss: 1.9904 - val_acc: 0.5897 Epoch 12/30 342/342 [==============================] - 30s 86ms/step - loss: 0.0431 - acc: 0.9854 - val_loss: 2.1962 - val_acc: 0.5385 Epoch 13/30 342/342 [==============================] - 27s 80ms/step - loss: 0.0483 - acc: 0.9825 - val_loss: 1.5308 - val_acc: 0.6923 Epoch 14/30 342/342 [==============================] - 29s 83ms/step - loss: 0.0164 - acc: 0.9971 - val_loss: 1.3372 - val_acc: 0.6923 Epoch 15/30 342/342 [==============================] - 29s 86ms/step - loss: 0.0318 - acc: 0.9942 - val_loss: 1.5257 - val_acc: 0.6923 Epoch 16/30 342/342 [==============================] - 30s 87ms/step - loss: 0.0186 - acc: 0.9942 - val_loss: 0.8881 - val_acc: 0.7179 Epoch 17/30 342/342 [==============================] - 31s 90ms/step - loss: 0.0467 - acc: 0.9854 - val_loss: 4.5571 - val_acc: 0.5128 Epoch 18/30 342/342 [==============================] - 31s 90ms/step - loss: 0.0257 - acc: 0.9854 - val_loss: 1.7651 - val_acc: 0.5385 Epoch 19/30 342/342 [==============================] - 34s 98ms/step - loss: 0.0359 - acc: 0.9854 - val_loss: 1.6619 - val_acc: 0.6667 Epoch 20/30 342/342 [==============================] - 33s 95ms/step - loss: 0.0175 - acc: 0.9971 - val_loss: 1.4698 - val_acc: 0.6923 Epoch 21/30 342/342 [==============================] - 33s 95ms/step - loss: 0.0121 - acc: 0.9971 - val_loss: 1.3313 - val_acc: 0.7436 Epoch 22/30 342/342 [==============================] - 32s 94ms/step - loss: 0.0120 - acc: 1.0000 - val_loss: 2.5492 - val_acc: 0.5897 Epoch 23/30 342/342 [==============================] - 32s 95ms/step - loss: 0.0180 - acc: 0.9971 - val_loss: 2.7641 - val_acc: 0.5385 Epoch 24/30 342/342 [==============================] - 32s 94ms/step - loss: 0.0128 - acc: 0.9942 - val_loss: 2.9518 - val_acc: 0.5641 Epoch 25/30 342/342 [==============================] - 31s 91ms/step - loss: 0.0433 - acc: 0.9883 - val_loss: 2.7912 - val_acc: 0.5641 Epoch 26/30 342/342 [==============================] - 31s 92ms/step - loss: 0.0321 - acc: 0.9854 - val_loss: 1.7816 - val_acc: 0.6154 Epoch 27/30 342/342 [==============================] - 31s 89ms/step - loss: 0.0247 - acc: 0.9912 - val_loss: 1.0928 - val_acc: 0.6923 Epoch 28/30 342/342 [==============================] - 27s 79ms/step - loss: 0.0099 - acc: 1.0000 - val_loss: 1.3549 - val_acc: 0.7179 Epoch 29/30 342/342 [==============================] - 26s 75ms/step - loss: 0.0143 - acc: 0.9971 - val_loss: 1.0779 - val_acc: 0.7949 Epoch 30/30 342/342 [==============================] - 26s 76ms/step - loss: 0.0058 - acc: 1.0000 - val_loss: 1.1930 - val_acc: 0.7436
Out[307]:
<keras.callbacks.History at 0x14e0b8fd0>
In [308]:
# optimizer : Adam / Dropout = 0.3 추가 / epochs : 30 / batch_size = 20
loss, acc = model2.evaluate(X_test,y_test,verbose=0)
print("Accuracy : %0.2f" % (acc*100))
Accuracy : 73.68
Epoch를 늘렸더니 성능이 조금 나아지긴 한다. Epoch를 100이상 주고싶지만 지금의 환경에서는 더 성능이 나올 것 같지는 않아서 Data Augmentation을 거치고 다시 학습해볼 것이다. 데이터 수가 많아지면 Validation에도 데이터를 더 부여해서 학습하는 과정을 가질 것이다.
