SENet架构-通道注意力机制

开启掘金成长之旅！这是我参与「掘金日新计划 · 2 月更文挑战」的第1天

参考论文：Squeeze-and-Excitation Networks

作者：Jie Hu，Li Shen，Samuel Albanie，Gang Sun，Enhua Wu

论文中给出的源码链接：github.com/hujie-frank…

1、SeNet简介

  SENet 是 ImageNet Challenge 图像识别比赛 2017 年的冠军，是来自 Momenta 公司 的团队完成。他们提出了 Squeeze-and-Excitation Networks（简称 SENet）。SENet 不是独立的模型设计，只对模型的一种优化。一般 SENet 都会结合其它模型一起使用，比如 SENet 用于 ResNet-50 中我们就把这个模型称为 SE-ResNet-50，比如 SENet 用于 Inception-ResNet-v2 中我们就把这个模型称为 SE- Inception-ResNet-v2。最早提出 SENet 的论文是《Squeeze-and-Excitation Networks》。

  SENet 的模型优化思路很有意思，主要是针对特征的 channel 进行优化。

  我们可以想象在进行图像识别的时候，卷积计算后生成了很多特征图，不同的滤波器会 得到不同的特征图，不同的特征图代表从图像中提取的不同的特征。我们得到了这么多的特征图，按理来说某些特征图的应该更重要，某些特征图应该没这么重要，并不是所有特征图 都一样的重要。所以 SENet 的核心思想就是给特征图增加注意力和门控机制，增强重要的特征图的信息，减弱不重要的特征图的信息。

2、注意力和门控机制

  我们看一下 SENet 的名字 Squeeze-and-Excitation Networks。其中的“Squeeze”中文意思是“挤压”，在模型中 的实际操作其实是压缩特征图的特征，作者使用的压缩特征图的特征的方式是 avg pooling 平均池化。这个大家应该很熟悉了，求一个特征图所有值的平均值，把 avg pooling 计算后的结果作为这个特征图压缩后的特征。比如一共有 64 个特征图，“Squeeze”计算后我们 就会得到 64 个值，代表 64 个特征图压缩后的特征。

  “Excitation”中文意思是“激发”，在模型中的实际操作是调节特征图信号强弱，作者 使用的方式是给“Squeeze”计算后的结果加上两个全连接层，最终输出每个特征图对应的 激活值，激活值可以改变特征图信号的强弱。每个特征图乘以它所对应的激活值，得到特征 图的输出，然后再传给下一层。

2.1 SENet block

原论文中给出的图如下所示：

具体实现如下：

• 对输入进来的特征层进行全局平均池化。
• 然后进行两次全连接(这两个全连接可用1*1卷积代替)，第一次全连接神经元个数较少，第二次全连接神经元个数和输入特征层个数相同。
• 在完成两次全连接之后，再取一次sigmoid讲值固定到0-1之间，此时我们获得了输入特征层每一个通道的权值(0-1之间)。
• 在获得这个权值之后，讲这个权值与原输入特征层相乘即可。

2.2 SENet block的Tensorflow实现

import math

import tensorflow as tf
from tensorflow.keras import backend as K
from tensorflow.keras.layers import (Activation, Add, Concatenate, Conv1D, Conv2D, Dense,
                                     GlobalAveragePooling2D, GlobalMaxPooling2D, Lambda, BatchNormalization,
                                     Reshape, multiply, Input)
from tensorflow.keras.models import Model
from plot_model import plot_model

def se_block(input_feature, ratio=16, name=""):
    # 获得通道数[h,w,c]
    channel = input_feature.shape[-1]
    # 全局平均池化 [c]
    se_feature = GlobalAveragePooling2D()(input_feature)
    # [1,1,c]
    se_feature = Reshape((1, 1, channel))(se_feature)
    # 接下来两次全连接(也可以使用两个1*1卷积)
    # 使用两个全连接
    se_feature = Dense(channel // ratio,
                       activation='relu',
                       kernel_initializer='he_normal',
                       use_bias=False,
                       name="se_block_one_" + str(name))(se_feature)

    se_feature = Dense(channel,
                       kernel_initializer='he_normal',
                       use_bias=False,
                       name="se_block_two_" + str(name))(se_feature)
    # 使用两个1*1卷积
    # se_feature = Conv2D(channel // ratio,
    #                    activation='relu',
    #                    kernel_size=[1, 1],
    #                    kernel_initializer='he_normal',
    #                    use_bias=False,
    #                    name="se_block_one_" + str(name))(se_feature)
    #
    # se_feature = Conv2D(channel,
    #                     kernel_size=[1,1],
    #                    kernel_initializer='he_normal',
    #                    use_bias=False,
    #                    name="se_block_two_" + str(name))(se_feature)

    # 使用sigmoid获得输入特征层每一个通道的权值(0-1之间)
    se_feature = Activation('sigmoid')(se_feature)
    # 将权值乘上原输入特征层即可。
    out = multiply([input_feature, se_feature])
    return out

这里可以将两个全连接层替换为两个1*1卷积，就是注释掉的那部分代码。

if __name__ == '__main__':
    inputs = Input([26, 26, 512])
    x = se_block(inputs)
    model = Model(inputs, x)
    model.summary()
    plot_model(model,to_file='img/SeNet.png',show_shapes=True)

