GoogLeNet
위 논문에 따라 Keras로 구현해보았다. 논문의 unit과 num_class는 각각 1024, 1000 이다.
Auxiliary Classifier
def AuxiliaryClassifier(inputs, name=None):
x = AveragePooling2D((5,5), strides=3, padding='valid')(inputs)
x = Conv2D(128, (1,1), strides=1, padding='same', activation='relu',
kernel_initializer=kernel_initializer, bias_initializer=bias_initializer)(x)
x = Flatten()(x)
x = Dense(unit, activation='relu',
kernel_initializer=kernel_initializer, bias_initializer=bias_initializer)(x)
x = Dropout(0.7)(x)
x = Dense(num_class, activation='softmax', name=name)(x)
return x사이사이에 존재하는 gradient vanishing 문제를 해결하기 위한 Auxiliary Classifier.
Inception Module
급격히 불어나는 연산량을 줄이기 위한 Bottleneck 구조를 사용.
def InceptionModule(inputs, f1, f3r, f3, f5r, f5, fp, name=None):
conv_1x1 = Conv2D(f1, (1,1), padding='same', activation='relu',
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer)(inputs)
x1 = Conv2D(f3r, (1,1), padding='same', activation='relu',
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer)(inputs)
conv_3x3 = Conv2D(f3, (3,3), padding='same', activation='relu',
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer)(x1)
x2 = Conv2D(f5r, (1,1), padding='same', activation='relu',
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer)(inputs)
conv_5x5 = Conv2D(f5, (5,5), padding='same', activation='relu',
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer)(x2)
x3 = MaxPool2D((3,3), strides=1, padding='same')(inputs)
conv_pool = Conv2D(fp, (1,1), padding='same', activation='relu',
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer)(x3)
concat = Concatenate(axis=-1)([conv_1x1, conv_3x3, conv_5x5, conv_pool])
return concatGoogLeNet
위의 GoogLeNet 구조를 보며 구현했다.
def GoogLeNet():
inputs = Input(shape=(224,224,3), name='inputs')
x = Conv2D(64, (7,7), strides=2, padding='same', activation='relu', name='conv_1',
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer)(inputs)
x = MaxPool2D((3,3), strides=2, padding='same', name='maxpool_1')(x)
x = BatchNormalization()(x)
x = tf.nn.local_response_normalization(x, alpha=.0001, beta=.75, name='lrn_1')
x = Conv2D(64, (1,1), strides=1, padding='valid', activation='relu', name='conv_2',
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer)(x)
x = Conv2D(192, (3,3), strides=1, padding='same', activation='relu', name='conv_3',
kernel_initializer=kernel_initializer,
bias_initializer=bias_initializer)(x)
x = BatchNormalization()(x)
x = tf.nn.local_response_normalization(x, alpha=.0001, beta=.75, name='lrn_2')
x = MaxPool2D((3,3), strides=2, padding='same', name='maxpool_2')(x)
x = InceptionModule(x, 64, 96, 128, 16, 32, 32, name='inception (3a)')
x = InceptionModule(x, 128, 128, 192, 32, 96, 64, name='inception (3b)')
x = MaxPool2D((3,3), strides=2, padding='same', name='maxpool_3')(x)
x = InceptionModule(x, 192, 96, 208, 16, 48, 64, name='inception (4a)')
aux1 = AuxiliaryClassifier(x, name='aux1')
x = InceptionModule(x, 160, 112, 224, 24, 64, 64, name='inception (4b)')
x = InceptionModule(x, 128, 128, 256, 24, 64, 64, name='inception (4c)')
x = InceptionModule(x, 112, 144, 288, 32, 64, 64, name='inception (4d)')
aux2 = AuxiliaryClassifier(x, name='aux2')
x = InceptionModule(x, 256, 160, 320, 32, 128, 128, name='inception (4e)')
x = MaxPool2D((3,3), strides=2, padding='same', name='maxpool_4')(x)
x = InceptionModule(x, 256, 160, 320, 32, 128, 128, name='inception (5a)')
x = InceptionModule(x, 384, 192, 384, 48, 128, 128, name='inception (5b)')
x = GlobalAveragePooling2D(name='gap')(x)
x = Dropout(0.4, name='dropout')(x)
main = Dense(num_class, activation='softmax', name='main')(x)
model = Model(inputs=inputs, outputs=[main, aux1, aux2])
return modelModel Summary
model.compile(optimizer=optimizer,
loss='sparse_categorical_crossentropy',
loss_weights={'main':1.0, 'aux1':0.3, 'aux2':0.3},
metrics=['acc'])
model.summary()
'''
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
inputs (InputLayer) [(None, 224, 224, 3) 0
__________________________________________________________________________________________________
conv_1 (Conv2D) (None, 112, 112, 64) 9472 inputs[0][0]
__________________________________________________________________________________________________
maxpool_1 (MaxPooling2D) (None, 56, 56, 64) 0 conv_1[0][0]
