devbox
盐析白兔
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

【神经网络】(12) MobileNetV2 代码复现，网络解析，附Tensorflow完整代码_mobilenetv2代码

作者：盐析白兔 | 2024-06-04 17:46:06

赞

踩

mobilenetv2代码

各位同学好，今天和大家分享一下如何使用 Tensorflow 复现谷歌轻量化神经网络 MobileNetV2。

在上一篇中我介绍了MobileNetV1，探讨了深度可分离卷积，感兴趣的可以看一下：https://blog.csdn.net/dgvv4/article/details/123415708，本节还会继续用到深度可分离卷积的知识。那我们开始吧。

1. MobileNetV1的局限

上一节中我们认识到，深度可分离卷积能显著地减少参数量和计算量，占用更少的内存。

但是，由于MobileNet研发的时间比较早，没有和许多前沿的技术相结合，比如没有使用残差连接。

此外，网络训练时，很多深度可分离卷积核训练出来的卷积核权重是0，（1）因为卷积核权重数量和通道数太少，卷积核只有一个通道，太单薄；（2）rule激活函数会将所有小于0的输入都变成0，导致前向传播的张量为0，反向传播的梯度也为0，权重永远是0，梯度消失；（3）低精度的浮点数表示。如果使用16bit或8bit等很少的bit表示一个数，表示的范围很有限，导致训练出来的是一个非常小的数。

2. MobileNetV2网络

2.1 改进的深度可分离卷积

MobileNetV1 先使用3*3的深度卷积再使用1*1的逐点卷积，全部采用relu6激活函数。

MobileNetV2 先使用1*1卷积升维，在高维空间下使用3*3的深度卷积，再用1*1卷积降维，在降维时采用线性激活函数。当步长为1时，残差连接输入和输出；当步长为2时，不使用残差连接，因为为此时的输出和输出特征图的size不相同。

将 MobileNetV2 深度可分离卷积 和 RestNet 残差模块 做对比。ResNet 先使用1*1卷积降维，在降维空间下使用标准卷积，再用1*1卷积升维，残差连接两个高维的部分。

2.2 ReLu6激活函数

ReLu6激活函数是，当x>6时 将y值限制在6，那么relu的值域为[0,6]。因为在低精度浮点数表示下，如8bit或16bit表示一个数，表示的数范围比较小，而如果使用传统的relu激活函数，函数值域是[0,∞]，这时用比较少的bit数无法表示很大的数。因此用relu6函数，在低精度浮点数下有比较好的表示性能

2.3 总体流程

如下图，输入图像，先使用1*1逐点卷积升维（增加通道数）；然后使用3*3的深度卷积；再使用1*1逐点卷积降维，使用线性激活函数。通过控制1*1卷积核个数来控制输出特征图的通道数。

举个例子如下图：

输出的图像shape为5*5*64；先通过1*1*(64*6)的逐点卷积升维，将通道数提升6倍，使用relu6激活函数；在384个通道数的高维空间下使用3*3*1的卷积核进行深度卷积，生成特征图shape为3*3*384；再使用1*1*128的逐点卷积降维，线性激活函数；输出特征图shape为3*3*128；残差连接输入和输出

3. 代码复现

论文中给出的MobileNetV2网络结构如下图所示。Conv2d代表标准卷积；bottleneck代表残差连接优化后的深度可分离卷积（inverted residual block）；avgpool代表全局平均池化；t 列代表1*1卷积升维倍数（特征图通道数提升多少倍）；c 列代表1*1卷积降维的通道数；n 列代表该网络层重复次数；s列代表步长。

3.1 构建残差模块

残差连接的深度可分离卷积模块是由四部分构成。下面代码中的第（1）部分的①代表是1*1卷积在通道维度上升维，block_id用来判断是否需要升维，如上面的结构图，网络的第一个bottleneck就不需要升维操作，让它的block_id=0，直接进入深度卷积操作。第②部分是深度卷积操作。第③部分是1*1卷积降维。第④部分shotcut首尾短接。如果步长为1且输入和输出特征图的shape相同，那就使用残差连接，将输出输出特征图对应像素值相加。