3、Se-Inception模块与SE-ResNet模块

图 2. 原始 Inception 模块（左）和 SEInception 模块（右）的架构。

  图中的 Global pooling 表示全局池化；W 表示图片宽度；H 表示图片高度；C 表示图片 通道数；FC 表示全连接层；r 表示缩减率，意思是通道数在第一个全连接层缩减多少，总之 就是一个超参数，不用细究，一般取值为 16。

图 3. 原始残差模块（左）和 SEResNet 模块（右）的架构。

  图中的 Global pooling 表示全局池化；W 表示图片宽度；H 表示图片高度；C 表示图片 通道数；FC 表示全连接层；r 表示缩减率，意思是通道数在第一个全连接层缩减多少，总之 就是一个超参数，不用细究，一般取值为 16。

4、SE-残差结构的理解

图片来源：《深度学习从0到1》

普通的 ResNet 中的残差结构如下图所示

  图中的 Conv 表示卷积；Batch Norm 表示批量标准化；Identity mapping 表示恒等映射；batch 表示批次。

  普通的残差结构我们就不需要多说了，我前面也有论文的精读博客。下面我们看一下加上了 Squeeze-and-Excitation 模块后的残差结构如下图所示。

  加上 Squeeze-and-Excitation 模块后的残差结构主要变化是在原来的残差结构最后一个 卷积层后面进行 Squeeze-and-Excitation 的操作。Squeeze 就是先做平均池化，得到每一个特征图的压缩特征。图中特征图大小为 56×56，所以池化的窗口大小也是 56×56。池化 过后就是 Excitation 操作，前面我们有提到 Excitation 操作有两个全连接层，这是 SENet 原始论文中的做法，实际我们在写程序的时候也可以用两个窗口大小 1×1 的卷积层的替代，效果跟全连接是一样的。

  Excitation 操作部分最后的激活函数是 Sigmoid 函数，作者在这里 使用 Sigmoid 函数主要是利用 Sigmoid 函数输出范围是 0-1 这个特性，让 Excitation 的输 出激活值可以起到一个门控的作用。Excitation 的输出的激活值会乘以原始残差结构最后一 个卷积层的输出结果，对特征图的数值大小进行控制。如果是重要的特征图，会保持比较大 的数值；如果是不重要的特征图，特征图的数值就会变小。

5、ResNet-50，SE-ResNet-50，SE-ResNeXt-50(32×4d)模型结构

  （左）ResNet-50 [13]。 （中）SE-ResNet-50。 （右）带有 32×4d 模板的 SE-ResNeXt-50。括号内列出了残差构建块的特定参数设置的形状和操作，外部显示了阶段中堆叠块的数量。 fc后面的内括号表示SE模块中两个全连接层的输出维度。

   图中的 Output size 表示输出大小；conv 表示卷积；max pool 表示最大池化；stride 表示步长；fc 表示全连接层；global average pool 表示全局平均池化；C 表示分组数。

6、SeNet结果分析

  基础模型增加 SE 模块后会使得整体模型的参数增加 10%左右，计算量增加不多，一般 来说模型的效果也会有所提升。作者使用多个模型在 ImageNet 数据集上进行了测试，下图为多个模型在 ImageNet 验证集测试结果。

  图中 original 表示模型原始论文中的结果；re-implementation 表示 SENet 作者重新训 练模型的结果；SENet 表示给这些模型加上 SE 模块后的结果；top-1 err.表示 top1 错误 率；top-5 err.表示 top5 错误率；GFLOPs 表示计算量。

  图中结果可以看出，图中测试的所有模型只要加上 SE 模块，错误率都能降低，并且模型 浮点计算量没有太大变化。

  下图 也能看出加上 SE 模块后模型效果可以变得更好：

   图中的 epochs 表示周期；Top-1 error 表示 Top1 错误率；train 表示训练集；val 表示 验证集。

  SENet 论文的最后，作者还给了一组很有意思的图。作者用 ImageNet 数据集训练了一 个 SE-ResNet-50，然后选出 4 个种类(goldfish,pug,plane,cliff)的图片，统计这 4 个种类在 SE-ResNet-50 模型的每个 SE 模块的特征图的激活情况，如下图所示。

  图中的 all 表示所有 1000 个种类的平均值；goldfish 表示金鱼；pug 表示哈巴狗； plane 表示飞机；cliff 表示悬崖；channel index 表示通道；activation 表示激活值

作者观察实验结果得到 3 个结论：

• 第一，不同种类的物体在浅层激活分布情况是类似的，如图中的 SE_2_3 和 SE_3_4。也 就是不管是识别哪种物体，浅层的卷积层中，重要的特征图总是比较固定的那些。

• 第二，在更深层一些的位置，不同种类在不同的特征图激活分布不同，因为不同类别对 特征有不同的偏好，如图中的 SE_4_6 和 SE_5_1。低层特征通常更普遍，识别不同种类物体可以使用类似的滤波器，而高层特征通常包含更多细节，识别不同种类物体需要使用不同的滤波器。

• 第三，在模型的最后阶段，SE_5_2 呈现出饱和状态，其中大部分激活值都接近于 1，也 有一些接近于 0。对于激活值为 1 的特征图，相当于 SE 模块不存在。在网络的最后一个 SE 模块 SE_5_3，不同种类有着类似的分布，只是尺度不同。也就是说 SE_5_2 和 SE_5_3 相对 来说没有前面的一些 SE 模块重要，作者通过实验发现删除最后一个阶段的 SE 模块，总体参 数可以显著减少，性能只有一点损失(<0.1%的 Top1 错误率)。

