비행기, 자동차, 자전거의 google이미지를 받아서 딥러닝으로 학습한 후에 사진을 분류하는 소스입니다.
사진을 많이 수집해서 하면 좀더 높은 정확도가 나올 것 같은데 사진을 추려내는 데 시간이 많이 걸려서 각 400개 정도의 사진만으로 학습을 시켰기 때문에 정확도가 그리 높게 나오진 않았어요^^
깃 허브에도 동일한 소스가 있습니다.
https://github.com/llejo3/deep-learning/blob/master/vehicle_classification_learning.md
from google_images_download import google_images_download
response = google_images_download.googleimagesdownload()
구글 사진을 검색어로 가져오는데, 검색어와 맞지 않는 사진은 삭제하는 작업이 좀 필요합니다.
#arguments = {"keywords":"forsythia,cherry blossoms,magnolia,azalea,tulip",
arguments = {"keywords":"Planes, Bicycles, Cars",
"limit":600,
"print_urls":True,
format: "jpg,png",
"chromedriver" : "./chromedriver.exe",
"exact_size":"200,200"}
paths = response.download(arguments)
from keras.layers import Conv2D, Dense, Flatten, MaxPooling2D, Dropout
from keras.layers.normalization import BatchNormalization
from keras.models import Sequential
from keras.preprocessing.image import ImageDataGenerator
from keras.optimizers import Adam, RMSprop, SGD
from keras.callbacks import EarlyStopping, TensorBoard, ModelCheckpoint
train_datagen = ImageDataGenerator(
rescale=1./255,
rotation_range=40, # 40도까지 회전
width_shift_range=0.2, # 20%까지 좌우 이동
height_shift_range=0.2, # 20%까지 상하 이동
shear_range=0.2, # 20%까지 기울임
zoom_range=0.2, # 20%까지 확대
horizontal_flip=True # 좌우 뒤집기
)
train = train_datagen.flow_from_directory(
'vehicles/train',
target_size=(200, 200),
#batch_size=32,
class_mode='categorical')
valid = ImageDataGenerator(rescale=1.0/255).flow_from_directory(
'vehicles/validation',
target_size=(200, 200),
#batch_size=32,
class_mode='categorical',
shuffle=False)
Found 884 images belonging to 3 classes.
Found 298 images belonging to 3 classes.
model1 = Sequential()
model1.add(Conv2D(32, (3, 3), activation='relu', input_shape=(200, 200, 3)))
model1.add(MaxPooling2D((2, 2)))
model1.add(Flatten())
model1.add(Dense(3, activation='softmax'))
model1.summary()
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_8 (Conv2D) (None, 198, 198, 32) 896
_________________________________________________________________
max_pooling2d_8 (MaxPooling2 (None, 99, 99, 32) 0
_________________________________________________________________
flatten_7 (Flatten) (None, 313632) 0
_________________________________________________________________
dense_7 (Dense) (None, 3) 940899
=================================================================
Total params: 941,795
Trainable params: 941,795
Non-trainable params: 0
_________________________________________________________________
sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model1.compile(loss='categorical_crossentropy', metrics=['accuracy'], optimizer=sgd)
history1 = model1.fit_generator(
train, validation_data=valid, epochs=30,
callbacks=[
EarlyStopping(monitor = "val_loss", patience=2),
TensorBoard(log_dir='tensorboard_logs/log_model1')
])
Epoch 1/30
1/28 [>.............................] - ETA: 46s - loss: 1.1692 - acc: 0.3438
C:\Users\south\Anaconda3\lib\site-packages\PIL\Image.py:918: UserWarning: Palette images with Transparency expressed in bytes should be converted to RGBA images
'to RGBA images')
28/28 [==============================] - 36s 1s/step - loss: 3.7392 - acc: 0.3803 - val_loss: 1.0593 - val_acc: 0.4564
Epoch 2/30
28/28 [==============================] - 35s 1s/step - loss: 1.1686 - acc: 0.4723 - val_loss: 1.0099 - val_acc: 0.5839
Epoch 3/30
28/28 [==============================] - 36s 1s/step - loss: 0.9614 - acc: 0.5998 - val_loss: 0.9786 - val_acc: 0.5772
Epoch 4/30
28/28 [==============================] - 36s 1s/step - loss: 0.9075 - acc: 0.6209 - val_loss: 0.9295 - val_acc: 0.5805
Epoch 5/30
28/28 [==============================] - 36s 1s/step - loss: 0.8156 - acc: 0.6610 - val_loss: 0.8961 - val_acc: 0.6242
Epoch 6/30
