>- **🍨 本文为[🔗365天深度学习训练营]中的学习记录博客**
>- **🍖 原作者：[K同学啊]**
📌 本周任务：
●1.请根据本文 Pytorch 代码，编写出相应的 TensorFlow 代码（建议使用上周的数据测试一下模型是否构建正确）
●2.了解并研究 DenseNet与ResNetV 的区别
●3.改进思路是否可以迁移到其他地方呢（自由探索，虽然不强求但是请认真对待这个哦）
🚀我的环境：
- 语言环境：Python3.11.7
- 编译器：jupyter notebook
- 深度学习框架：TensorFlow2.13.0
本文完全根据 第J3周：DenseNet算法实战与解析（pytorch版）中的内容转换为TensorFlow，所以前述性的内容不在一一重复，仅就TensorFlow的内容进行叙述。
一、前期工作
1、设置CPU（也可以是GPU）
import tensorflow as tf
gpus=tf.config.list_physical_devices("GPU")if gpus:tf.config.experimental.set_memory_growth(gpus[0],True)tf.config.set_visible_devices([gpus[0]],"GPU")2、导入数据
import pathlibdata_dir=r'D:\THE MNIST DATABASE\J-series\J1\bird_photos'
data_dir=pathlib.Path(data_dir)3、查看数据
image_count=len(list(data_dir.glob('*/*')))
print("图片总数为：",image_count)运行结果：
图片总数为： 565二、数据预处理
1、加载数据
加载训练集：
train_ds=tf.keras.preprocessing.image_dataset_from_directory(data_dir,validation_split=0.2,subset="training",seed=123,image_size=(224,224),batch_size=8
)运行结果：
Found 565 files belonging to 4 classes.
Using 452 files for training.
val_ds=tf.keras.preprocessing.image_dataset_from_directory(data_dir,validation_split=0.2,subset="validation",seed=123,image_size=(224,224),batch_size=8
)
val_ds=tf.keras.preprocessing.image_datas加载验证集：
val_ds=tf.keras.preprocessing.image_dataset_from_directory(data_dir,validation_split=0.2,subset="validation",seed=123,image_size=(224,224),batch_size=8
)运行结果：
ound 565 files belonging to 4 classes.
Using 113 files for validation.查看分类名称
classNames=train_ds.class_names
print(classNames)运行结果：
['Bananaquit', 'Black Skimmer', 'Black Throated Bushtiti', 'Cockatoo']2、可视化数据
import matplotlib.pyplot as plt
plt.rcParams['font.sans-serif']=['SimHei'] #正常显示中文标签
plt.rcParams['axes.unicode_minus']=False #正常显示负号plt.figure(figsize=(10,5))
plt.suptitle("OreoCC的案例")for images,labels in train_ds.take(1):for i in range(8):ax=plt.subplot(2,4,i+1)plt.imshow(images[i].numpy().astype("uint8"))plt.title(classNames[labels[i]])plt.axis("off")运行结果：
单独查看其中一张图片
plt.imshow(images[1].numpy().astype("uint8"))运行结果：
3、再次检查数据
for image_batch,labels_batch in train_ds:print(image_batch.shape)print(labels_batch.shape)break运行结果：
(8, 224, 224, 3)
(8,)image_batch是形状的张量(8,224,224,3)。这是一批形状224*224*4的8张图片（最后一维指的是彩色通道RGB）
labels_batch是形状(8,)的张量，这些标签对应8张图片。
4、配置数据集
shuffle() ： 打乱数据，关于此函数的详细介绍可以参考：https://zhuanlan.zhihu.com/p/42417456
prefetch() ：预取数据，加速运行，其详细介绍可以参考前面文章，里面都有讲解。
cache() ：将数据集缓存到内存当中，加速运行
AUTOTUNE=tf.data.AUTOTUNEtrain_ds=train_ds.cache().shuffle(1000).prefetch(buffer_size=AUTOTUNE)
val_ds=val_ds.cache().prefetch(buffer_size=AUTOTUNE)三、构建DenseNet网络模型
1、搭建DenseLayer
from tensorflow import keras
from keras.layers import Input,Activation,BatchNormalization,Flatten
from keras.layers import Dense,Conv2D,MaxPooling2D,ZeroPadding2D,AveragePooling2D
from keras.models import Model#DenseLayer