Data Augmentation
Augmentation에는 다양한 방법이 존재한다. Flip_vertical / flip_horizontal / rotate / crop / translation, contrast, bright, inverse등이 있다.
추후에 labeling이 어려우니 시바견 따로, 진돗개 따로 Augmentation을 할 것이다.
In [69]:shiba_dog = list()
for fileName in os.listdir('./shiba'):
if fileName.startswith('google'):
shiba_dog.append(fileName)
shiba_img = list()
for dog in shiba_dog:
img = cv2.imread('./shiba/' + dog,cv2.IMREAD_GRAYSCALE)
resize = cv2.resize(img,(224,224))
shiba_img.append(resize)
shiba_img = np.array(shiba_img)
shiba_labels = list(list([1,0] for _ in range(n_shiba)))
In [70]:len(shiba_dog), len(shiba_img)
Out[70]:(232, 232)
In [71]:jindot_dog = list()
for fileName in os.listdir('./진돗개'):
jindot_dog.append(fileName)
jindot_img = list()
for dog in jindot_dog:
img = cv2.imread('./진돗개/' + dog,cv2.IMREAD_GRAYSCALE)
resize = cv2.resize(img,(224,224))
jindot_img.append(resize)
jindot_img = np.array(jindot_img)
jindo_labels = list(list([0,1] for _ in range(n_jindo)))
In [72]:len(jindot_dog), len(jindot_img)
Out[72]:(244, 244)
In [57]:import matplotlib.pyplot as plt
for i in range(0,9):
plt.subplot(330 + 1 + i)
plt.imshow(X_train[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
먼저 기존의 학습데이터 시바견과 진돗개들 출력했습니다.
Augmentation에는 다양한 방법이 존재한다. Flip_vertical / flip_horizontal / rotate / crop / translation, contrast, bright, inverse등이 있다.
추후에 labeling이 어려우니 시바견 따로, 진돗개 따로 Augmentation을 할 것이다.
In [69]:
shiba_dog = list()
for fileName in os.listdir('./shiba'):
if fileName.startswith('google'):
shiba_dog.append(fileName)
shiba_img = list()
for dog in shiba_dog:
img = cv2.imread('./shiba/' + dog,cv2.IMREAD_GRAYSCALE)
resize = cv2.resize(img,(224,224))
shiba_img.append(resize)
shiba_img = np.array(shiba_img)
shiba_labels = list(list([1,0] for _ in range(n_shiba)))
In [70]:
len(shiba_dog), len(shiba_img)
Out[70]:
(232, 232)
In [71]:
jindot_dog = list()
for fileName in os.listdir('./진돗개'):
jindot_dog.append(fileName)
jindot_img = list()
for dog in jindot_dog:
img = cv2.imread('./진돗개/' + dog,cv2.IMREAD_GRAYSCALE)
resize = cv2.resize(img,(224,224))
jindot_img.append(resize)
jindot_img = np.array(jindot_img)
jindo_labels = list(list([0,1] for _ in range(n_jindo)))
In [72]:
len(jindot_dog), len(jindot_img)
Out[72]:
(244, 244)
In [57]:
import matplotlib.pyplot as plt
for i in range(0,9):
plt.subplot(330 + 1 + i)
plt.imshow(X_train[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
먼저 기존의 학습데이터 시바견과 진돗개들 출력했습니다.
1. Feature Standardization
이미지의 픽셀을 전체 데이터셋에 대해 표준화하는 것이다.
In [66]:from keras.preprocessing.image import ImageDataGenerator
from keras import backend as K
K.set_image_dim_ordering('th')
이미지의 픽셀을 전체 데이터셋에 대해 표준화하는 것이다.