__________________________________________________________________________________________________
batch_normalization_2 (BatchNor (None, 56, 56, 64) 256 maxpool_1[0][0]
__________________________________________________________________________________________________
tf.nn.local_response_normalizat (None, 56, 56, 64) 0 batch_normalization_2[0][0]
__________________________________________________________________________________________________
conv_2 (Conv2D) (None, 56, 56, 64) 4160 tf.nn.local_response_normalizatio
__________________________________________________________________________________________________
conv_3 (Conv2D) (None, 56, 56, 192) 110784 conv_2[0][0]
__________________________________________________________________________________________________
batch_normalization_3 (BatchNor (None, 56, 56, 192) 768 conv_3[0][0]
__________________________________________________________________________________________________
tf.nn.local_response_normalizat (None, 56, 56, 192) 0 batch_normalization_3[0][0]
__________________________________________________________________________________________________
maxpool_2 (MaxPooling2D) (None, 28, 28, 192) 0 tf.nn.local_response_normalizatio
__________________________________________________________________________________________________
conv2d_57 (Conv2D) (None, 28, 28, 96) 18528 maxpool_2[0][0]
__________________________________________________________________________________________________
conv2d_59 (Conv2D) (None, 28, 28, 16) 3088 maxpool_2[0][0]
__________________________________________________________________________________________________
max_pooling2d_9 (MaxPooling2D) (None, 28, 28, 192) 0 maxpool_2[0][0]
__________________________________________________________________________________________________
conv2d_56 (Conv2D) (None, 28, 28, 64) 12352 maxpool_2[0][0]
__________________________________________________________________________________________________
conv2d_58 (Conv2D) (None, 28, 28, 128) 110720 conv2d_57[0][0]
__________________________________________________________________________________________________
conv2d_60 (Conv2D) (None, 28, 28, 32) 12832 conv2d_59[0][0]
__________________________________________________________________________________________________
conv2d_61 (Conv2D) (None, 28, 28, 32) 6176 max_pooling2d_9[0][0]
__________________________________________________________________________________________________
concatenate_9 (Concatenate) (None, 28, 28, 256) 0 conv2d_56[0][0]
conv2d_58[0][0]
conv2d_60[0][0]
conv2d_61[0][0]
__________________________________________________________________________________________________
conv2d_63 (Conv2D) (None, 28, 28, 128) 32896 concatenate_9[0][0]
__________________________________________________________________________________________________
conv2d_65 (Conv2D) (None, 28, 28, 32) 8224 concatenate_9[0][0]
__________________________________________________________________________________________________
max_pooling2d_10 (MaxPooling2D) (None, 28, 28, 256) 0 concatenate_9[0][0]
__________________________________________________________________________________________________
conv2d_62 (Conv2D) (None, 28, 28, 128) 32896 concatenate_9[0][0]
__________________________________________________________________________________________________
conv2d_64 (Conv2D) (None, 28, 28, 192) 221376 conv2d_63[0][0]
__________________________________________________________________________________________________
conv2d_66 (Conv2D) (None, 28, 28, 96) 76896 conv2d_65[0][0]
__________________________________________________________________________________________________
conv2d_67 (Conv2D) (None, 28, 28, 64) 16448 max_pooling2d_10[0][0]
__________________________________________________________________________________________________
concatenate_10 (Concatenate) (None, 28, 28, 480) 0 conv2d_62[0][0]
conv2d_64[0][0]
conv2d_66[0][0]
conv2d_67[0][0]
__________________________________________________________________________________________________
maxpool_3 (MaxPooling2D) (None, 14, 14, 480) 0 concatenate_10[0][0]