对深度可分离卷积有疑问的可见我的上一篇文章：【神经网络】(11) 轻量化网络MobileNetV1代码复现、解析，附Tensorflow完整代码

  1. # 定义函数，使超参数处理过后的卷积核个数满足网络要求，被8整除
  2. def make_divisible(value, divisor):
  3.  
  4.     # 与value最接近的8的倍数。若value=30，那么处理后的卷积核=((30+4)//8)*8=32
  5.     new_value = max(divisor, int(value + divisor/2) // divisor * divisor)
  6.  
  7.     # 如果计算出的结果比原值要小，那就再加上一个倍数
  8.     if new_value < 0.9 * value:
  9.         new_value += divisor
  10.  
  11.     # 返回新的卷积核个数
  12.     return new_value
  13.  
  14.  
  15. #（1）残差连接深度可分离卷积块
  16. # expension代表升维的倍数，超参数alpha控制卷积核个数，block_id控制是否需要升维
  17. def inverted_res_block(input_tensor, expansion, stride, alpha, filters, block_id):
  18.  
  19.     x = input_tensor  # 先用x表示输入，如果不执行①的话，那么x没机会接收input_tensor
  20.  
  21.     # 超参数控制逐点卷积的卷积核个数，使用int()函数取整，防止alpha是小数
  22.     points_filters = int(filters*alpha)
  23.  
  24.     # 由于网络结构图中特征图的个数为32、16、24、64...都是8的倍数，
  25.     # 保证超参数处理过后的卷积核个数能满足网络特征图个数的要求，因此需要保证逐点卷积的卷积核个数为8的倍数
  26.     points_filters = make_divisible(points_filters, 8)
  27.  
  28.     # keras.backend.int_shape得到图像的shape，这里只需要最后一个维度即通道维度的大小
  29.     in_channels = keras.backend.int_shape(input_tensor)[-1]
  30.  
  31.     # ① 在通道维度上升维
  32.     if block_id:
  33.  
  34.         # 1*1卷积提升通道数，
  35.         x = layers.Conv2D(expansion * in_channels,  # 卷积核个数，通道维度提升expansion倍 
  36.                           kernel_size = (1,1),  # 卷积核大小1*1
  37.                           strides = (1,1),  # 步长1*1
  38.                           padding = 'same',  # 卷积过程中特征图size不变
  39.                           use_bias = False)(input_tensor)  # 有BN层就不需要计算偏置
  40.  
  41.         x = layers.BatchNormalization()(x)  # 批标准化
  42.  
  43.         x = layers.ReLU(6.0)(x)  # relu6激活函数
  44.     
  45.     # ② 深度可分离卷积
  46.     x = layers.DepthwiseConv2D(kernel_size=(3,3),  # 卷积核个数默认3*3
  47.                                strides=stride,  # 步长
  48.                                use_bias=False,  # 有BN层就不需要偏置
  49.                                padding = 'same')(x)  # 卷积过程中，步长=1，size不变；步长=2，size减半
  50.  
  51.     x = layers.BatchNormalization()(x)  # 批标准化
  52.  
  53.     x = layers.ReLU(6.0)(x)  # relu6激活函数
  54.  
  55.     # ③ 1*1卷积下降通道数（压缩特征），不使用relu激活函数（保证特征不被破坏）
  56.     x = layers.Conv2D(points_filters,  # 卷积核个数（即通道数）下降到起始状态
  57.                       kernel_size = (1,1),  # 卷积核大小为1*1
  58.                       strides = (1,1),  # 步长为1*1
  59.                       padding = 'same',  # 卷积过程中特征图size不变
  60.                       use_bias = False)(x)  # 有BN层就不需要偏置
  61.     
  62.     x = layers.BatchNormalization()(x)  # 批标准化
  63.  
  64.     # ④ 残差连接，输入和输出短接
  65.     # 步长为1，且输入特征图的shape和输出特征图的shape相同时，残差连接
  66.     if stride == 1 and in_channels == points_filters:
  67.  
  68.         # 将输入和输出的tensor的对应元素值相加，要求shape完全一致
  69.         x = layers.Add()([input_tensor, x])
  70.         return x
  71.     
  72.     # 步长为2，不需要进行短接
  73.     return x

3.2 MobileNetV2 网络完整代码

网络核心部分就是残差连接优化的深度可分离卷积，inverted_res_block()模块，上面已经解释过了。其次网络涉及两个超参数，alpha：宽度超参数，控制卷积核个数； depth_multiplier：分辨率超参数，控制输入图像的尺寸，进而控制中间层特征图的大小。具体可以看我的前一篇MobileNetV1的文章（文章开头有链接），这里就不作过多解释了。