7、SE-ResNet50图像识别

import numpy as np
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Input,Dense,Dropout,Conv2D,MaxPool2D,Flatten,GlobalAvgPool2D,BatchNormalization,Activation,Add,ZeroPadding2D,Multiply
from tensorflow.keras.optimizers import Adam
import matplotlib.pyplot as plt
from tensorflow.keras.callbacks import LearningRateScheduler
from tensorflow.keras.models import Model
from tensorflow.keras.utils import plot_model

# 类别数
num_classes = 17
# 批次大小
batch_size = 32
# 周期数
epochs = 100
# 图片大小
image_size = 224

# 训练集数据进行数据增强
train_datagen = ImageDataGenerator(
    rotation_range = 20,     # 随机旋转度数
    width_shift_range = 0.1, # 随机水平平移
    height_shift_range = 0.1,# 随机竖直平移
    rescale = 1/255,         # 数据归一化
    shear_range = 10,       # 随机错切变换
    zoom_range = 0.1,        # 随机放大
    horizontal_flip = True,  # 水平翻转
    brightness_range=(0.7, 1.3), # 亮度变化
    fill_mode = 'nearest',   # 填充方式
) 
# 测试集数据只需要归一化就可以
test_datagen = ImageDataGenerator(
    rescale = 1/255,         # 数据归一化
) 

# 训练集数据生成器，可以在训练时自动产生数据进行训练
# 从'data/train'获得训练集数据
# 获得数据后会把图片resize为image_size×image_size的大小
# generator每次会产生batch_size个数据
train_generator = train_datagen.flow_from_directory(
    'data/train',
    target_size=(image_size,image_size),
    batch_size=batch_size,
    )

# 测试集数据生成器
test_generator = test_datagen.flow_from_directory(
    'data/test',
    target_size=(image_size,image_size),
    batch_size=batch_size,
    )

# 字典的键为17个文件夹的名字，值为对应的分类编号
train_generator.class_indices

# SE模块
def ChannelSE(input_tensor, reduction=16):
    # 获得信号通道数
    channels = input_tensor.shape[-1]
    # SE模块
    x = GlobalAvgPool2D()(input_tensor)
    # 把2维数据再变成4维(?,1,1,?)
    x = x[:, None, None, :]
    # 卷积替代全连接层
    x = Conv2D(filters=channels//reduction,kernel_size=1,strides=1)(x)
    x = Activation('relu')(x)
    x = Conv2D(filters=channels,kernel_size=1,strides=1)(x)
    x = Activation('sigmoid')(x)
    x = Multiply()([input_tensor, x])
    return x


# 定义残差单元
def block(x, filters, strides=1, conv_shortcut=True, reduction=16): 
    # projection shortcut
    if conv_shortcut == True:
        shortcut = Conv2D(filters*4,kernel_size=1,strides=strides,padding='valid')(x)
        # epsilon为BN公式中防止分母为零的值
        shortcut = BatchNormalization(epsilon=1.001e-5)(shortcut)
    else:
        # identity_shortcut
        shortcut = x
    # 3个卷积层
    x = Conv2D(filters=filters,kernel_size=1,strides=strides,padding='valid')(x)
    x = BatchNormalization(epsilon=1.001e-5)(x)
    x = Activation('relu')(x)
 
    x = Conv2D(filters=filters,kernel_size=3,strides=1,padding='same')(x)
    x = BatchNormalization(epsilon=1.001e-5)(x)
    x = Activation('relu')(x)
 
    x = Conv2D(filters=filters*4,kernel_size=1,strides=1,padding='valid')(x)
    x = BatchNormalization(epsilon=1.001e-5)(x)
    
    # SE模块
    x = ChannelSE(x, reduction=reduction)
 
    x = Add()([x, shortcut])
    x = Activation('relu')(x)
    return x

# 堆叠残差单元
def stack(x, filters, blocks, strides):
    x = block(x, filters, strides=strides)
    for i in range(blocks-1):
        x = block(x, filters, conv_shortcut=False)
    return x
    
# 定义SE-ResNet50
inputs = Input(shape=(image_size,image_size,3))
# 填充3圈0，填充后图像从224×224变成230×230
x = ZeroPadding2D((3, 3))(inputs)
x= Conv2D(filters=64,kernel_size=7,strides=2,padding='valid')(x)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = Activation('relu')(x)
# 填充1圈0
x = ZeroPadding2D((1, 1))(x)
x = MaxPool2D(pool_size=3,strides=2,padding='valid')(x)
# 堆叠残差结构
# blocks表示堆叠数量
x = stack(x, filters=64, blocks=3, strides=1)
x = stack(x, filters=128, blocks=4, strides=2)
x = stack(x, filters=256, blocks=6, strides=2)
x = stack(x, filters=512, blocks=3, strides=2)
# 根据特征图大小进行平均池化，池化后得到2维数据
x = GlobalAvgPool2D()(x)
x = Dense(num_classes, activation='softmax')(x)
# 定义模型
model = Model(inputs=inputs,outputs=x)