__________________________________________________________________________________________________
conv2d_69 (Conv2D) (None, 14, 14, 96) 46176 maxpool_3[0][0]
__________________________________________________________________________________________________
conv2d_71 (Conv2D) (None, 14, 14, 16) 7696 maxpool_3[0][0]
__________________________________________________________________________________________________
max_pooling2d_11 (MaxPooling2D) (None, 14, 14, 480) 0 maxpool_3[0][0]
__________________________________________________________________________________________________
conv2d_68 (Conv2D) (None, 14, 14, 192) 92352 maxpool_3[0][0]
__________________________________________________________________________________________________
conv2d_70 (Conv2D) (None, 14, 14, 208) 179920 conv2d_69[0][0]
__________________________________________________________________________________________________
conv2d_72 (Conv2D) (None, 14, 14, 48) 19248 conv2d_71[0][0]
__________________________________________________________________________________________________
conv2d_73 (Conv2D) (None, 14, 14, 64) 30784 max_pooling2d_11[0][0]
__________________________________________________________________________________________________
concatenate_11 (Concatenate) (None, 14, 14, 512) 0 conv2d_68[0][0]
conv2d_70[0][0]
conv2d_72[0][0]
conv2d_73[0][0]
__________________________________________________________________________________________________
conv2d_76 (Conv2D) (None, 14, 14, 112) 57456 concatenate_11[0][0]
__________________________________________________________________________________________________
conv2d_78 (Conv2D) (None, 14, 14, 24) 12312 concatenate_11[0][0]
__________________________________________________________________________________________________
max_pooling2d_12 (MaxPooling2D) (None, 14, 14, 512) 0 concatenate_11[0][0]
__________________________________________________________________________________________________
conv2d_75 (Conv2D) (None, 14, 14, 160) 82080 concatenate_11[0][0]
__________________________________________________________________________________________________
conv2d_77 (Conv2D) (None, 14, 14, 224) 226016 conv2d_76[0][0]
__________________________________________________________________________________________________
conv2d_79 (Conv2D) (None, 14, 14, 64) 38464 conv2d_78[0][0]
__________________________________________________________________________________________________
conv2d_80 (Conv2D) (None, 14, 14, 64) 32832 max_pooling2d_12[0][0]
__________________________________________________________________________________________________
concatenate_12 (Concatenate) (None, 14, 14, 512) 0 conv2d_75[0][0]
conv2d_77[0][0]
conv2d_79[0][0]
conv2d_80[0][0]
__________________________________________________________________________________________________
conv2d_82 (Conv2D) (None, 14, 14, 128) 65664 concatenate_12[0][0]
__________________________________________________________________________________________________
conv2d_84 (Conv2D) (None, 14, 14, 24) 12312 concatenate_12[0][0]
__________________________________________________________________________________________________
max_pooling2d_13 (MaxPooling2D) (None, 14, 14, 512) 0 concatenate_12[0][0]
__________________________________________________________________________________________________
conv2d_81 (Conv2D) (None, 14, 14, 128) 65664 concatenate_12[0][0]
__________________________________________________________________________________________________
conv2d_83 (Conv2D) (None, 14, 14, 256) 295168 conv2d_82[0][0]
__________________________________________________________________________________________________
conv2d_85 (Conv2D) (None, 14, 14, 64) 38464 conv2d_84[0][0]
__________________________________________________________________________________________________
conv2d_86 (Conv2D) (None, 14, 14, 64) 32832 max_pooling2d_13[0][0]
__________________________________________________________________________________________________
concatenate_13 (Concatenate) (None, 14, 14, 512) 0 conv2d_81[0][0]
conv2d_83[0][0]
conv2d_85[0][0]
conv2d_86[0][0]
__________________________________________________________________________________________________
conv2d_88 (Conv2D) (None, 14, 14, 144) 73872 concatenate_13[0][0]