其他就是基本的网络层的堆叠，每一行都有注释，有疑问的可以在评论区留言。

  1. import tensorflow as tf
  2. from tensorflow import keras
  3. from tensorflow.keras import layers, Model
  4.  
  5. # 定义函数，使超参数处理过后的卷积核个数满足网络要求，被8整除
  6. def make_divisible(value, divisor):
  7.  
  8.     # 与value最接近的8的倍数。若value=30，那么处理后的卷积核=((30+4)//8)*8=32
  9.     new_value = max(divisor, int(value + divisor/2) // divisor * divisor)
  10.  
  11.     # 如果计算出的结果比原值要小，那就再加上一个倍数
  12.     if new_value < 0.9 * value:
  13.         new_value += divisor
  14.  
  15.     # 返回新的卷积核个数
  16.     return new_value
  17.  
  18.  
  19. #（1）残差连接深度可分离卷积块
  20. # expension代表升维的倍数，超参数alpha控制卷积核个数，block_id控制是否需要升维
  21. def inverted_res_block(input_tensor, expansion, stride, alpha, filters, block_id):
  22.  
  23.     x = input_tensor  # 先用x表示输入，如果不执行①的话，那么x没机会接收input_tensor
  24.  
  25.     # 超参数控制逐点卷积的卷积核个数，使用int()函数取整，防止alpha是小数
  26.     points_filters = int(filters*alpha)
  27.  
  28.     # 由于网络结构图中特征图的个数为32、16、24、64...都是8的倍数，
  29.     # 保证超参数处理过后的卷积核个数能满足网络特征图个数的要求，因此需要保证逐点卷积的卷积核个数为8的倍数
  30.     points_filters = make_divisible(points_filters, 8)
  31.  
  32.     # keras.backend.int_shape得到图像的shape，这里只需要最后一个维度即通道维度的大小
  33.     in_channels = keras.backend.int_shape(input_tensor)[-1]
  34.  
  35.     # ① 在通道维度上升维
  36.     if block_id:
  37.  
  38.         # 1*1卷积提升通道数，
  39.         x = layers.Conv2D(expansion * in_channels,  # 卷积核个数，通道维度提升expansion倍 
  40.                           kernel_size = (1,1),  # 卷积核大小1*1
  41.                           strides = (1,1),  # 步长1*1
  42.                           padding = 'same',  # 卷积过程中特征图size不变
  43.                           use_bias = False)(input_tensor)  # 有BN层就不需要计算偏置
  44.  
  45.         x = layers.BatchNormalization()(x)  # 批标准化
  46.  
  47.         x = layers.ReLU(6.0)(x)  # relu6激活函数
  48.     
  49.     # ② 深度可分离卷积
  50.     x = layers.DepthwiseConv2D(kernel_size=(3,3),  # 卷积核个数默认3*3
  51.                                strides=stride,  # 步长
  52.                                use_bias=False,  # 有BN层就不需要偏置
  53.                                padding = 'same')(x)  # 卷积过程中，步长=1，size不变；步长=2，size减半
  54.  
  55.     x = layers.BatchNormalization()(x)  # 批标准化
  56.  
  57.     x = layers.ReLU(6.0)(x)  # relu6激活函数
  58.  
  59.     # ③ 1*1卷积下降通道数（压缩特征），不使用relu激活函数（保证特征不被破坏）
  60.     x = layers.Conv2D(points_filters,  # 卷积核个数（即通道数）下降到起始状态
  61.                       kernel_size = (1,1),  # 卷积核大小为1*1
  62.                       strides = (1,1),  # 步长为1*1
  63.                       padding = 'same',  # 卷积过程中特征图size不变
  64.                       use_bias = False)(x)  # 有BN层就不需要偏置
  65.     
  66.     x = layers.BatchNormalization()(x)  # 批标准化
  67.  
  68.     # ④ 残差连接，输入和输出短接
  69.     # 步长为1，且输入特征图的shape和输出特征图的shape相同时，残差连接
  70.     if stride == 1 and in_channels == points_filters:
  71.  
  72.         # 将输入和输出的tensor的对应元素值相加，要求shape完全一致
  73.         x = layers.Add()([input_tensor, x])
  74.         return x
  75.     
  76.     # 步长为2，不需要进行短接
  77.     return x
  78.  
  79.  
  80. #（2）网络模型结构
  81. # 传入classes分类数，输入图片大小，超参数alpha默认=1
  82. def MobileNetV2(classes, input_shape, alpha=1.0):
  83.  
  84.     # 规定超参数的取值范围
  85.     if alpha not in [0.5, 0.75, 1.0, 1.3]:
  86.         raise ValueError('alpha should use 0.5, 0.75, 1.0, 1.3')
  87.  
  88.     # 超参数控制卷积核个数，先处理第一次卷积的卷积核个数（32*alpha）使它能被8整除
  89.     first_conv_filters = make_divisible(32*alpha, 8)
  90.  
  91.     # 构建网络输入图像
  92.     inputs = keras.Input(shape=input_shape)
  93.  
  94.     # [224,224,3]==>[112,112,32]
  95.     x = layers.Conv2D(first_conv_filters,  # 卷积核个数
  96.                       kernel_size = (3,3),  # 卷积核大小3*3
  97.                       strides = (2,2),  # 步长=2，压缩特征图size
  98.                       padding = 'same',  # 步长=2，卷积过程中size减半