In [66]:
from keras.preprocessing.image import ImageDataGenerator
from keras import backend as K
K.set_image_dim_ordering('th')
시바견
In [73]:X_train_FS =shiba_img.reshape(shiba_img.shape[0],1,224,224)
X_train_FS = X_train_FS.astype('float32')
datagen = ImageDataGenerator(featurewise_center=True, featurewise_std_normalization=True)
datagen.fit(X_train_FS)
# datagen.flow(X_train_FS,y_train,batch_size=9)
for X_batch,y_batch in datagen.flow(X_train_FS,shiba_labels,batch_size=9,save_to_dir='shiba',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
/Users/charming/anaconda3/lib/python3.7/site-packages/matplotlib/text.py:1191: FutureWarning: elementwise comparison failed; returning scalar instead, but in the future will perform elementwise comparison
if s != self._text:
In [73]:
X_train_FS =shiba_img.reshape(shiba_img.shape[0],1,224,224)
X_train_FS = X_train_FS.astype('float32')
datagen = ImageDataGenerator(featurewise_center=True, featurewise_std_normalization=True)
datagen.fit(X_train_FS)
# datagen.flow(X_train_FS,y_train,batch_size=9)
for X_batch,y_batch in datagen.flow(X_train_FS,shiba_labels,batch_size=9,save_to_dir='shiba',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
/Users/charming/anaconda3/lib/python3.7/site-packages/matplotlib/text.py:1191: FutureWarning: elementwise comparison failed; returning scalar instead, but in the future will perform elementwise comparison if s != self._text:
진돗개
In [74]:X_train_FS =jindot_img.reshape(jindot_img.shape[0],1,224,224)
X_train_FS = X_train_FS.astype('float32')
datagen = ImageDataGenerator(featurewise_center=True, featurewise_std_normalization=True)
datagen.fit(X_train_FS)
# datagen.flow(X_train_FS,y_train,batch_size=9)
for X_batch,y_batch in datagen.flow(X_train_FS,jindo_labels,batch_size=9,save_to_dir='진돗개',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
/Users/charming/anaconda3/lib/python3.7/site-packages/matplotlib/text.py:1191: FutureWarning: elementwise comparison failed; returning scalar instead, but in the future will perform elementwise comparison
if s != self._text:
In [74]:
X_train_FS =jindot_img.reshape(jindot_img.shape[0],1,224,224)
X_train_FS = X_train_FS.astype('float32')
datagen = ImageDataGenerator(featurewise_center=True, featurewise_std_normalization=True)
datagen.fit(X_train_FS)
# datagen.flow(X_train_FS,y_train,batch_size=9)
for X_batch,y_batch in datagen.flow(X_train_FS,jindo_labels,batch_size=9,save_to_dir='진돗개',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
/Users/charming/anaconda3/lib/python3.7/site-packages/matplotlib/text.py:1191: FutureWarning: elementwise comparison failed; returning scalar instead, but in the future will perform elementwise comparison if s != self._text:
2. Random Rotations
시바견
In [75]:X_train_S =shiba_img.reshape(shiba_img.shape[0],1,224,224)
X_train_S = X_train_S.astype('float32')
datagen = ImageDataGenerator(rotation_range=90)
datagen.fit(X_train_S)
for X_batch,y_batch in datagen.flow(X_train_S,shiba_labels,batch_size=9,save_to_dir='shiba',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
In [75]:
X_train_S =shiba_img.reshape(shiba_img.shape[0],1,224,224)
X_train_S = X_train_S.astype('float32')
datagen = ImageDataGenerator(rotation_range=90)
datagen.fit(X_train_S)
for X_batch,y_batch in datagen.flow(X_train_S,shiba_labels,batch_size=9,save_to_dir='shiba',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
진돗개
In [76]:X_train_F =jindot_img.reshape(jindot_img.shape[0],1,224,224)
X_train_F = X_train_F.astype('float32')
datagen = ImageDataGenerator(rotation_range=90)
datagen.fit(X_train_F)
# datagen.flow(X_train_FS,y_train,batch_size=9)
for X_batch,y_batch in datagen.flow(X_train_F,jindo_labels,batch_size=9,save_to_dir='진돗개',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
