人工智能深度学习：如何实现CNN变体网络？

时间：2020-06-29 09:45:56 阅读：46 评论：0 收藏：0 [点我收藏+]

标签：简单的 empty add des sof normal cross 结构网络

1.载入数据

(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
x_train = x_train.reshape((-1,28,28,1))
x_test = x_test.reshape((-1,28,28,1))
print(x_train.shape, ‘ ‘, y_train.shape)
print(x_test.shape, ‘ ‘, y_test.shape)
(60000, 28, 28, 1)   (60000,)
(10000, 28, 28, 1)   (10000,)

2.简单的深度网络

如AlexNet,VggNet

x_shape  = x_train.shape
deep_model = keras.Sequential(
[
    layers.Conv2D(input_shape=((x_shape[1], x_shape[2], x_shape[3])),
                 filters=32, kernel_size=(3,3), strides=(1,1), padding=‘same‘, activation=‘relu‘),
    layers.Conv2D(filters=32, kernel_size=(3,3), strides=(1,1), padding=‘same‘, activation=‘relu‘),
    layers.MaxPool2D(pool_size=(2,2)),
    layers.Conv2D(filters=32, kernel_size=(3,3), strides=(1,1), padding=‘same‘, activation=‘relu‘),
    layers.Conv2D(filters=32, kernel_size=(3,3), strides=(1,1), padding=‘same‘, activation=‘relu‘),
    layers.MaxPool2D(pool_size=(2,2)),
    layers.Flatten(),
    layers.Dense(32, activation=‘relu‘),
    layers.Dense(10, activation=‘softmax‘)

])
deep_model.compile(optimizer=keras.optimizers.Adam(),
             loss=keras.losses.SparseCategoricalCrossentropy(),
            metrics=[‘accuracy‘])
deep_model.summary()
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 28, 28, 32)        320       
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 28, 28, 32)        9248      
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 14, 14, 32)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 14, 14, 32)        9248      
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 14, 14, 32)        9248      
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 7, 7, 32)          0         
_________________________________________________________________
flatten (Flatten)            (None, 1568)              0         
_________________________________________________________________
dense (Dense)                (None, 32)                50208     
_________________________________________________________________
dense_1 (Dense)              (None, 10)                330       
=================================================================
Total params: 78,602
Trainable params: 78,602
Non-trainable params: 0
_________________________________________________________________
history = deep_model.fit(x_train, y_train, batch_size=64, epochs=5, validation_split=0.1)


[0.04454196666887728, 0.9863]
import matplotlib.pyplot as plt
plt.plot(history.history[‘accuracy‘])
plt.plot(history.history[‘val_accuracy‘])
plt.legend([‘training‘, ‘valivation‘], loc=‘upper left‘)
plt.show()

技术图片

result = deep_model.evaluate(x_test, y_test)
10000/10000 [==============================] - 2s 219us/sample - loss: 0.0445 - accuracy: 0.9863

3.添加了其它功能层的深度卷积

x_shape  = x_train.shape
deep_model = keras.Sequential(
[
    layers.Conv2D(input_shape=((x_shape[1], x_shape[2], x_shape[3])),
                 filters=32, kernel_size=(3,3), strides=(1,1), padding=‘same‘, activation=‘relu‘),
    layers.BatchNormalization(),
    layers.Conv2D(filters=32, kernel_size=(3,3), strides=(1,1), padding=‘same‘, activation=‘relu‘),
    layers.BatchNormalization(),
    layers.MaxPool2D(pool_size=(2,2)),
    layers.Conv2D(filters=32, kernel_size=(3,3), strides=(1,1), padding=‘same‘, activation=‘relu‘),
    layers.BatchNormalization(),
    layers.BatchNormalization(),
    layers.Conv2D(filters=32, kernel_size=(3,3), strides=(1,1), padding=‘same‘, activation=‘relu‘),
    layers.MaxPool2D(pool_size=(2,2)),
    layers.Flatten(),
    layers.Dense(32, activation=‘relu‘),
    layers.Dropout(0.2),
    layers.Dense(10, activation=‘softmax‘)

])
deep_model.compile(optimizer=keras.optimizers.Adam(),
             loss=keras.losses.SparseCategoricalCrossentropy(),
            metrics=[‘accuracy‘])
deep_model.summary()
Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_4 (Conv2D)            (None, 28, 28, 32)        320       
_________________________________________________________________
batch_normalization_v2 (Batc (None, 28, 28, 32)        128       
_________________________________________________________________
conv2d_5 (Conv2D)            (None, 28, 28, 32)        9248      
_________________________________________________________________
batch_normalization_v2_1 (Ba (None, 28, 28, 32)        128       
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 14, 14, 32)        0         
_________________________________________________________________
conv2d_6 (Conv2D)            (None, 14, 14, 32)        9248      
_________________________________________________________________
batch_normalization_v2_2 (Ba (None, 14, 14, 32)        128       
_________________________________________________________________
batch_normalization_v2_3 (Ba (None, 14, 14, 32)        128       
_________________________________________________________________
conv2d_7 (Conv2D)            (None, 14, 14, 32)        9248      
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 7, 7, 32)          0         
_________________________________________________________________
flatten_1 (Flatten)          (None, 1568)              0         
_________________________________________________________________
dense_2 (Dense)              (None, 32)                50208     
_________________________________________________________________
dropout (Dropout)            (None, 32)                0         
_________________________________________________________________
dense_3 (Dense)              (None, 10)                330       
=================================================================
Total params: 79,114
Trainable params: 78,858
Non-trainable params: 256
_________________________________________________________________
history = deep_model.fit(x_train, y_train, batch_size=64, epochs=5, validation_split=0.1)
Train on 54000 samples, validate on 6000 samples
Epoch 1/5
54000/54000 [==============================] - 120s 2ms/sample - loss: 0.2683 - accuracy: 0.9163 - val_loss: 0.0470 - val_accuracy: 0.9880
...
Epoch 5/5
54000/54000 [==============================] - 115s 2ms/sample - loss: 0.0504 - accuracy: 0.9839 - val_loss: 0.0315 - val_accuracy: 0.9922
import matplotlib.pyplot as plt
plt.plot(history.history[‘accuracy‘])
plt.plot(history.history[‘val_accuracy‘])
plt.legend([‘training‘, ‘valivation‘], loc=‘upper left‘)
plt.show()

技术图片

result = deep_model.evaluate(x_test, y_test)
10000/10000 [==============================] - 4s 365us/sample - loss: 0.0288 - accuracy: 0.9909

4.NIN网络

Min等人在 2013年（https://arxiv.org/abs/1312.4400）提出了减少模型中参数数量的方法之一即“网络中的网络（NIN）”或“1X1卷积” 方法很简单 - 在其他卷积层之后添加卷积层具有降低图像空间的维度（深度）的效果，有效地减少了参数的数量

GoogleNet 中就用到了NIN结构