  99.                       use_bias = 'False')(inputs)  # 有BN层就不需要偏置
  100.  
  101.     x = layers.BatchNormalization()(x)  # 批标准化
  102.  
  103.     x = layers.ReLU(6.0)(x)  # relu6激活函数
  104.  
  105.     # [112, 112, 32]==>[112,112,16]
  106.     # 第一个残差卷积由于expansion=1，不用上升维度，因此不用执行第①步的标准卷积操作，所以block_id=0
  107.     x = inverted_res_block(x, expansion=1, stride=1, alpha=alpha, filters=16, block_id=0)
  108.  
  109.     # [112,112,16]==>[56,56,24]
  110.     x = inverted_res_block(x, expansion=6, stride=2, alpha=alpha, filters=24, block_id=1)
  111.     # [56,56,24]==>[56,56,24]
  112.     x = inverted_res_block(x, expansion=6, stride=1, alpha=alpha, filters=24, block_id=1)
  113.  
  114.     # [56,56,24]==>[28,28,32]
  115.     x = inverted_res_block(x, expansion=6, stride=2, alpha=alpha, filters=32, block_id=1)
  116.     # [28,28,32]==>[28,28,32]
  117.     x = inverted_res_block(x, expansion=6, stride=1, alpha=alpha, filters=32, block_id=1)
  118.     x = inverted_res_block(x, expansion=6, stride=1, alpha=alpha, filters=32, block_id=1)
  119.  
  120.     # [28,28,32]==>[14,14,64]
  121.     x = inverted_res_block(x, expansion=6, stride=2, alpha=alpha, filters=64, block_id=1)
  122.     # [14,14,64]==>[14,14,64]
  123.     x = inverted_res_block(x, expansion=6, stride=1, alpha=alpha, filters=64, block_id=1)
  124.     x = inverted_res_block(x, expansion=6, stride=1, alpha=alpha, filters=64, block_id=1)
  125.     x = inverted_res_block(x, expansion=6, stride=1, alpha=alpha, filters=64, block_id=1)
  126.  
  127.     # [14,14,64]==>[14,14,96]
  128.     x = inverted_res_block(x, expansion=6, stride=1, alpha=alpha, filters=96, block_id=1)
  129.     x = inverted_res_block(x, expansion=6, stride=1, alpha=alpha, filters=96, block_id=1)
  130.     x = inverted_res_block(x, expansion=6, stride=1, alpha=alpha, filters=96, block_id=1)
  131.  
  132.     # [14,14,96]==>[7,7,160]
  133.     x = inverted_res_block(x, expansion=6, stride=2, alpha=alpha, filters=160, block_id=1)
  134.     # [7,7,160]==>[7,7,160]
  135.     x = inverted_res_block(x, expansion=6, stride=1, alpha=alpha, filters=160, block_id=1)
  136.     x = inverted_res_block(x, expansion=6, stride=1, alpha=alpha, filters=160, block_id=1)
  137.  
  138.     # [7,7,160]==>[7,7,320]
  139.     x = inverted_res_block(x, expansion=1, stride=1, alpha=alpha, filters=320, block_id=1)
  140.  
  141.     # [7,7,320]==>[7,7,1280]
  142.     x = layers.Conv2D(filters=1280, kernel_size=(1,1), strides=1, padding='same', use_bias=False)(x)  # 普通卷积
  143.     x = layers.BatchNormalization()(x)  # 批标准化
  144.     x = layers.ReLU(6.0)(x)  # relu6激活函数
  145.  
  146.     # [7,7,1280]==>[None,1280]
  147.     x = layers.GlobalAveragePooling2D()(x)  # 通道维度上进行全剧平均池化
  148.  
  149.     # [None,1280]==>[1,1,1280]
  150.     # 超参数调整卷积核个数
  151.     shape = (1,1,make_divisible(1280*alpha, 8))
  152.     # 调整输出特征图的shape
  153.     x = layers.Reshape(target_shape=shape)(x)
  154.  
  155.     # Dropout层随机杀死神经元，防止过拟合
  156.     x = layers.Dropout(rate=1e-3)(x)
  157.  
  158.     # [1,1,1280]==>[1,1,1000]
  159.     # 卷积层，将特征图x的个数转换成分类数
  160.     x = layers.Conv2D(classes, kernel_size=(1,1), padding='same')(x)
  161.  
  162.     # 经过softmax函数，变成分类概率
  163.     x = layers.Activation('softmax')(x)
  164.  
  165.     # [1,1,1000]==>[None,1000]
  166.     # 重塑概率数排列形式
  167.     x = layers.Reshape(target_shape=(classes,))(x)
  168.  
  169.     # 构建网络
  170.     model = Model(inputs, x)
  171.  
  172.     return model # 返回网络模型
  173.  
  174. if __name__ == '__main__':
  175.  
  176.     # 获得网络模型结构
  177.     model = MobileNetV2(classes=1000,  # 1000个分类
  178.                         input_shape=[224,224,3],  # 网络输入的shape
  179.                         alpha=1.0)  # 超参数控制卷积核个数
  180.  
  181.     # 查看网络模型结构
  182.     model.summary()

