使用tensorflow实现inception v3

樱花的浪漫

已于 2022-05-12 21:16:36 修改

阅读量965

点赞数 1

CC 4.0 BY-SA版权

分类专栏：图像分类文章标签： tensorflow 人工智能 python 计算机视觉深度学习

于 2022-05-07 19:24:44 首次发布

本文链接：https://blog.youkuaiyun.com/qq_52053775/article/details/124635373

图像分类专栏收录该内容

19 篇文章

订阅专栏

本文详细解读了Inception V3网络结构，包括inception_1到inception_5模块的实现，从不同分支的卷积、池化到全连接层的串联过程。核心内容涉及多尺度特征提取和并行路径设计。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

1.inception整体结构

1.figure5实现

def inception_1(x,param1,param2,param3,param4):
    # branch1 1*1-3*3-3*3
    branch1 = conv_fn(x,filters=param1[0],kernel_size=(1,1),strides=1)
    branch1 = conv_fn(branch1,filters=param1[1],kernel_size=(3,3),strides=1)
    branch1 = conv_fn(branch1,filters=param1[1],kernel_size=(3,3),strides=1)
    
    # branch2 1*1-3*3
    branch2 = conv_fn(x,filters=param2[0],kernel_size=(1,1),strides=1)
    branch2 = conv_fn(branch2,filters=param2[1],kernel_size=(3,3),strides=1)
    
    # branch3 avg_pool 3*3-1*1
    branch3 = keras.layers.AveragePooling2D(pool_size=(3,3),strides=1,padding="same")(x)
    branch3 = conv_fn(branch3,filters=param3[0],kernel_size=(1,1),strides=1)
    
    # branch4 1*1
    branch4 = conv_fn(x,filters=param4[0],kernel_size=(1,1),strides=1)
    
    return keras.layers.Concatenate(axis=3)([branch1,branch2,branch3,branch4])

2.figure5与figure6中间部分实现

def inception_2(x,param1,param2):
    # branch1 conv 3*3
    branch1 = conv_fn(x,filters=param1[0],kernel_size=(3,3),strides=2,padding="valid")
    
    # branch2 1*1-3*3-3*3
    branch2 = conv_fn(x,filters=param2[0],kernel_size=(1,1),strides=1)
    branch2 = conv_fn(branch2,filters=param2[1],kernel_size=(3,3),strides=1)
    branch2 = conv_fn(branch2,filters=param2[1],kernel_size=(3,3),strides=2,padding="valid")
    
    # branch3 max_pool 3*3 2
    branch3 = keras.layers.MaxPooling2D(pool_size=(3,3),strides=2)(x)
    
    return keras.layers.Concatenate(axis=3)([branch1,branch2,branch3])

3.figure 6实现

def inception_3(x,param1,param2,param3,param4):
    # branch1 1*1-1*7-7*1-1*7-7*1
    branch1 = conv_fn(x,filters=param1[0],kernel_size=(1,1),strides=1)
    branch1 = conv_fn(branch1,filters=param1[0],kernel_size=(1,7),strides=1)
    branch1 = conv_fn(branch1,filters=param1[0],kernel_size=(7,1),strides=1)
    branch1 = conv_fn(branch1,filters=param1[0],kernel_size=(1,7),strides=1)
    branch1 = conv_fn(branch1,filters=param1[1],kernel_size=(7,1),strides=1)
    
    # branch2 1*1-1*7-7*1
    branch2 = conv_fn(x,filters=param2[0],kernel_size=(1,1),strides=1)
    branch2 = conv_fn(branch2,filters=param2[0],kernel_size=(1,7),strides=1)
    branch2 = conv_fn(branch2,filters=param2[1],kernel_size=(7,1),strides=1)
    
    # branch3 avg_pool-conv 1*1
    branch3 = keras.layers.AveragePooling2D(pool_size=(3,3),strides=1,padding="same")(x)
    branch3 = conv_fn(branch3,param3[0],kernel_size=(1,1),strides=1)
    
    # branch4
    branch4 = conv_fn(x,param4[0],kernel_size=(1,1),strides=1)
    return keras.layers.Concatenate(axis=3)([branch1,branch2,branch3,branch4])

4.figure6与figure7连接部分实现

def inception_4(x,param1,param2):
    # branch1 1*1-3*3
    branch1 = conv_fn(x,filters=param1[0],kernel_size=(1,1),strides=1)
    branch1 = conv_fn(branch1,filters=param1[1],kernel_size=(3,3),strides=2,padding='valid')
    
    # branch2 1*1-1*7-7*1-3*3
    branch2 = conv_fn(x,filters=param2[0],kernel_size=(1,1),strides=1)
    branch2 = conv_fn(branch2,filters=param2[0],kernel_size=(1,7),strides=1)
    branch2 = conv_fn(branch2,filters=param2[0],kernel_size=(7,1),strides=1)
    branch2 = conv_fn(branch2,filters=param2[0],kernel_size=(3,3),strides=2,padding='valid')
    