28/28 [==============================] - 36s 1s/step - loss: 0.7874 - acc: 0.6738 - val_loss: 0.9063 - val_acc: 0.5973
Epoch 7/30
28/28 [==============================] - 37s 1s/step - loss: 0.7597 - acc: 0.6724 - val_loss: 0.8911 - val_acc: 0.6074
Epoch 8/30
28/28 [==============================] - 37s 1s/step - loss: 0.7685 - acc: 0.6638 - val_loss: 0.9621 - val_acc: 0.5772
Epoch 9/30
28/28 [==============================] - 34s 1s/step - loss: 0.7566 - acc: 0.6674 - val_loss: 0.8563 - val_acc: 0.6376
Epoch 10/30
28/28 [==============================] - 34s 1s/step - loss: 0.6990 - acc: 0.7087 - val_loss: 0.9243 - val_acc: 0.6074
Epoch 11/30
28/28 [==============================] - 35s 1s/step - loss: 0.7083 - acc: 0.7040 - val_loss: 0.9144 - val_acc: 0.6208
model2 = Sequential()
model2.add(Conv2D(32, (3, 3), activation='relu', input_shape=(200, 200, 3)))
model2.add(MaxPooling2D((2, 2)))
model2.add(Conv2D(32, (3, 3), activation='relu'))
model2.add(MaxPooling2D((2, 2)))
model2.add(Flatten())
model2.add(Dense(3, activation='softmax'))
model2.summary()
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_11 (Conv2D) (None, 198, 198, 32) 896
_________________________________________________________________
max_pooling2d_11 (MaxPooling (None, 99, 99, 32) 0
_________________________________________________________________
conv2d_12 (Conv2D) (None, 97, 97, 32) 9248
_________________________________________________________________
max_pooling2d_12 (MaxPooling (None, 48, 48, 32) 0
_________________________________________________________________
flatten_9 (Flatten) (None, 73728) 0
_________________________________________________________________
dense_9 (Dense) (None, 3) 221187
=================================================================
Total params: 231,331
Trainable params: 231,331
Non-trainable params: 0
_________________________________________________________________
sgd = SGD(lr=0.001, decay=1e-6, momentum=0.9, nesterov=True)
model2.compile(loss='categorical_crossentropy', metrics=['accuracy'], optimizer=sgd)
history2 = model2.fit_generator(
train, validation_data=valid, epochs=100,
callbacks=[
EarlyStopping(monitor = "val_loss", patience=2),
TensorBoard(log_dir='tensorboard_logs/log_model2')
])
Epoch 1/30
4/28 [===>..........................] - ETA: 44s - loss: 1.4224 - acc: 0.3391
C:\Users\south\Anaconda3\lib\site-packages\PIL\Image.py:918: UserWarning: Palette images with Transparency expressed in bytes should be converted to RGBA images
'to RGBA images')
28/28 [==============================] - 57s 2s/step - loss: 1.1169 - acc: 0.4312 - val_loss: 1.3070 - val_acc: 0.3658
Epoch 2/30
28/28 [==============================] - 53s 2s/step - loss: 1.0442 - acc: 0.4781 - val_loss: 1.0760 - val_acc: 0.4262
Epoch 3/30
28/28 [==============================] - 55s 2s/step - loss: 0.9853 - acc: 0.5179 - val_loss: 0.9727 - val_acc: 0.5503
Epoch 4/30
28/28 [==============================] - 55s 2s/step - loss: 0.9722 - acc: 0.5405 - val_loss: 1.0100 - val_acc: 0.6107
Epoch 5/30
28/28 [==============================] - 54s 2s/step - loss: 0.9162 - acc: 0.5844 - val_loss: 0.9376 - val_acc: 0.6040
Epoch 6/30
28/28 [==============================] - 59s 2s/step - loss: 0.7842 - acc: 0.6691 - val_loss: 1.0358 - val_acc: 0.6107
Epoch 7/30
28/28 [==============================] - 55s 2s/step - loss: 0.7229 - acc: 0.6908 - val_loss: 0.7254 - val_acc: 0.7148
Epoch 8/30
28/28 [==============================] - 55s 2s/step - loss: 0.7381 - acc: 0.6953 - val_loss: 0.8614 - val_acc: 0.6342
Epoch 9/30
28/28 [==============================] - 54s 2s/step - loss: 0.6778 - acc: 0.7174 - val_loss: 0.7262 - val_acc: 0.6980
model3 = Sequential()
model3.add(Conv2D(32, (3, 3), activation='relu', input_shape=(200, 200, 3)))
model3.add(BatchNormalization())
model3.add(Conv2D(32, (3, 3), activation='relu'))