In [76]:
X_train_F =jindot_img.reshape(jindot_img.shape[0],1,224,224)
X_train_F = X_train_F.astype('float32')
datagen = ImageDataGenerator(rotation_range=90)
datagen.fit(X_train_F)
# datagen.flow(X_train_FS,y_train,batch_size=9)
for X_batch,y_batch in datagen.flow(X_train_F,jindo_labels,batch_size=9,save_to_dir='진돗개',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
3. Random Shifts
Shiba
In [77]:X_train_S =shiba_img.reshape(shiba_img.shape[0],1,224,224)
X_train_S = X_train_S.astype('float32')
shift = 0.2
datagen = ImageDataGenerator(width_shift_range=shift, height_shift_range=shift)
datagen.fit(X_train_S)
for X_batch,y_batch in datagen.flow(X_train_S,shiba_labels,batch_size=9,save_to_dir='shiba',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
In [77]:
X_train_S =shiba_img.reshape(shiba_img.shape[0],1,224,224)
X_train_S = X_train_S.astype('float32')
shift = 0.2
datagen = ImageDataGenerator(width_shift_range=shift, height_shift_range=shift)
datagen.fit(X_train_S)
for X_batch,y_batch in datagen.flow(X_train_S,shiba_labels,batch_size=9,save_to_dir='shiba',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
진돗개
In [78]:X_train_F =jindot_img.reshape(jindot_img.shape[0],1,224,224)
X_train_F = X_train_F.astype('float32')
shift = 0.2
datagen = ImageDataGenerator(width_shift_range=shift, height_shift_range=shift)
datagen.fit(X_train_F)
# datagen.flow(X_train_FS,y_train,batch_size=9)
for X_batch,y_batch in datagen.flow(X_train_F,jindo_labels,batch_size=9,save_to_dir='진돗개',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
In [78]:
X_train_F =jindot_img.reshape(jindot_img.shape[0],1,224,224)
X_train_F = X_train_F.astype('float32')
shift = 0.2
datagen = ImageDataGenerator(width_shift_range=shift, height_shift_range=shift)
datagen.fit(X_train_F)
# datagen.flow(X_train_FS,y_train,batch_size=9)
for X_batch,y_batch in datagen.flow(X_train_F,jindo_labels,batch_size=9,save_to_dir='진돗개',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
4. Random Flips
시바견
In [79]:X_train_S =shiba_img.reshape(shiba_img.shape[0],1,224,224)
X_train_S = X_train_S.astype('float32')
datagen = ImageDataGenerator(horizontal_flip=True,vertical_flip=True)
datagen.fit(X_train_S)
for X_batch,y_batch in datagen.flow(X_train_S,shiba_labels,batch_size=9,save_to_dir='shiba',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
In [79]:
X_train_S =shiba_img.reshape(shiba_img.shape[0],1,224,224)
X_train_S = X_train_S.astype('float32')
datagen = ImageDataGenerator(horizontal_flip=True,vertical_flip=True)
datagen.fit(X_train_S)
for X_batch,y_batch in datagen.flow(X_train_S,shiba_labels,batch_size=9,save_to_dir='shiba',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
진돗개
In [80]:X_train_F =jindot_img.reshape(jindot_img.shape[0],1,224,224)
X_train_F = X_train_F.astype('float32')
datagen = ImageDataGenerator(horizontal_flip=True,vertical_flip=True)
datagen.fit(X_train_F)
# datagen.flow(X_train_FS,y_train,batch_size=9)
for X_batch,y_batch in datagen.flow(X_train_F,jindo_labels,batch_size=9,save_to_dir='진돗개',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
Data Augmentation 이후의 학습과정과 평가 과정은 2부에서...
In [80]:
X_train_F =jindot_img.reshape(jindot_img.shape[0],1,224,224)
X_train_F = X_train_F.astype('float32')
datagen = ImageDataGenerator(horizontal_flip=True,vertical_flip=True)
datagen.fit(X_train_F)
# datagen.flow(X_train_FS,y_train,batch_size=9)
for X_batch,y_batch in datagen.flow(X_train_F,jindo_labels,batch_size=9,save_to_dir='진돗개',save_prefix='aug',save_format="jpeg"):
for i in range(0,9):
plt.subplot(330+1+i)
plt.title(y_batch[i])
plt.imshow(X_batch[i].reshape(224,224),cmap=plt.cm.gray)
plt.show()
Data Augmentation 이후의 학습과정과 평가 과정은 2부에서...
728x90
반응형