3.3 查看网络结构

通过model.summary()查看网络模型结构，可见MobileNetV2的参数量只有三百万个，相比于MobileNetV1四百万的参数量，模型变得更加轻量化。

  1. Model: "model"
  2. __________________________________________________________________________________________________
  3. Layer (type)                    Output Shape         Param #     Connected to                     
  4. ==================================================================================================
  5. input_1 (InputLayer)            [(None, 224, 224, 3) 0                                            
  6. __________________________________________________________________________________________________
  7. conv2d (Conv2D)                 (None, 112, 112, 32) 896         input_1[0][0]                    
  8. __________________________________________________________________________________________________
  9. batch_normalization (BatchNorma (None, 112, 112, 32) 128         conv2d[0][0]                     
  10. __________________________________________________________________________________________________
  11. re_lu (ReLU)                    (None, 112, 112, 32) 0           batch_normalization[0][0]        
  12. __________________________________________________________________________________________________
  13. depthwise_conv2d (DepthwiseConv (None, 112, 112, 32) 288         re_lu[0][0]                      
  14. __________________________________________________________________________________________________
  15. batch_normalization_1 (BatchNor (None, 112, 112, 32) 128         depthwise_conv2d[0][0]           
  16. __________________________________________________________________________________________________
  17. re_lu_1 (ReLU)                  (None, 112, 112, 32) 0           batch_normalization_1[0][0]      
  18. __________________________________________________________________________________________________
  19. conv2d_1 (Conv2D)               (None, 112, 112, 16) 512         re_lu_1[0][0]                    
  20. __________________________________________________________________________________________________
  21. batch_normalization_2 (BatchNor (None, 112, 112, 16) 64          conv2d_1[0][0]                   
  22. __________________________________________________________________________________________________
  23. conv2d_2 (Conv2D)               (None, 112, 112, 96) 1536        batch_normalization_2[0][0]      
  24. __________________________________________________________________________________________________
  25. batch_normalization_3 (BatchNor (None, 112, 112, 96) 384         conv2d_2[0][0]                   
  26. __________________________________________________________________________________________________
  27. re_lu_2 (ReLU)                  (None, 112, 112, 96) 0           batch_normalization_3[0][0]      
  28. __________________________________________________________________________________________________
  29. depthwise_conv2d_1 (DepthwiseCo (None, 56, 56, 96)   864         re_lu_2[0][0]                    
  30. __________________________________________________________________________________________________
  31. batch_normalization_4 (BatchNor (None, 56, 56, 96)   384         depthwise_conv2d_1[0][0]         
  32. __________________________________________________________________________________________________
  33. re_lu_3 (ReLU)                  (None, 56, 56, 96)   0           batch_normalization_4[0][0]      
  34. __________________________________________________________________________________________________
  35. conv2d_3 (Conv2D)               (None, 56, 56, 24)   2304        re_lu_3[0][0]                    
  36. __________________________________________________________________________________________________
  37. batch_normalization_5 (BatchNor (None, 56, 56, 24)   96          conv2d_3[0][0]                   
  38. __________________________________________________________________________________________________
  39. conv2d_4 (Conv2D)               (None, 56, 56, 144)  3456        batch_normalization_5[0][0]      
  40. __________________________________________________________________________________________________
  41. batch_normalization_6 (BatchNor (None, 56, 56, 144)  576         conv2d_4[0][0]                   
  42. __________________________________________________________________________________________________
  43. re_lu_4 (ReLU)                  (None, 56, 56, 144)  0           batch_normalization_6[0][0]      
  44. __________________________________________________________________________________________________
  45. depthwise_conv2d_2 (DepthwiseCo (None, 56, 56, 144)  1296        re_lu_4[0][0]                    
  46. __________________________________________________________________________________________________
  47. batch_normalization_7 (BatchNor (None, 56, 56, 144)  576         depthwise_conv2d_2[0][0]         
  48. __________________________________________________________________________________________________
  49. re_lu_5 (ReLU)                  (None, 56, 56, 144)  0           batch_normalization_7[0][0]      
  50. __________________________________________________________________________________________________
  51. conv2d_5 (Conv2D)               (None, 56, 56, 24)   3456        re_lu_5[0][0]                    
  52. __________________________________________________________________________________________________
  53. batch_normalization_8 (BatchNor (None, 56, 56, 24)   96          conv2d_5[0][0]                   
  54. __________________________________________________________________________________________________
  55. add (Add)                       (None, 56, 56, 24)   0           batch_normalization_5[0][0]      
  56.                                                                  batch_normalization_8[0][0]      
  57. __________________________________________________________________________________________________
  58. conv2d_6 (Conv2D)               (None, 56, 56, 144)  3456        add[0][0]                        
  59. __________________________________________________________________________________________________
  60. batch_normalization_9 (BatchNor (None, 56, 56, 144)  576         conv2d_6[0][0]                   
  61. __________________________________________________________________________________________________
  62. re_lu_6 (ReLU)                  (None, 56, 56, 144)  0           batch_normalization_9[0][0]      
  63. __________________________________________________________________________________________________
  64. depthwise_conv2d_3 (DepthwiseCo (None, 28, 28, 144)  1296        re_lu_6[0][0]                    
  65. __________________________________________________________________________________________________
  66. batch_normalization_10 (BatchNo (None, 28, 28, 144)  576         depthwise_conv2d_3[0][0]         
  67. __________________________________________________________________________________________________
  68. re_lu_7 (ReLU)                  (None, 28, 28, 144)  0           batch_normalization_10[0][0]     
  69. __________________________________________________________________________________________________
  70. conv2d_7 (Conv2D)               (None, 28, 28, 32)   4608        re_lu_7[0][0]                    
  71. __________________________________________________________________________________________________
  72. batch_normalization_11 (BatchNo (None, 28, 28, 32)   128         conv2d_7[0][0]                   
  73. __________________________________________________________________________________________________
  74. conv2d_8 (Conv2D)               (None, 28, 28, 192)  6144        batch_normalization_11[0][0]     
  75. __________________________________________________________________________________________________
  76. batch_normalization_12 (BatchNo (None, 28, 28, 192)  768         conv2d_8[0][0]                   
  77. __________________________________________________________________________________________________
  78. re_lu_8 (ReLU)                  (None, 28, 28, 192)  0           batch_normalization_12[0][0]     
  79. __________________________________________________________________________________________________
  80. depthwise_conv2d_4 (DepthwiseCo (None, 28, 28, 192)  1728        re_lu_8[0][0]                    
  81. __________________________________________________________________________________________________
  82. batch_normalization_13 (BatchNo (None, 28, 28, 192)  768         depthwise_conv2d_4[0][0]         
  83. __________________________________________________________________________________________________