    # branch3 max_pool 3*3
    branch3 = keras.layers.MaxPooling2D(pool_size=(3,3),strides=2)(x)
    return keras.layers.Concatenate(axis=3)([branch1,branch2,branch3])

5.辅助输出部分实现

def aux_logit(x,n_classes=1000):
    x = keras.layers.AveragePooling2D(pool_size=(5,5),strides=3)(x)
    x = conv_fn(x,filters=128,kernel_size=(1,1),strides=1)
    x = keras.layers.Flatten()(x)
    x = keras.layers.BatchNormalization()(x)
    aux_output = keras.layers.Dense(n_classes,activation="softmax")(x)
    return aux_output

7.figure7实现

def inception_5(x,param1,param2,param3,param4):
    # branch1 1*1
    branch1 = conv_fn(x,filters=param1[0],kernel_size=(1,1),strides=1)
    
    # branch 1*1-1*3 and 3*1
    branch2 = conv_fn(x,filters=param2[0],kernel_size=(1,1),strides=1)
    branch2_1 = conv_fn(branch2,filters=param2[1],kernel_size=(1,3),strides=1)
    branch2_2 = conv_fn(branch2,filters=param2[1],kernel_size=(3,1),strides=1)
    branch2 = keras.layers.Concatenate(axis=3)([branch2_1,branch2_2])
    
    # branch3 1*1-3*3-1*3 and 3*1
    branch3 = conv_fn(x,filters=param3[0],kernel_size=(1,1),strides=1)
    branch3 = conv_fn(branch3,filters=param3[1],kernel_size=(3,3),strides=1)
    branch3_1 = conv_fn(branch3,filters=param3[1],kernel_size=(3,1),strides=1)
    branch3_2 = conv_fn(branch3,filters=param3[1],kernel_size=(1,3),strides=1)
    branch3 = keras.layers.Concatenate(axis=3)([branch3_1,branch3_2])
    
    # branch avg_pool 3*3 - conv 1*1
    branch4 = keras.layers.AveragePooling2D(pool_size=(3,3),strides=1,padding="same")(x)
    branch4 = conv_fn(branch3,filters=param4[0],kernel_size=(1,1),strides=1)
    
    return keras.layers.Concatenate(axis=3)([branch1,branch2,branch3,branch4])

8、串联整个网络

def incepetion_v3(input_shape,n_classes=1000):
    x_input = keras.layers.Input(shape=input_shape)
    x = conv_fn(x_input,filters=32,kernel_size=(3,3),strides=2,padding="valid")
    x = conv_fn(x,filters=32,kernel_size=(3,3),strides=1,padding="valid")
    x = conv_fn(x,filters=64,kernel_size=(3,3),strides=1)
    x = keras.layers.MaxPooling2D(pool_size=(3,3),strides=2)(x)
    
    x = conv_fn(x,filters=80,kernel_size=(1,1),strides=1,padding="valid")
    x = conv_fn(x,filters=192,kernel_size=(3,3),strides=1,padding="valid")
    x = keras.layers.MaxPooling2D(pool_size=(3,3),strides=2)(x)
    
    # mixed 0: 35 x 35 x 256
    x = inception_1(x,[64,96],[48,64],[32],[64])
    
    #  mixed 1: 35 x 35 x 288
    x = inception_1(x,[64,96],[48,64],[64],[64])
    
    # mixed 2: 35 x 35 x 288
    x = inception_1(x,[64,96],[48,64],[64],[64])    
    
    # mixed 3: 17 x 17 x 768
    x = inception_2(x,[384],[64,96])
    
    # mixed 4: 17 x 17 x 768
    x = inception_3(x,[128,192],[128,192],[192],[192])
    
    # mixed 5, 6: 17 x 17 x 768
    x = inception_3(x,[160,192],[160,192],[192],[192])
    x = inception_3(x,[160,192],[160,192],[192],[192])
    
    # mixed 7: 17 x 17 x 768
    x = inception_3(x,[192,192],[192,192],[192],[192])
    
    # 辅助输出
    aux_out = aux_logit(x,n_classes)
    
    # mixed 8: 8 x 8 x 1280
    x = inception_4(x,[192,320],[192])
    
    # mixed 9、10: 8 x 8 x 2048
    x = inception_5(x,[320],[384,384],[448,384],[192])
    x = inception_5(x,[320],[384,384],[448,384],[192])
    
    # 输出层
    x = keras.layers.GlobalAveragePooling2D()(x)
    output = keras.layers.Dense(n_classes,activation="relu")(x)
    
    model = keras.models.Model(inputs=[x_input],outputs=[aux_out,output])
    return model