# 学习率调节函数，逐渐减小学习率
def adjust_learning_rate(epoch):
    # 前40周期
    if epoch<=40:
        lr = 1e-4
    # 前40到80周期
    elif epoch>40 and epoch<=80:
        lr = 1e-5
    # 80到100周期
    else:
        lr = 1e-6
    return lr

# 定义优化器
adam = Adam(lr=1e-4)

# 定义学习率衰减策略
callbacks = []
callbacks.append(LearningRateScheduler(adjust_learning_rate))

# 定义优化器，loss function，训练过程中计算准确率
model.compile(optimizer=adam,loss='categorical_crossentropy',metrics=['accuracy'])

# Tensorflow2.1版本之前可以使用fit_generator训练模型
# history = model.fit_generator(train_generator,steps_per_epoch=len(train_generator),epochs=epochs,validation_data=test_generator,validation_steps=len(test_generator))

# Tensorflow2.1版本(包括2.1)之后可以直接使用fit训练模型
history = model.fit(x=train_generator,epochs=epochs,validation_data=test_generator,callbacks=callbacks)

Epoch 1/100
34/34 [==============================] - 19s 549ms/step - loss: 2.5603 - accuracy: 0.1930 - val_loss: 2.8569 - val_accuracy: 0.0588
Epoch 2/100
34/34 [==============================] - 17s 501ms/step - loss: 1.9323 - accuracy: 0.3603 - val_loss: 3.0023 - val_accuracy: 0.0588
Epoch 3/100
34/34 [==============================] - 18s 519ms/step - loss: 1.5821 - accuracy: 0.4614 - val_loss: 3.3108 - val_accuracy: 0.0588
Epoch 4/100
34/34 [==============================] - 17s 508ms/step - loss: 1.4403 - accuracy: 0.5285 - val_loss: 3.5824 - val_accuracy: 0.0588
Epoch 5/100
34/34 [==============================] - 17s 510ms/step - loss: 1.2837 - accuracy: 0.5781 - val_loss: 4.0358 - val_accuracy: 0.0625
Epoch 6/100
34/34 [==============================] - 17s 507ms/step - loss: 1.2065 - accuracy: 0.6075 - val_loss: 4.3552 - val_accuracy: 0.0588
Epoch 7/100
34/34 [==============================] - 18s 515ms/step - loss: 1.1144 - accuracy: 0.6259 - val_loss: 4.0170 - val_accuracy: 0.0846
Epoch 8/100
34/34 [==============================] - 17s 507ms/step - loss: 1.0141 - accuracy: 0.6572 - val_loss: 4.4925 - val_accuracy: 0.0625
Epoch 9/100
34/34 [==============================] - 17s 510ms/step - loss: 0.9487 - accuracy: 0.6857 - val_loss: 4.7573 - val_accuracy: 0.0993
Epoch 10/100
34/34 [==============================] - 17s 503ms/step - loss: 0.8503 - accuracy: 0.7132 - val_loss: 4.5246 - val_accuracy: 0.0919
Epoch 11/100
34/34 [==============================] - 17s 499ms/step - loss: 0.8037 - accuracy: 0.7298 - val_loss: 4.4767 - val_accuracy: 0.1213
Epoch 12/100
34/34 [==============================] - 17s 507ms/step - loss: 0.7193 - accuracy: 0.7518 - val_loss: 4.1744 - val_accuracy: 0.2243
Epoch 13/100
34/34 [==============================] - 17s 506ms/step - loss: 0.6467 - accuracy: 0.7904 - val_loss: 3.3856 - val_accuracy: 0.2537
Epoch 14/100
34/34 [==============================] - 17s 511ms/step - loss: 0.6490 - accuracy: 0.7748 - val_loss: 3.3443 - val_accuracy: 0.2941
Epoch 15/100
34/34 [==============================] - 17s 501ms/step - loss: 0.6287 - accuracy: 0.7914 - val_loss: 2.8897 - val_accuracy: 0.3603
Epoch 16/100
34/34 [==============================] - 17s 510ms/step - loss: 0.5383 - accuracy: 0.8088 - val_loss: 2.5651 - val_accuracy: 0.4338
Epoch 17/100
34/34 [==============================] - 18s 515ms/step - loss: 0.5573 - accuracy: 0.8033 - val_loss: 2.0221 - val_accuracy: 0.5000
Epoch 18/100
34/34 [==============================] - 17s 506ms/step - loss: 0.5236 - accuracy: 0.8217 - val_loss: 2.5051 - val_accuracy: 0.4485
Epoch 19/100
34/34 [==============================] - 17s 511ms/step - loss: 0.5003 - accuracy: 0.8355 - val_loss: 2.3877 - val_accuracy: 0.5000
Epoch 20/100
34/34 [==============================] - 17s 513ms/step - loss: 0.4758 - accuracy: 0.8290 - val_loss: 1.5135 - val_accuracy: 0.6728
Epoch 21/100
34/34 [==============================] - 17s 511ms/step - loss: 0.4111 - accuracy: 0.8621 - val_loss: 1.4329 - val_accuracy: 0.6176