def DenseLayer(x,growth_rate):f=BatchNormalization()(x)f=Activation('relu')(f)f=Conv2D(4*growth_rate,kernel_size=1,strides=1,padding='same',use_bias=False)(f)f=BatchNormalization()(f)f=Activation('relu')(f)f=Conv2D(growth_rate,kernel_size=3,strides=1,padding='same',use_bias=False)(f)return layers.Concatenate(axis=3)([x,f])2、搭建DenseBlock模块
#DenseBlock
def DenseBlock(x,block,growth_rate=32):for i in range(block):x=DenseLayer(x,growth_rate)return x3、搭建TransitionBlock模块
#Transition
k=keras.backend
def Transition(x,theta):f=BatchNormalization()(x)f=Activation('relu')(f)f=Conv2D(int(k.int_shape(x)[3]*theta),kernel_size=1,strides=1,use_bias=False)(f)f=AveragePooling2D(pool_size=(2,2),strides=2,padding='valid')(f)return f4、搭建DenseNet网络模型
#DenseNet
def DenseNet(input_shape,block,num_classes=4):#56*56*64img_input=Input(shape=input_shape)x=Conv2D(64,kernel_size=(7,7),strides=2,padding='same',use_bias=False)(img_input)x=BatchNormalization()(x)x=MaxPooling2D(pool_size=3,strides=2,padding='same')(x)x=DenseBlock(x,block[0])x=Transition(x,0.5) #28*28x=DenseBlock(x,block[1])x=Transition(x,0.5) #14*14x=DenseBlock(x,block[2])x=Transition(x,0.5) #7*7x=DenseBlock(x,block[3])x=BatchNormalization()(x)x=Activation('relu')(x)x=GlobalAveragePooling2D()(x)outputs=Dense(num_classes,activation='softmax')(x)model=Model(inputs=[img_input],outputs=[outputs])return model5、建立DenseNet-121模型
model_121=DenseNet([224,224,3],[6,12,24,16]) #DenseNet-121
model_169=DenseNet([224,224,3],[6,12,32,32]) #DenseNet-169
model_201=DenseNet([224,224,3],[6,12,48,32]) #DenseNet-201
model_269=DenseNet([224,224,3],[6,12,64,48]) #DenseNet-269
model=model_121
model.summary()查看模型结构：
Model: "model_4"
__________________________________________________________________________________________________Layer (type) Output Shape Param # Connected to
==================================================================================================input_12 (InputLayer) [(None, 224, 224, 3)] 0 [] conv2d_1112 (Conv2D) (None, 112, 112, 64) 9408 ['input_12[0][0]'] batch_normalization_1118 ( (None, 112, 112, 64) 256 ['conv2d_1112[0][0]'] BatchNormalization) max_pooling2d_7 (MaxPoolin (None, 56, 56, 64) 0 ['batch_normalization_1118[0][g2D) 0]'] batch_normalization_1119 ( (None, 56, 56, 64) 256 ['max_pooling2d_7[0][0]'] BatchNormalization) activation_1112 (Activatio (None, 56, 56, 64) 0 ['batch_normalization_1119[0][n) 0]'] conv2d_1113 (Conv2D) (None, 56, 56, 128) 8192 ['activation_1112[0][0]'] batch_normalization_1120 ( (None, 56, 56, 128) 512 ['conv2d_1113[0][0]'] BatchNormalization) activation_1113 (Activatio (None, 56, 56, 128) 0 ['batch_normalization_1120[0][n) 0]'] conv2d_1114 (Conv2D) (None, 56, 56, 32) 36864 ['activation_1113[0][0]'] concatenate_484 (Concatena (None, 56, 56, 96) 0 ['max_pooling2d_7[0][0]', te) 'conv2d_1114[0][0]'] batch_normalization_1121 ( (None, 56, 56, 96) 384 ['concatenate_484[0][0]'] BatchNormalization) activation_1114 (Activatio (None, 56, 56, 96) 0 ['batch_normalization_1121[0][n) 0]'] conv2d_1115 (Conv2D) (None, 56, 56, 128) 12288 ['activation_1114[0][0]'] batch_normalization_1122 ( (None, 56, 56, 128) 512 ['conv2d_1115[0][0]'] BatchNormalization) activation_1115 (Activatio (None, 56, 56, 128) 0 ['batch_normalization_1122[0][n) 0]'] conv2d_1116 (Conv2D) (None, 56, 56, 32) 36864 ['activation_1115[0][0]'] concatenate_485 (Concatena (None, 56, 56, 128) 0 ['concatenate_484[0][0]', te) 'conv2d_1116[0][0]'] batch_normalization_1123 ( (None, 56, 56, 128) 512 ['concatenate_485[0][0]'] BatchNormalization) activation_1116 (Activatio (None, 56, 56, 128) 0 ['batch_normalization_1123[0][n) 0]'] conv2d_1117 (Conv2D) (None, 56, 56, 128) 16384 ['activation_1116[0][0]'] batch_normalization_1124 ( (None, 56, 56, 128) 512 ['conv2d_1117[0][0]'] BatchNormalization) activation_1117 (Activatio (None, 56, 56, 128) 0 ['batch_normalization_1124[0][n) 0]'] conv2d_1118 (Conv2D) (None, 56, 56, 32) 36864 ['activation_1117[0][0]'] concatenate_486 (Concatena (None, 56, 56, 160) 0 ['concatenate_485[0][0]', te) 'conv2d_1118[0][0]'] batch_normalization_1125 ( (None, 56, 56, 160) 640 ['concatenate_486[0][0]'] BatchNormalization) activation_1118 (Activatio (None, 56, 56, 160) 0 ['batch_normalization_1125[0][n) 0]'] conv2d_1119 (Conv2D) (None, 56, 56, 128) 20480 ['activation_1118[0][0]'] batch_normalization_1126 ( (None, 56, 56, 128) 512 ['conv2d_1119[0][0]'] BatchNormalization) activation_1119 (Activatio (None, 56, 56, 128) 0 ['batch_normalization_1126[0][n) 0]'] conv2d_1120 (Conv2D) (None, 56, 56, 32) 36864 ['activation_1119[0][0]'] concatenate_487 (Concatena (None, 56, 56, 192) 0 ['concatenate_486[0][0]', te) 'conv2d_1120[0][0]'] batch_normalization_1127 ( (None, 56, 56, 192) 768 ['concatenate_487[0][0]'] BatchNormalization) activation_1120 (Activatio (None, 56, 56, 192) 0 ['batch_normalization_1127[0][n) 0]'] conv2d_1121 (Conv2D) (None, 56, 56, 128) 24576 ['activation_1120[0][0]'] batch_normalization_1128 ( (None, 56, 56, 128) 512 ['conv2d_1121[0][0]'] BatchNormalization) activation_1121 (Activatio (None, 56, 56, 128) 0 ['batch_normalization_1128[0][n) 0]'] conv2d_1122 (Conv2D) (None, 56, 56, 32) 36864 ['activation_1121[0][0]'] concatenate_488 (Concatena (None, 56, 56, 224) 0 ['concatenate_487[0][0]', te) 'conv2d_1122[0][0]'] batch_normalization_1129 ( (None, 56, 56, 224) 896 ['concatenate_488[0][0]'] BatchNormalization) activation_1122 (Activatio (None, 56, 56, 224) 0 ['batch_normalization_1129[0][n) 0]'] conv2d_1123 (Conv2D) (None, 56, 56, 128) 28672 ['activation_1122[0][0]'] batch_normalization_1130 ( (None, 56, 56, 128) 512 ['conv2d_1123[0][0]'] BatchNormalization) activation_1123 (Activatio (None, 56, 56, 128) 0 ['batch_normalization_1130[0][n) 0]'] conv2d_1124 (Conv2D) (None, 56, 56, 32) 36864 ['activation_1123[0][0]'] concatenate_489 (Concatena (None, 56, 56, 256) 0 ['concatenate_488[0][0]', te) 'conv2d_1124[0][0]'] batch_normalization_1131 ( (None, 56, 56, 256) 1024 ['concatenate_489[0][0]'] BatchNormalization) activation_1124 (Activatio (None, 56, 56, 256) 0 ['batch_normalization_1131[0][n) 0]'] conv2d_1125 (Conv2D) (None, 56, 56, 128) 32768 ['activation_1124[0][0]'] average_pooling2d_21 (Aver (None, 28, 28, 128) 0 ['conv2d_1125[0][0]'] agePooling2D) batch_normalization_1132 ( (None, 28, 28, 128) 512 ['average_pooling2d_21[0][0]']BatchNormalization) activation_1125 (Activatio (None, 28, 28, 128) 0 ['batch_normalization_1132[0][n) 0]'] conv2d_1126 (Conv2D) (None, 28, 28, 128) 16384 ['activation_1125[0][0]'] batch_normalization_1133 ( (None, 28, 28, 128) 512 ['conv2d_1126[0][0]'] BatchNormalization) activation_1126 (Activatio (None, 28, 28, 128) 0 ['batch_normalization_1133[0][n) 0]'] conv2d_1127 (Conv2D) (None, 28, 28, 32) 36864 ['activation_1126[0][0]'] concatenate_490 (Concatena (None, 28, 28, 160) 0 ['average_pooling2d_21[0][0]',te) 'conv2d_1127[0][0]'] batch_normalization_1134 ( (None, 28, 28, 160) 640 ['concatenate_490[0][0]'] BatchNormalization) activation_1127 (Activatio (None, 28, 28, 160) 0 ['batch_normalization_1134[0][n) 0]'] conv2d_1128 (Conv2D) (None, 28, 28, 128) 20480 ['activation_1127[0][0]'] batch_normalization_1135 ( (None, 28, 28, 128) 512 ['conv2d_1128[0][0]'] BatchNormalization) activation_1128 (Activatio (None, 28, 28, 128) 0 ['batch_normalization_1135[0][n) 0]'] conv2d_1129 (Conv2D) (None, 28, 28, 32) 36864 ['activation_1128[0][0]'] concatenate_491 (Concatena (None, 28, 28, 192) 0 ['concatenate_490[0][0]', te) 'conv2d_1129[0][0]'] batch_normalization_1136 ( (None, 28, 28, 192) 768 ['concatenate_491[0][0]'] BatchNormalization) activation_1129 (Activatio (None, 28, 28, 192) 0 ['batch_normalization_1136[0][n) 0]'] conv2d_1130 (Conv2D) (None, 28, 28, 128) 24576 ['activation_1129[0][0]'] batch_normalization_1137 ( (None, 28, 28, 128) 512 ['conv2d_1130[0][0]'] BatchNormalization) activation_1130 (Activatio (None, 28, 28, 128) 0 ['batch_normalization_1137[0][n) 0]'] conv2d_1131 (Conv2D) (None, 28, 28, 32) 36864 ['activation_1130[0][0]'] concatenate_492 (Concatena (None, 28, 28, 224) 0 ['concatenate_491[0][0]', te) 'conv2d_1131[0][0]'] batch_normalization_1138 ( (None, 28, 28, 224) 896 ['concatenate_492[0][0]'] BatchNormalization) activation_1131 (Activatio (None, 28, 28, 224) 0 ['batch_normalization_1138[0][n) 0]'] conv2d_1132 (Conv2D) (None, 28, 28, 128) 28672 ['activation_1131[0][0]'] batch_normalization_1139 ( (None, 28, 28, 128) 512 ['conv2d_1132[0][0]'] BatchNormalization) activation_1132 (Activatio (None, 28, 28, 128) 0 ['batch_normalization_1139[0][n) 0]'] conv2d_1133 (Conv2D) (None, 28, 28, 32) 36864 ['activation_1132[0][0]'] concatenate_493 (Concatena (None, 28, 28, 256) 0 ['concatenate_492[0][0]', te) 'conv2d_1133[0][0]'] batch_normalization_1140 ( (None, 