__________________________________________________________________________________________________
conv2d_90 (Conv2D) (None, 14, 14, 32) 16416 concatenate_13[0][0]
__________________________________________________________________________________________________
max_pooling2d_14 (MaxPooling2D) (None, 14, 14, 512) 0 concatenate_13[0][0]
__________________________________________________________________________________________________
conv2d_87 (Conv2D) (None, 14, 14, 112) 57456 concatenate_13[0][0]
__________________________________________________________________________________________________
conv2d_89 (Conv2D) (None, 14, 14, 288) 373536 conv2d_88[0][0]
__________________________________________________________________________________________________
conv2d_91 (Conv2D) (None, 14, 14, 64) 51264 conv2d_90[0][0]
__________________________________________________________________________________________________
conv2d_92 (Conv2D) (None, 14, 14, 64) 32832 max_pooling2d_14[0][0]
__________________________________________________________________________________________________
concatenate_14 (Concatenate) (None, 14, 14, 528) 0 conv2d_87[0][0]
conv2d_89[0][0]
conv2d_91[0][0]
conv2d_92[0][0]
__________________________________________________________________________________________________
conv2d_95 (Conv2D) (None, 14, 14, 160) 84640 concatenate_14[0][0]
__________________________________________________________________________________________________
conv2d_97 (Conv2D) (None, 14, 14, 32) 16928 concatenate_14[0][0]
__________________________________________________________________________________________________
max_pooling2d_15 (MaxPooling2D) (None, 14, 14, 528) 0 concatenate_14[0][0]
__________________________________________________________________________________________________
conv2d_94 (Conv2D) (None, 14, 14, 256) 135424 concatenate_14[0][0]
__________________________________________________________________________________________________
conv2d_96 (Conv2D) (None, 14, 14, 320) 461120 conv2d_95[0][0]
__________________________________________________________________________________________________
conv2d_98 (Conv2D) (None, 14, 14, 128) 102528 conv2d_97[0][0]
__________________________________________________________________________________________________
conv2d_99 (Conv2D) (None, 14, 14, 128) 67712 max_pooling2d_15[0][0]
__________________________________________________________________________________________________
concatenate_15 (Concatenate) (None, 14, 14, 832) 0 conv2d_94[0][0]
conv2d_96[0][0]
conv2d_98[0][0]
conv2d_99[0][0]
__________________________________________________________________________________________________
maxpool_4 (MaxPooling2D) (None, 7, 7, 832) 0 concatenate_15[0][0]
__________________________________________________________________________________________________
conv2d_101 (Conv2D) (None, 7, 7, 160) 133280 maxpool_4[0][0]
__________________________________________________________________________________________________
conv2d_103 (Conv2D) (None, 7, 7, 32) 26656 maxpool_4[0][0]
__________________________________________________________________________________________________
max_pooling2d_16 (MaxPooling2D) (None, 7, 7, 832) 0 maxpool_4[0][0]
__________________________________________________________________________________________________
conv2d_100 (Conv2D) (None, 7, 7, 256) 213248 maxpool_4[0][0]
__________________________________________________________________________________________________
conv2d_102 (Conv2D) (None, 7, 7, 320) 461120 conv2d_101[0][0]
__________________________________________________________________________________________________
conv2d_104 (Conv2D) (None, 7, 7, 128) 102528 conv2d_103[0][0]
__________________________________________________________________________________________________
conv2d_105 (Conv2D) (None, 7, 7, 128) 106624 max_pooling2d_16[0][0]
__________________________________________________________________________________________________
concatenate_16 (Concatenate) (None, 7, 7, 832) 0 conv2d_100[0][0]
conv2d_102[0][0]
conv2d_104[0][0]
conv2d_105[0][0]