  84. re_lu_9 (ReLU)                  (None, 28, 28, 192)  0           batch_normalization_13[0][0]     
  85. __________________________________________________________________________________________________
  86. conv2d_9 (Conv2D)               (None, 28, 28, 32)   6144        re_lu_9[0][0]                    
  87. __________________________________________________________________________________________________
  88. batch_normalization_14 (BatchNo (None, 28, 28, 32)   128         conv2d_9[0][0]                   
  89. __________________________________________________________________________________________________
  90. add_1 (Add)                     (None, 28, 28, 32)   0           batch_normalization_11[0][0]     
  91.                                                                  batch_normalization_14[0][0]     
  92. __________________________________________________________________________________________________
  93. conv2d_10 (Conv2D)              (None, 28, 28, 192)  6144        add_1[0][0]                      
  94. __________________________________________________________________________________________________
  95. batch_normalization_15 (BatchNo (None, 28, 28, 192)  768         conv2d_10[0][0]                  
  96. __________________________________________________________________________________________________
  97. re_lu_10 (ReLU)                 (None, 28, 28, 192)  0           batch_normalization_15[0][0]     
  98. __________________________________________________________________________________________________
  99. depthwise_conv2d_5 (DepthwiseCo (None, 28, 28, 192)  1728        re_lu_10[0][0]                   
  100. __________________________________________________________________________________________________
  101. batch_normalization_16 (BatchNo (None, 28, 28, 192)  768         depthwise_conv2d_5[0][0]         
  102. __________________________________________________________________________________________________
  103. re_lu_11 (ReLU)                 (None, 28, 28, 192)  0           batch_normalization_16[0][0]     
  104. __________________________________________________________________________________________________
  105. conv2d_11 (Conv2D)              (None, 28, 28, 32)   6144        re_lu_11[0][0]                   
  106. __________________________________________________________________________________________________
  107. batch_normalization_17 (BatchNo (None, 28, 28, 32)   128         conv2d_11[0][0]                  
  108. __________________________________________________________________________________________________
  109. add_2 (Add)                     (None, 28, 28, 32)   0           add_1[0][0]                      
  110.                                                                  batch_normalization_17[0][0]     
  111. __________________________________________________________________________________________________
  112. conv2d_12 (Conv2D)              (None, 28, 28, 192)  6144        add_2[0][0]                      
  113. __________________________________________________________________________________________________
  114. batch_normalization_18 (BatchNo (None, 28, 28, 192)  768         conv2d_12[0][0]                  
  115. __________________________________________________________________________________________________
  116. re_lu_12 (ReLU)                 (None, 28, 28, 192)  0           batch_normalization_18[0][0]     
  117. __________________________________________________________________________________________________
  118. depthwise_conv2d_6 (DepthwiseCo (None, 14, 14, 192)  1728        re_lu_12[0][0]                   
  119. __________________________________________________________________________________________________
  120. batch_normalization_19 (BatchNo (None, 14, 14, 192)  768         depthwise_conv2d_6[0][0]         
  121. __________________________________________________________________________________________________
  122. re_lu_13 (ReLU)                 (None, 14, 14, 192)  0           batch_normalization_19[0][0]     
  123. __________________________________________________________________________________________________
  124. conv2d_13 (Conv2D)              (None, 14, 14, 64)   12288       re_lu_13[0][0]                   
  125. __________________________________________________________________________________________________
  126. batch_normalization_20 (BatchNo (None, 14, 14, 64)   256         conv2d_13[0][0]                  
  127. __________________________________________________________________________________________________
  128. conv2d_14 (Conv2D)              (None, 14, 14, 384)  24576       batch_normalization_20[0][0]     
  129. __________________________________________________________________________________________________
  130. batch_normalization_21 (BatchNo (None, 14, 14, 384)  1536        conv2d_14[0][0]                  
  131. __________________________________________________________________________________________________
  132. re_lu_14 (ReLU)                 (None, 14, 14, 384)  0           batch_normalization_21[0][0]     
  133. __________________________________________________________________________________________________
  134. depthwise_conv2d_7 (DepthwiseCo (None, 14, 14, 384)  3456        re_lu_14[0][0]                   
  135. __________________________________________________________________________________________________
  136. batch_normalization_22 (BatchNo (None, 14, 14, 384)  1536        depthwise_conv2d_7[0][0]         
  137. __________________________________________________________________________________________________
  138. re_lu_15 (ReLU)                 (None, 14, 14, 384)  0           batch_normalization_22[0][0]     
  139. __________________________________________________________________________________________________
  140. conv2d_15 (Conv2D)              (None, 14, 14, 64)   24576       re_lu_15[0][0]                   
  141. __________________________________________________________________________________________________
  142. batch_normalization_23 (BatchNo (None, 14, 14, 64)   256         conv2d_15[0][0]                  
  143. __________________________________________________________________________________________________
  144. add_3 (Add)                     (None, 14, 14, 64)   0           batch_normalization_20[0][0]     
  145.                                                                  batch_normalization_23[0][0]     
  146. __________________________________________________________________________________________________
  147. conv2d_16 (Conv2D)              (None, 14, 14, 384)  24576       add_3[0][0]                      
  148. __________________________________________________________________________________________________
  149. batch_normalization_24 (BatchNo (None, 14, 14, 384)  1536        conv2d_16[0][0]                  
  150. __________________________________________________________________________________________________
  151. re_lu_16 (ReLU)                 (None, 14, 14, 384)  0           batch_normalization_24[0][0]     
  152. __________________________________________________________________________________________________
  153. depthwise_conv2d_8 (DepthwiseCo (None, 14, 14, 384)  3456        re_lu_16[0][0]                   
  154. __________________________________________________________________________________________________
  155. batch_normalization_25 (BatchNo (None, 14, 14, 384)  1536        depthwise_conv2d_8[0][0]         
  156. __________________________________________________________________________________________________
  157. re_lu_17 (ReLU)                 (None, 14, 14, 384)  0           batch_normalization_25[0][0]     
  158. __________________________________________________________________________________________________
  159. conv2d_17 (Conv2D)              (None, 14, 14, 64)   24576       re_lu_17[0][0]                   
  160. __________________________________________________________________________________________________
  161. batch_normalization_26 (BatchNo (None, 14, 14, 64)   256         conv2d_17[0][0]                  
  162. __________________________________________________________________________________________________
  163. add_4 (Add)                     (None, 14, 14, 64)   0           add_3[0][0]                      
  164.                                                                  batch_normalization_26[0][0]     
  165. __________________________________________________________________________________________________