Epoch 22/100
34/34 [==============================] - 17s 509ms/step - loss: 0.3885 - accuracy: 0.8768 - val_loss: 1.5033 - val_accuracy: 0.6434
Epoch 23/100
34/34 [==============================] - 17s 507ms/step - loss: 0.3462 - accuracy: 0.8851 - val_loss: 2.2281 - val_accuracy: 0.5000
Epoch 24/100
34/34 [==============================] - 17s 510ms/step - loss: 0.3518 - accuracy: 0.8824 - val_loss: 2.0414 - val_accuracy: 0.6176
Epoch 25/100
34/34 [==============================] - 17s 506ms/step - loss: 0.3619 - accuracy: 0.8704 - val_loss: 1.8381 - val_accuracy: 0.6287
Epoch 26/100
34/34 [==============================] - 17s 513ms/step - loss: 0.3551 - accuracy: 0.8814 - val_loss: 1.4278 - val_accuracy: 0.6507
Epoch 27/100
34/34 [==============================] - 18s 516ms/step - loss: 0.3022 - accuracy: 0.9017 - val_loss: 1.3334 - val_accuracy: 0.6838
Epoch 28/100
34/34 [==============================] - 17s 508ms/step - loss: 0.2860 - accuracy: 0.9072 - val_loss: 1.3219 - val_accuracy: 0.6875
Epoch 29/100
34/34 [==============================] - 17s 510ms/step - loss: 0.2219 - accuracy: 0.9219 - val_loss: 0.9909 - val_accuracy: 0.7132
Epoch 30/100
34/34 [==============================] - 17s 507ms/step - loss: 0.2143 - accuracy: 0.9228 - val_loss: 1.1023 - val_accuracy: 0.7353
Epoch 31/100
34/34 [==============================] - 17s 514ms/step - loss: 0.2513 - accuracy: 0.9136 - val_loss: 1.4077 - val_accuracy: 0.6691
Epoch 32/100
34/34 [==============================] - 17s 511ms/step - loss: 0.2759 - accuracy: 0.8980 - val_loss: 2.3026 - val_accuracy: 0.5772
Epoch 33/100
34/34 [==============================] - 17s 511ms/step - loss: 0.2204 - accuracy: 0.9219 - val_loss: 1.7835 - val_accuracy: 0.6213
Epoch 34/100
34/34 [==============================] - 17s 506ms/step - loss: 0.2390 - accuracy: 0.9228 - val_loss: 1.9852 - val_accuracy: 0.6103
Epoch 35/100
34/34 [==============================] - 17s 504ms/step - loss: 0.1942 - accuracy: 0.9301 - val_loss: 1.1830 - val_accuracy: 0.6765
Epoch 36/100
34/34 [==============================] - 18s 516ms/step - loss: 0.1756 - accuracy: 0.9403 - val_loss: 1.2185 - val_accuracy: 0.6985
Epoch 37/100
34/34 [==============================] - 17s 510ms/step - loss: 0.1713 - accuracy: 0.9403 - val_loss: 1.2769 - val_accuracy: 0.7243
Epoch 38/100
34/34 [==============================] - 17s 509ms/step - loss: 0.1709 - accuracy: 0.9430 - val_loss: 1.6684 - val_accuracy: 0.6654
Epoch 39/100
34/34 [==============================] - 17s 506ms/step - loss: 0.1908 - accuracy: 0.9467 - val_loss: 1.4263 - val_accuracy: 0.7096
Epoch 40/100
34/34 [==============================] - 17s 502ms/step - loss: 0.1768 - accuracy: 0.9338 - val_loss: 1.2879 - val_accuracy: 0.7096
Epoch 41/100
34/34 [==============================] - 17s 513ms/step - loss: 0.1483 - accuracy: 0.9504 - val_loss: 1.0808 - val_accuracy: 0.7390
Epoch 42/100
34/34 [==============================] - 17s 511ms/step - loss: 0.1057 - accuracy: 0.9623 - val_loss: 0.8448 - val_accuracy: 0.7868
Epoch 43/100
34/34 [==============================] - 17s 512ms/step - loss: 0.0670 - accuracy: 0.9807 - val_loss: 0.7289 - val_accuracy: 0.8162
Epoch 44/100
34/34 [==============================] - 17s 512ms/step - loss: 0.0768 - accuracy: 0.9779 - val_loss: 0.6560 - val_accuracy: 0.8162
Epoch 45/100
34/34 [==============================] - 17s 505ms/step - loss: 0.0568 - accuracy: 0.9871 - val_loss: 0.6800 - val_accuracy: 0.8199
Epoch 46/100
34/34 [==============================] - 18s 515ms/step - loss: 0.0612 - accuracy: 0.9871 - val_loss: 0.6611 - val_accuracy: 0.8309
Epoch 47/100
34/34 [==============================] - 17s 505ms/step - loss: 0.0499 - accuracy: 0.9899 - val_loss: 0.6868 - val_accuracy: 0.8015
Epoch 48/100