__________________________________________________________________________________________________
conv2d_107 (Conv2D) (None, 7, 7, 192) 159936 concatenate_16[0][0]
__________________________________________________________________________________________________
conv2d_109 (Conv2D) (None, 7, 7, 48) 39984 concatenate_16[0][0]
__________________________________________________________________________________________________
max_pooling2d_17 (MaxPooling2D) (None, 7, 7, 832) 0 concatenate_16[0][0]
__________________________________________________________________________________________________
average_pooling2d_2 (AveragePoo (None, 4, 4, 512) 0 concatenate_11[0][0]
__________________________________________________________________________________________________
average_pooling2d_3 (AveragePoo (None, 4, 4, 528) 0 concatenate_14[0][0]
__________________________________________________________________________________________________
conv2d_106 (Conv2D) (None, 7, 7, 384) 319872 concatenate_16[0][0]
__________________________________________________________________________________________________
conv2d_108 (Conv2D) (None, 7, 7, 384) 663936 conv2d_107[0][0]
__________________________________________________________________________________________________
conv2d_110 (Conv2D) (None, 7, 7, 128) 153728 conv2d_109[0][0]
__________________________________________________________________________________________________
conv2d_111 (Conv2D) (None, 7, 7, 128) 106624 max_pooling2d_17[0][0]
__________________________________________________________________________________________________
conv2d_74 (Conv2D) (None, 4, 4, 128) 65664 average_pooling2d_2[0][0]
__________________________________________________________________________________________________
conv2d_93 (Conv2D) (None, 4, 4, 128) 67712 average_pooling2d_3[0][0]
__________________________________________________________________________________________________
concatenate_17 (Concatenate) (None, 7, 7, 1024) 0 conv2d_106[0][0]
conv2d_108[0][0]
conv2d_110[0][0]
conv2d_111[0][0]
__________________________________________________________________________________________________
flatten_2 (Flatten) (None, 2048) 0 conv2d_74[0][0]
__________________________________________________________________________________________________
flatten_3 (Flatten) (None, 2048) 0 conv2d_93[0][0]
__________________________________________________________________________________________________
gap (GlobalAveragePooling2D) (None, 1024) 0 concatenate_17[0][0]
__________________________________________________________________________________________________
dense_2 (Dense) (None, 1024) 2098176 flatten_2[0][0]
__________________________________________________________________________________________________
dense_3 (Dense) (None, 1024) 2098176 flatten_3[0][0]
__________________________________________________________________________________________________
dropout (Dropout) (None, 1024) 0 gap[0][0]
__________________________________________________________________________________________________
dropout_2 (Dropout) (None, 1024) 0 dense_2[0][0]
__________________________________________________________________________________________________
dropout_3 (Dropout) (None, 1024) 0 dense_3[0][0]
__________________________________________________________________________________________________
main (Dense) (None, 10) 10250 dropout[0][0]
__________________________________________________________________________________________________
aux1 (Dense) (None, 10) 10250 dropout_2[0][0]
__________________________________________________________________________________________________
aux2 (Dense) (None, 10) 10250 dropout_3[0][0]
==================================================================================================
Total params: 10,335,054
Trainable params: 10,334,542
Non-trainable params: 512
__________________________________________________________________________________________________
''''ML > CNN' 카테고리의 다른 글
| [CNN] DenseNet Implementation (Keras) (0) | 2021.08.10 |
|---|---|
| [CNN] ResNet50 Implementation(Keras) (0) | 2021.08.09 |
| [CNN] DenseNet 요약 (0) | 2021.08.04 |
| [CNN] ResNet 요약 (0) | 2021.08.03 |
| [CNN] GoogLeNet 요약 (0) | 2021.08.02 |
댓글