28, 28, 256) 1024 ['concatenate_493[0][0]'] BatchNormalization) activation_1133 (Activatio (None, 28, 28, 256) 0 ['batch_normalization_1140[0][n) 0]'] conv2d_1134 (Conv2D) (None, 28, 28, 128) 32768 ['activation_1133[0][0]'] batch_normalization_1141 ( (None, 28, 28, 128) 512 ['conv2d_1134[0][0]'] BatchNormalization) activation_1134 (Activatio (None, 28, 28, 128) 0 ['batch_normalization_1141[0][n) 0]'] conv2d_1135 (Conv2D) (None, 28, 28, 32) 36864 ['activation_1134[0][0]'] concatenate_494 (Concatena (None, 28, 28, 288) 0 ['concatenate_493[0][0]', te) 'conv2d_1135[0][0]'] batch_normalization_1142 ( (None, 28, 28, 288) 1152 ['concatenate_494[0][0]'] BatchNormalization) activation_1135 (Activatio (None, 28, 28, 288) 0 ['batch_normalization_1142[0][n) 0]'] conv2d_1136 (Conv2D) (None, 28, 28, 128) 36864 ['activation_1135[0][0]'] batch_normalization_1143 ( (None, 28, 28, 128) 512 ['conv2d_1136[0][0]'] BatchNormalization) activation_1136 (Activatio (None, 28, 28, 128) 0 ['batch_normalization_1143[0][n) 0]'] conv2d_1137 (Conv2D) (None, 28, 28, 32) 36864 ['activation_1136[0][0]'] concatenate_495 (Concatena (None, 28, 28, 320) 0 ['concatenate_494[0][0]', te) 'conv2d_1137[0][0]'] batch_normalization_1144 ( (None, 28, 28, 320) 1280 ['concatenate_495[0][0]'] BatchNormalization) activation_1137 (Activatio (None, 28, 28, 320) 0 ['batch_normalization_1144[0][n) 0]'] conv2d_1138 (Conv2D) (None, 28, 28, 128) 40960 ['activation_1137[0][0]'] batch_normalization_1145 ( (None, 28, 28, 128) 512 ['conv2d_1138[0][0]'] BatchNormalization) activation_1138 (Activatio (None, 28, 28, 128) 0 ['batch_normalization_1145[0][n) 0]'] conv2d_1139 (Conv2D) (None, 28, 28, 32) 36864 ['activation_1138[0][0]'] concatenate_496 (Concatena (None, 28, 28, 352) 0 ['concatenate_495[0][0]', te) 'conv2d_1139[0][0]'] batch_normalization_1146 ( (None, 28, 28, 352) 1408 ['concatenate_496[0][0]'] BatchNormalization) activation_1139 (Activatio (None, 28, 28, 352) 0 ['batch_normalization_1146[0][n) 0]'] conv2d_1140 (Conv2D) (None, 28, 28, 128) 45056 ['activation_1139[0][0]'] batch_normalization_1147 ( (None, 28, 28, 128) 512 ['conv2d_1140[0][0]'] BatchNormalization) activation_1140 (Activatio (None, 28, 28, 128) 0 ['batch_normalization_1147[0][n) 0]'] conv2d_1141 (Conv2D) (None, 28, 28, 32) 36864 ['activation_1140[0][0]'] concatenate_497 (Concatena (None, 28, 28, 384) 0 ['concatenate_496[0][0]', te) 'conv2d_1141[0][0]'] batch_normalization_1148 ( (None, 28, 28, 384) 1536 ['concatenate_497[0][0]'] BatchNormalization) activation_1141 (Activatio (None, 28, 28, 384) 0 ['batch_normalization_1148[0][n) 0]'] conv2d_1142 (Conv2D) (None, 28, 28, 128) 49152 ['activation_1141[0][0]'] batch_normalization_1149 ( (None, 28, 28, 128) 512 ['conv2d_1142[0][0]'] BatchNormalization) activation_1142 (Activatio (None, 28, 28, 128) 0 ['batch_normalization_1149[0][n) 0]'] conv2d_1143 (Conv2D) (None, 