34/34 [==============================] - 17s 506ms/step - loss: 0.0561 - accuracy: 0.9816 - val_loss: 0.6607 - val_accuracy: 0.8051
Epoch 49/100
34/34 [==============================] - 17s 505ms/step - loss: 0.0418 - accuracy: 0.9926 - val_loss: 0.6790 - val_accuracy: 0.8015
Epoch 50/100
34/34 [==============================] - 18s 516ms/step - loss: 0.0383 - accuracy: 0.9917 - val_loss: 0.6684 - val_accuracy: 0.8162
Epoch 51/100
34/34 [==============================] - 17s 506ms/step - loss: 0.0416 - accuracy: 0.9917 - val_loss: 0.6851 - val_accuracy: 0.8051
Epoch 52/100
34/34 [==============================] - 17s 503ms/step - loss: 0.0435 - accuracy: 0.9881 - val_loss: 0.6635 - val_accuracy: 0.8051
Epoch 53/100
34/34 [==============================] - 17s 503ms/step - loss: 0.0457 - accuracy: 0.9917 - val_loss: 0.6653 - val_accuracy: 0.8199
Epoch 54/100
34/34 [==============================] - 17s 508ms/step - loss: 0.0444 - accuracy: 0.9862 - val_loss: 0.7060 - val_accuracy: 0.8125
Epoch 55/100
34/34 [==============================] - 17s 506ms/step - loss: 0.0497 - accuracy: 0.9862 - val_loss: 0.7086 - val_accuracy: 0.8051
Epoch 56/100
34/34 [==============================] - 18s 516ms/step - loss: 0.0355 - accuracy: 0.9926 - val_loss: 0.6972 - val_accuracy: 0.8162
Epoch 57/100
34/34 [==============================] - 18s 515ms/step - loss: 0.0407 - accuracy: 0.9917 - val_loss: 0.7006 - val_accuracy: 0.8088
Epoch 58/100
34/34 [==============================] - 17s 514ms/step - loss: 0.0429 - accuracy: 0.9890 - val_loss: 0.6793 - val_accuracy: 0.8162
Epoch 59/100
34/34 [==============================] - 17s 509ms/step - loss: 0.0327 - accuracy: 0.9936 - val_loss: 0.6627 - val_accuracy: 0.8199
Epoch 60/100
34/34 [==============================] - 17s 504ms/step - loss: 0.0449 - accuracy: 0.9890 - val_loss: 0.6477 - val_accuracy: 0.8088
Epoch 61/100
34/34 [==============================] - 17s 507ms/step - loss: 0.0349 - accuracy: 0.9926 - val_loss: 0.6539 - val_accuracy: 0.8125
Epoch 62/100
34/34 [==============================] - 17s 500ms/step - loss: 0.0273 - accuracy: 0.9963 - val_loss: 0.6503 - val_accuracy: 0.8235
Epoch 63/100
34/34 [==============================] - 17s 509ms/step - loss: 0.0287 - accuracy: 0.9972 - val_loss: 0.6243 - val_accuracy: 0.8346
Epoch 64/100
34/34 [==============================] - 18s 515ms/step - loss: 0.0306 - accuracy: 0.9954 - val_loss: 0.6159 - val_accuracy: 0.8272
Epoch 65/100
34/34 [==============================] - 17s 510ms/step - loss: 0.0272 - accuracy: 0.9936 - val_loss: 0.6473 - val_accuracy: 0.8125
Epoch 66/100
34/34 [==============================] - 17s 508ms/step - loss: 0.0257 - accuracy: 0.9972 - val_loss: 0.6518 - val_accuracy: 0.8088
Epoch 67/100
34/34 [==============================] - 17s 507ms/step - loss: 0.0325 - accuracy: 0.9954 - val_loss: 0.6649 - val_accuracy: 0.8015
Epoch 68/100
34/34 [==============================] - 17s 514ms/step - loss: 0.0269 - accuracy: 0.9963 - val_loss: 0.6461 - val_accuracy: 0.8088
Epoch 69/100
34/34 [==============================] - 17s 506ms/step - loss: 0.0278 - accuracy: 0.9954 - val_loss: 0.6525 - val_accuracy: 0.8088
Epoch 70/100
34/34 [==============================] - 17s 513ms/step - loss: 0.0347 - accuracy: 0.9899 - val_loss: 0.6667 - val_accuracy: 0.8088
Epoch 71/100
34/34 [==============================] - 17s 500ms/step - loss: 0.0296 - accuracy: 0.9936 - val_loss: 0.6512 - val_accuracy: 0.8235
Epoch 72/100
34/34 [==============================] - 17s 507ms/step - loss: 0.0373 - accuracy: 0.9899 - val_loss: 0.6657 - val_accuracy: 0.8125
Epoch 73/100
34/34 [==============================] - 17s 500ms/step - loss: 0.0307 - accuracy: 0.9954 - val_loss: 0.6822 - val_accuracy: 0.8051
Epoch 74/100
34/34 [==============================] - 17s 499ms/step - loss: 0.0332 - accuracy: 0.9899 - val_loss: 0.6935 - val_accuracy: 0.8162