  166. conv2d_18 (Conv2D)              (None, 14, 14, 384)  24576       add_4[0][0]                      
  167. __________________________________________________________________________________________________
  168. batch_normalization_27 (BatchNo (None, 14, 14, 384)  1536        conv2d_18[0][0]                  
  169. __________________________________________________________________________________________________
  170. re_lu_18 (ReLU)                 (None, 14, 14, 384)  0           batch_normalization_27[0][0]     
  171. __________________________________________________________________________________________________
  172. depthwise_conv2d_9 (DepthwiseCo (None, 14, 14, 384)  3456        re_lu_18[0][0]                   
  173. __________________________________________________________________________________________________
  174. batch_normalization_28 (BatchNo (None, 14, 14, 384)  1536        depthwise_conv2d_9[0][0]         
  175. __________________________________________________________________________________________________
  176. re_lu_19 (ReLU)                 (None, 14, 14, 384)  0           batch_normalization_28[0][0]     
  177. __________________________________________________________________________________________________
  178. conv2d_19 (Conv2D)              (None, 14, 14, 64)   24576       re_lu_19[0][0]                   
  179. __________________________________________________________________________________________________
  180. batch_normalization_29 (BatchNo (None, 14, 14, 64)   256         conv2d_19[0][0]                  
  181. __________________________________________________________________________________________________
  182. add_5 (Add)                     (None, 14, 14, 64)   0           add_4[0][0]                      
  183.                                                                  batch_normalization_29[0][0]     
  184. __________________________________________________________________________________________________
  185. conv2d_20 (Conv2D)              (None, 14, 14, 384)  24576       add_5[0][0]                      
  186. __________________________________________________________________________________________________
  187. batch_normalization_30 (BatchNo (None, 14, 14, 384)  1536        conv2d_20[0][0]                  
  188. __________________________________________________________________________________________________
  189. re_lu_20 (ReLU)                 (None, 14, 14, 384)  0           batch_normalization_30[0][0]     
  190. __________________________________________________________________________________________________
  191. depthwise_conv2d_10 (DepthwiseC (None, 14, 14, 384)  3456        re_lu_20[0][0]                   
  192. __________________________________________________________________________________________________
  193. batch_normalization_31 (BatchNo (None, 14, 14, 384)  1536        depthwise_conv2d_10[0][0]        
  194. __________________________________________________________________________________________________
  195. re_lu_21 (ReLU)                 (None, 14, 14, 384)  0           batch_normalization_31[0][0]     
  196. __________________________________________________________________________________________________
  197. conv2d_21 (Conv2D)              (None, 14, 14, 96)   36864       re_lu_21[0][0]                   
  198. __________________________________________________________________________________________________
  199. batch_normalization_32 (BatchNo (None, 14, 14, 96)   384         conv2d_21[0][0]                  
  200. __________________________________________________________________________________________________
  201. conv2d_22 (Conv2D)              (None, 14, 14, 576)  55296       batch_normalization_32[0][0]     
  202. __________________________________________________________________________________________________
  203. batch_normalization_33 (BatchNo (None, 14, 14, 576)  2304        conv2d_22[0][0]                  
  204. __________________________________________________________________________________________________
  205. re_lu_22 (ReLU)                 (None, 14, 14, 576)  0           batch_normalization_33[0][0]     
  206. __________________________________________________________________________________________________
  207. depthwise_conv2d_11 (DepthwiseC (None, 14, 14, 576)  5184        re_lu_22[0][0]                   
  208. __________________________________________________________________________________________________
  209. batch_normalization_34 (BatchNo (None, 14, 14, 576)  2304        depthwise_conv2d_11[0][0]        
  210. __________________________________________________________________________________________________
  211. re_lu_23 (ReLU)                 (None, 14, 14, 576)  0           batch_normalization_34[0][0]     
  212. __________________________________________________________________________________________________
  213. conv2d_23 (Conv2D)              (None, 14, 14, 96)   55296       re_lu_23[0][0]                   
  214. __________________________________________________________________________________________________
  215. batch_normalization_35 (BatchNo (None, 14, 14, 96)   384         conv2d_23[0][0]                  
  216. __________________________________________________________________________________________________
  217. add_6 (Add)                     (None, 14, 14, 96)   0           batch_normalization_32[0][0]     
  218.                                                                  batch_normalization_35[0][0]     
  219. __________________________________________________________________________________________________
  220. conv2d_24 (Conv2D)              (None, 14, 14, 576)  55296       add_6[0][0]                      
  221. __________________________________________________________________________________________________
  222. batch_normalization_36 (BatchNo (None, 14, 14, 576)  2304        conv2d_24[0][0]                  
  223. __________________________________________________________________________________________________
  224. re_lu_24 (ReLU)                 (None, 14, 14, 576)  0           batch_normalization_36[0][0]     
  225. __________________________________________________________________________________________________
  226. depthwise_conv2d_12 (DepthwiseC (None, 14, 14, 576)  5184        re_lu_24[0][0]                   
  227. __________________________________________________________________________________________________
  228. batch_normalization_37 (BatchNo (None, 14, 14, 576)  2304        depthwise_conv2d_12[0][0]        
  229. __________________________________________________________________________________________________
  230. re_lu_25 (ReLU)                 (None, 14, 14, 576)  0           batch_normalization_37[0][0]     
  231. __________________________________________________________________________________________________
  232. conv2d_25 (Conv2D)              (None, 14, 14, 96)   55296       re_lu_25[0][0]                   
  233. __________________________________________________________________________________________________
  234. batch_normalization_38 (BatchNo (None, 14, 14, 96)   384         conv2d_25[0][0]                  
  235. __________________________________________________________________________________________________
  236. add_7 (Add)                     (None, 14, 14, 96)   0           add_6[0][0]                      
  237.                                                                  batch_normalization_38[0][0]     
  238. __________________________________________________________________________________________________
  239. conv2d_26 (Conv2D)              (None, 14, 14, 576)  55296       add_7[0][0]                      
  240. __________________________________________________________________________________________________
  241. batch_normalization_39 (BatchNo (None, 14, 14, 576)  2304        conv2d_26[0][0]                  
  242. __________________________________________________________________________________________________
  243. re_lu_26 (ReLU)                 (None, 14, 14, 576)  0           batch_normalization_39[0][0]     
  244. __________________________________________________________________________________________________
  245. depthwise_conv2d_13 (DepthwiseC (None, 7, 7, 576)    5184        re_lu_26[0][0]                   
  246. __________________________________________________________________________________________________
  247. batch_normalization_40 (BatchNo (None, 7, 7, 576)    2304        depthwise_conv2d_13[0][0]        