model3.add(MaxPooling2D(pool_size=(2, 2)))
model3.add(Dropout(0.25))
model3.add(Conv2D(64, (3, 3), activation='relu'))
model3.add(BatchNormalization())
model3.add(Conv2D(64, (3, 3), activation='relu'))
model3.add(MaxPooling2D(pool_size=(2, 2)))
model3.add(Dropout(0.25))
model3.add(Flatten())
model3.add(Dense(256, activation='relu'))
model3.add(Dropout(0.5))
model3.add(Dense(3, activation='softmax'))
model3.summary()
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_19 (Conv2D) (None, 198, 198, 32) 896
_________________________________________________________________
max_pooling2d_16 (MaxPooling (None, 99, 99, 32) 0
_________________________________________________________________
dropout_4 (Dropout) (None, 99, 99, 32) 0
_________________________________________________________________
conv2d_20 (Conv2D) (None, 97, 97, 64) 18496
_________________________________________________________________
max_pooling2d_17 (MaxPooling (None, 48, 48, 64) 0
_________________________________________________________________
dropout_5 (Dropout) (None, 48, 48, 64) 0
_________________________________________________________________
flatten_11 (Flatten) (None, 147456) 0
_________________________________________________________________
dense_12 (Dense) (None, 256) 37748992
_________________________________________________________________
dropout_6 (Dropout) (None, 256) 0
_________________________________________________________________
dense_13 (Dense) (None, 3) 771
=================================================================
Total params: 37,769,155
Trainable params: 37,769,155
Non-trainable params: 0
_________________________________________________________________
sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model3.compile(loss='categorical_crossentropy', metrics=['accuracy'], optimizer=sgd)
history3 = model3.fit_generator(
train, validation_data=valid, epochs=30,
callbacks=[
EarlyStopping(monitor = "val_loss", patience=2),
ModelCheckpoint('model3-{epoch:02d}.hdf5', save_best_only=True),
TensorBoard(log_dir='tensorboard_logs/log_model3')
])
Epoch 1/30
1/28 [>.............................] - ETA: 1:34 - loss: 1.1490 - acc: 0.2500
C:\Users\south\Anaconda3\lib\site-packages\PIL\Image.py:918: UserWarning: Palette images with Transparency expressed in bytes should be converted to RGBA images
'to RGBA images')
28/28 [==============================] - 85s 3s/step - loss: 1.2178 - acc: 0.3502 - val_loss: 1.0796 - val_acc: 0.3893
Epoch 2/30
28/28 [==============================] - 86s 3s/step - loss: 1.0463 - acc: 0.4851 - val_loss: 1.0564 - val_acc: 0.4899
Epoch 3/30
28/28 [==============================] - 82s 3s/step - loss: 0.9830 - acc: 0.5198 - val_loss: 1.0233 - val_acc: 0.5839
Epoch 4/30
28/28 [==============================] - 86s 3s/step - loss: 0.8876 - acc: 0.6224 - val_loss: 0.9437 - val_acc: 0.5336
Epoch 5/30
28/28 [==============================] - 84s 3s/step - loss: 0.7544 - acc: 0.6600 - val_loss: 0.8263 - val_acc: 0.6577
Epoch 6/30
28/28 [==============================] - 86s 3s/step - loss: 0.7302 - acc: 0.7052 - val_loss: 0.7034 - val_acc: 0.6879
Epoch 7/30
28/28 [==============================] - 89s 3s/step - loss: 0.6840 - acc: 0.7004 - val_loss: 0.6810 - val_acc: 0.7047
Epoch 8/30
28/28 [==============================] - 89s 3s/step - loss: 0.6499 - acc: 0.7214 - val_loss: 0.6697 - val_acc: 0.7248
Epoch 9/30
28/28 [==============================] - 87s 3s/step - loss: 0.6163 - acc: 0.7403 - val_loss: 0.6561 - val_acc: 0.7383
Epoch 10/30
28/28 [==============================] - 107s 4s/step - loss: 0.6387 - acc: 0.7238 - val_loss: 0.6992 - val_acc: 0.6846
Epoch 11/30
28/28 [==============================] - 90s 3s/step - loss: 0.5976 - acc: 0.7457 - val_loss: 0.6568 - val_acc: 0.7114