Epoch 75/100
34/34 [==============================] - 17s 510ms/step - loss: 0.0187 - accuracy: 0.9991 - val_loss: 0.6491 - val_accuracy: 0.8162
Epoch 76/100
34/34 [==============================] - 17s 503ms/step - loss: 0.0243 - accuracy: 0.9963 - val_loss: 0.6576 - val_accuracy: 0.8272
Epoch 77/100
34/34 [==============================] - 17s 514ms/step - loss: 0.0227 - accuracy: 0.9945 - val_loss: 0.6988 - val_accuracy: 0.8162
Epoch 78/100
34/34 [==============================] - 17s 514ms/step - loss: 0.0205 - accuracy: 0.9954 - val_loss: 0.7424 - val_accuracy: 0.8162
Epoch 79/100
34/34 [==============================] - 17s 506ms/step - loss: 0.0298 - accuracy: 0.9936 - val_loss: 0.6804 - val_accuracy: 0.8309
Epoch 80/100
34/34 [==============================] - 17s 514ms/step - loss: 0.0246 - accuracy: 0.9936 - val_loss: 0.7717 - val_accuracy: 0.8015
Epoch 81/100
34/34 [==============================] - 17s 505ms/step - loss: 0.0294 - accuracy: 0.9945 - val_loss: 0.6915 - val_accuracy: 0.7978
Epoch 82/100
34/34 [==============================] - 17s 512ms/step - loss: 0.0223 - accuracy: 0.9954 - val_loss: 0.6843 - val_accuracy: 0.8015
Epoch 83/100
34/34 [==============================] - 17s 510ms/step - loss: 0.0260 - accuracy: 0.9945 - val_loss: 0.6777 - val_accuracy: 0.8051
Epoch 84/100
34/34 [==============================] - 17s 507ms/step - loss: 0.0250 - accuracy: 0.9926 - val_loss: 0.6703 - val_accuracy: 0.8051
Epoch 85/100
34/34 [==============================] - 17s 508ms/step - loss: 0.0263 - accuracy: 0.9954 - val_loss: 0.6659 - val_accuracy: 0.8088
Epoch 86/100
34/34 [==============================] - 17s 510ms/step - loss: 0.0189 - accuracy: 0.9963 - val_loss: 0.6651 - val_accuracy: 0.8088
Epoch 87/100
34/34 [==============================] - 17s 512ms/step - loss: 0.0250 - accuracy: 0.9954 - val_loss: 0.6708 - val_accuracy: 0.8199
Epoch 88/100
34/34 [==============================] - 17s 506ms/step - loss: 0.0247 - accuracy: 0.9945 - val_loss: 0.6672 - val_accuracy: 0.8199
Epoch 89/100
34/34 [==============================] - 17s 508ms/step - loss: 0.0237 - accuracy: 0.9936 - val_loss: 0.6719 - val_accuracy: 0.8235
Epoch 90/100
34/34 [==============================] - 17s 507ms/step - loss: 0.0268 - accuracy: 0.9936 - val_loss: 0.6678 - val_accuracy: 0.8235
Epoch 91/100
34/34 [==============================] - 17s 510ms/step - loss: 0.0228 - accuracy: 0.9936 - val_loss: 0.6671 - val_accuracy: 0.8199
Epoch 92/100
34/34 [==============================] - 18s 515ms/step - loss: 0.0154 - accuracy: 1.0000 - val_loss: 0.6638 - val_accuracy: 0.8235
Epoch 93/100
34/34 [==============================] - 17s 512ms/step - loss: 0.0206 - accuracy: 0.9991 - val_loss: 0.6617 - val_accuracy: 0.8199
Epoch 94/100
34/34 [==============================] - 17s 501ms/step - loss: 0.0249 - accuracy: 0.9954 - val_loss: 0.6695 - val_accuracy: 0.8235
Epoch 95/100
34/34 [==============================] - 17s 509ms/step - loss: 0.0282 - accuracy: 0.9945 - val_loss: 0.6846 - val_accuracy: 0.8235
Epoch 96/100
34/34 [==============================] - 17s 513ms/step - loss: 0.0196 - accuracy: 0.9963 - val_loss: 0.6784 - val_accuracy: 0.8272
Epoch 97/100
34/34 [==============================] - 18s 523ms/step - loss: 0.0255 - accuracy: 0.9917 - val_loss: 0.6732 - val_accuracy: 0.8346
Epoch 98/100
34/34 [==============================] - 17s 513ms/step - loss: 0.0278 - accuracy: 0.9899 - val_loss: 0.6697 - val_accuracy: 0.8309
Epoch 99/100
34/34 [==============================] - 17s 512ms/step - loss: 0.0163 - accuracy: 0.9963 - val_loss: 0.6711 - val_accuracy: 0.8309
Epoch 100/100
34/34 [==============================] - 17s 503ms/step - loss: 0.0238 - accuracy: 0.9945 - val_loss: 0.6632 - val_accuracy: 0.8272