  248. __________________________________________________________________________________________________
  249. re_lu_27 (ReLU)                 (None, 7, 7, 576)    0           batch_normalization_40[0][0]     
  250. __________________________________________________________________________________________________
  251. conv2d_27 (Conv2D)              (None, 7, 7, 160)    92160       re_lu_27[0][0]                   
  252. __________________________________________________________________________________________________
  253. batch_normalization_41 (BatchNo (None, 7, 7, 160)    640         conv2d_27[0][0]                  
  254. __________________________________________________________________________________________________
  255. conv2d_28 (Conv2D)              (None, 7, 7, 960)    153600      batch_normalization_41[0][0]     
  256. __________________________________________________________________________________________________
  257. batch_normalization_42 (BatchNo (None, 7, 7, 960)    3840        conv2d_28[0][0]                  
  258. __________________________________________________________________________________________________
  259. re_lu_28 (ReLU)                 (None, 7, 7, 960)    0           batch_normalization_42[0][0]     
  260. __________________________________________________________________________________________________
  261. depthwise_conv2d_14 (DepthwiseC (None, 7, 7, 960)    8640        re_lu_28[0][0]                   
  262. __________________________________________________________________________________________________
  263. batch_normalization_43 (BatchNo (None, 7, 7, 960)    3840        depthwise_conv2d_14[0][0]        
  264. __________________________________________________________________________________________________
  265. re_lu_29 (ReLU)                 (None, 7, 7, 960)    0           batch_normalization_43[0][0]     
  266. __________________________________________________________________________________________________
  267. conv2d_29 (Conv2D)              (None, 7, 7, 160)    153600      re_lu_29[0][0]                   
  268. __________________________________________________________________________________________________
  269. batch_normalization_44 (BatchNo (None, 7, 7, 160)    640         conv2d_29[0][0]                  
  270. __________________________________________________________________________________________________
  271. add_8 (Add)                     (None, 7, 7, 160)    0           batch_normalization_41[0][0]     
  272.                                                                  batch_normalization_44[0][0]     
  273. __________________________________________________________________________________________________
  274. conv2d_30 (Conv2D)              (None, 7, 7, 960)    153600      add_8[0][0]                      
  275. __________________________________________________________________________________________________
  276. batch_normalization_45 (BatchNo (None, 7, 7, 960)    3840        conv2d_30[0][0]                  
  277. __________________________________________________________________________________________________
  278. re_lu_30 (ReLU)                 (None, 7, 7, 960)    0           batch_normalization_45[0][0]     
  279. __________________________________________________________________________________________________
  280. depthwise_conv2d_15 (DepthwiseC (None, 7, 7, 960)    8640        re_lu_30[0][0]                   
  281. __________________________________________________________________________________________________
  282. batch_normalization_46 (BatchNo (None, 7, 7, 960)    3840        depthwise_conv2d_15[0][0]        
  283. __________________________________________________________________________________________________
  284. re_lu_31 (ReLU)                 (None, 7, 7, 960)    0           batch_normalization_46[0][0]     
  285. __________________________________________________________________________________________________
  286. conv2d_31 (Conv2D)              (None, 7, 7, 160)    153600      re_lu_31[0][0]                   
  287. __________________________________________________________________________________________________
  288. batch_normalization_47 (BatchNo (None, 7, 7, 160)    640         conv2d_31[0][0]                  
  289. __________________________________________________________________________________________________
  290. add_9 (Add)                     (None, 7, 7, 160)    0           add_8[0][0]                      
  291.                                                                  batch_normalization_47[0][0]     
  292. __________________________________________________________________________________________________
  293. conv2d_32 (Conv2D)              (None, 7, 7, 160)    25600       add_9[0][0]                      
  294. __________________________________________________________________________________________________
  295. batch_normalization_48 (BatchNo (None, 7, 7, 160)    640         conv2d_32[0][0]                  
  296. __________________________________________________________________________________________________
  297. re_lu_32 (ReLU)                 (None, 7, 7, 160)    0           batch_normalization_48[0][0]     
  298. __________________________________________________________________________________________________
  299. depthwise_conv2d_16 (DepthwiseC (None, 7, 7, 160)    1440        re_lu_32[0][0]                   
  300. __________________________________________________________________________________________________
  301. batch_normalization_49 (BatchNo (None, 7, 7, 160)    640         depthwise_conv2d_16[0][0]        
  302. __________________________________________________________________________________________________
  303. re_lu_33 (ReLU)                 (None, 7, 7, 160)    0           batch_normalization_49[0][0]     
  304. __________________________________________________________________________________________________
  305. conv2d_33 (Conv2D)              (None, 7, 7, 320)    51200       re_lu_33[0][0]                   
  306. __________________________________________________________________________________________________
  307. batch_normalization_50 (BatchNo (None, 7, 7, 320)    1280        conv2d_33[0][0]                  
  308. __________________________________________________________________________________________________
  309. conv2d_34 (Conv2D)              (None, 7, 7, 1280)   409600      batch_normalization_50[0][0]     
  310. __________________________________________________________________________________________________
  311. batch_normalization_51 (BatchNo (None, 7, 7, 1280)   5120        conv2d_34[0][0]                  
  312. __________________________________________________________________________________________________
  313. re_lu_34 (ReLU)                 (None, 7, 7, 1280)   0           batch_normalization_51[0][0]     
  314. __________________________________________________________________________________________________
  315. global_average_pooling2d (Globa (None, 1280)         0           re_lu_34[0][0]                   
  316. __________________________________________________________________________________________________
  317. reshape (Reshape)               (None, 1, 1, 1280)   0           global_average_pooling2d[0][0]   
  318. __________________________________________________________________________________________________
  319. dropout (Dropout)               (None, 1, 1, 1280)   0           reshape[0][0]                    
  320. __________________________________________________________________________________________________
  321. conv2d_35 (Conv2D)              (None, 1, 1, 1000)   1281000     dropout[0][0]                    
  322. __________________________________________________________________________________________________
  323. activation (Activation)         (None, 1, 1, 1000)   0           conv2d_35[0][0]                  
  324. __________________________________________________________________________________________________
  325. reshape_1 (Reshape)             (None, 1000)         0           activation[0][0]                 
  326. ==================================================================================================
  327. Total params: 3,141,416
  328. Trainable params: 3,110,504
  329. Non-trainable params: 30,912
  330. __________________________________________________________________________________________________
声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/盐析白兔/article/detail/672855
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号