可视化acc

# 画出训练集准确率曲线图
plt.plot(np.arange(epochs),history.history['accuracy'],c='b',label='train_accuracy')
# 画出验证集准确率曲线图
plt.plot(np.arange(epochs),history.history['val_accuracy'],c='y',label='val_accuracy')
# 图例
plt.legend()
# x坐标描述
plt.xlabel('epochs')
# y坐标描述
plt.ylabel('accuracy')
# 显示图像
plt.show()
# 模型保存
model.save('SE-ResNet50.h5')

模型结构如下所示：

Model: "functional_1"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            [(None, 224, 224, 3) 0                                            
__________________________________________________________________________________________________
zero_padding2d (ZeroPadding2D)  (None, 230, 230, 3)  0           input_1[0][0]                    
__________________________________________________________________________________________________
conv2d (Conv2D)                 (None, 112, 112, 64) 9472        zero_padding2d[0][0]             
__________________________________________________________________________________________________
batch_normalization (BatchNorma (None, 112, 112, 64) 256         conv2d[0][0]                     
__________________________________________________________________________________________________
activation (Activation)         (None, 112, 112, 64) 0           batch_normalization[0][0]        
__________________________________________________________________________________________________
zero_padding2d_1 (ZeroPadding2D (None, 114, 114, 64) 0           activation[0][0]                 
__________________________________________________________________________________________________
max_pooling2d (MaxPooling2D)    (None, 56, 56, 64)   0           zero_padding2d_1[0][0]           
__________________________________________________________________________________________________
conv2d_2 (Conv2D)               (None, 56, 56, 64)   4160        max_pooling2d[0][0]              
__________________________________________________________________________________________________
batch_normalization_2 (BatchNor (None, 56, 56, 64)   256         conv2d_2[0][0]                   
__________________________________________________________________________________________________
activation_1 (Activation)       (None, 56, 56, 64)   0           batch_normalization_2[0][0]      
__________________________________________________________________________________________________
conv2d_3 (Conv2D)               (None, 56, 56, 64)   36928       activation_1[0][0]               
__________________________________________________________________________________________________
batch_normalization_3 (BatchNor (None, 56, 56, 64)   256         conv2d_3[0][0]                   
__________________________________________________________________________________________________
activation_2 (Activation)       (None, 56, 56, 64)   0           batch_normalization_3[0][0]      
__________________________________________________________________________________________________
conv2d_4 (Conv2D)               (None, 56, 56, 256)  16640       activation_2[0][0]               
__________________________________________________________________________________________________
batch_normalization_4 (BatchNor (None, 56, 56, 256)  1024        conv2d_4[0][0]                   
__________________________________________________________________________________________________
global_average_pooling2d (Globa (None, 256)          0           batch_normalization_4[0][0]      
__________________________________________________________________________________________________
tf_op_layer_strided_slice (Tens (None, 1, 1, 256)    0           global_average_pooling2d[0][0]   
__________________________________________________________________________________________________
conv2d_5 (Conv2D)               (None, 1, 1, 16)     4112        tf_op_layer_strided_slice[0][0]  
__________________________________________________________________________________________________
activation_3 (Activation)       (None, 1, 1, 16)     0           conv2d_5[0][0]                   
__________________________________________________________________________________________________
conv2d_6 (Conv2D)               (None, 1, 1, 256)    4352        activation_3[0][0]               
__________________________________________________________________________________________________
activation_4 (Activation)       (None, 1, 1, 256)    0           conv2d_6[0][0]                   
__________________________________________________________________________________________________
conv2d_1 (Conv2D)               (None, 56, 56, 256)  16640       max_pooling2d[0][0]              
__________________________________________________________________________________________________
multiply (Multiply)             (None, 56, 56, 256)  0           batch_normalization_4[0][0]      
                                                                 activation_4[0][0]               
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 56, 56, 256)  1024        conv2d_1[0][0]                   
__________________________________________________________________________________________________
add (Add)                       (None, 56, 56, 256)  0           multiply[0][0]                   
                                                                 batch_normalization_1[0][0]      
__________________________________________________________________________________________________
activation_5 (Activation)       (None, 56, 56, 256)  0           add[0][0]                        
__________________________________________________________________________________________________
conv2d_7 (Conv2D)               (None, 56, 56, 64)   16448       activation_5[0][0]               
__________________________________________________________________________________________________
batch_normalization_5 (BatchNor (None, 56, 56, 64)   256         conv2d_7[0][0]                   
__________________________________________________________________________________________________
activation_6 (Activation)       (None, 56, 56, 64)   0           batch_normalization_5[0][0]      
__________________________________________________________________________________________________
conv2d_8 (Conv2D)               (None, 56, 56, 64)   36928       activation_6[0][0]               
__________________________________________________________________________________________________
batch_normalization_6 (BatchNor (None, 56, 56, 64)   256         conv2d_8[0][0]                   
__________________________________________________________________________________________________
activation_7 (Activation)       (None, 56, 56, 64)   0           batch_normalization_6[0][0]      
__________________________________________________________________________________________________
conv2d_9 (Conv2D)               (None, 56, 56, 256)  16640       activation_7[0][0]               
__________________________________________________________________________________________________
batch_normalization_7 (BatchNor (None, 56, 56, 256)  1024        conv2d_9[0][0]                   
__________________________________________________________________________________________________
global_average_pooling2d_1 (Glo (None, 256)          0           batch_normalization_7[0][0]      
__________________________________________________________________________________________________
tf_op_layer_strided_slice_1 (Te (None, 1, 1, 256)    0           global_average_pooling2d_1[0][0] 
__________________________________________________________________________________________________
conv2d_10 (Conv2D)              (None, 1, 1, 16)     4112        tf_op_layer_strided_slice_1[0][0]
__________________________________________________________________________________________________
activation_8 (Activation)       (None, 1, 1, 16)     0           conv2d_10[0][0]                  
__________________________________________________________________________________________________
conv2d_11 (Conv2D)              (None, 1, 1, 256)    4352        activation_8[0][0]               
__________________________________________________________________________________________________
activation_9 (Activation)       (None, 1, 1, 256)    0           conv2d_11[0][0]                  
__________________________________________________________________________________________________
multiply_1 (Multiply)           (None, 56, 56, 256)  0           batch_normalization_7[0][0]      
                                                                 activation_9[0][0]               
__________________________________________________________________________________________________
add_1 (Add)                     (None, 56, 56, 256)  0           multiply_1[0][0]                 
                                                                 activation_5[0][0]               
__________________________________________________________________________________________________
activation_10 (Activation)      (None, 56, 56, 256)  0           add_1[0][0]                      
_______________________