28, 28, 32) 36864 ['activation_1142[0][0]'] concatenate_498 (Concatena (None, 28, 28, 416) 0 ['concatenate_497[0][0]', te) 'conv2d_1143[0][0]'] batch_normalization_1150 ( (None, 28, 28, 416) 1664 ['concatenate_498[0][0]'] BatchNormalization) activation_1143 (Activatio (None, 28, 28, 416) 0 ['batch_normalization_1150[0][n) 0]'] conv2d_1144 (Conv2D) (None, 28, 28, 128) 53248 ['activation_1143[0][0]'] batch_normalization_1151 ( (None, 28, 28, 128) 512 ['conv2d_1144[0][0]'] BatchNormalization) activation_1144 (Activatio (None, 28, 28, 128) 0 ['batch_normalization_1151[0][n) 0]'] conv2d_1145 (Conv2D) (None, 28, 28, 32) 36864 ['activation_1144[0][0]'] concatenate_499 (Concatena (None, 28, 28, 448) 0 ['concatenate_498[0][0]', te) 'conv2d_1145[0][0]'] batch_normalization_1152 ( (None, 28, 28, 448) 1792 ['concatenate_499[0][0]'] BatchNormalization) activation_1145 (Activatio (None, 28, 28, 448) 0 ['batch_normalization_1152[0][n) 0]'] conv2d_1146 (Conv2D) (None, 28, 28, 128) 57344 ['activation_1145[0][0]'] batch_normalization_1153 ( (None, 28, 28, 128) 512 ['conv2d_1146[0][0]'] BatchNormalization) activation_1146 (Activatio (None, 28, 28, 128) 0 ['batch_normalization_1153[0][n) 0]'] conv2d_1147 (Conv2D) (None, 28, 28, 32) 36864 ['activation_1146[0][0]'] concatenate_500 (Concatena (None, 28, 28, 480) 0 ['concatenate_499[0][0]', te) 'conv2d_1147[0][0]'] batch_normalization_1154 ( (None, 28, 28, 480) 1920 ['concatenate_500[0][0]'] BatchNormalization) activation_1147 (Activatio (None, 28, 28, 480) 0 ['batch_normalization_1154[0][n) 0]'] conv2d_1148 (Conv2D) (None, 28, 28, 128) 61440 ['activation_1147[0][0]'] batch_normalization_1155 ( (None, 28, 28, 128) 512 ['conv2d_1148[0][0]'] BatchNormalization) activation_1148 (Activatio (None, 28, 28, 128) 0 ['batch_normalization_1155[0][n) 0]'] conv2d_1149 (Conv2D) (None, 28, 28, 32) 36864 ['activation_1148[0][0]'] concatenate_501 (Concatena (None, 28, 28, 512) 0 ['concatenate_500[0][0]', te) 'conv2d_1149[0][0]'] batch_normalization_1156 ( (None, 28, 28, 512) 2048 ['concatenate_501[0][0]'] BatchNormalization) activation_1149 (Activatio (None, 28, 28, 512) 0 ['batch_normalization_1156[0][n) 0]'] conv2d_1150 (Conv2D) (None, 28, 28, 256) 131072 ['activation_1149[0][0]'] average_pooling2d_22 (Aver (None, 14, 14, 256) 0 ['conv2d_1150[0][0]'] agePooling2D) batch_normalization_1157 ( (None, 14, 14, 256) 1024 ['average_pooling2d_22[0][0]']BatchNormalization) activation_1150 (Activatio (None, 14, 14, 256) 0 ['batch_normalization_1157[0][n) 0]'] conv2d_1151 (Conv2D) (None, 14, 14, 128) 32768 ['activation_1150[0][0]'] batch_normalization_1158 ( (None, 14, 14, 128) 512 ['conv2d_1151[0][0]'] BatchNormalization) activation_1151 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1158[0][n) 0]'] conv2d_1152 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1151[0][0]'] concatenate_502 (Concatena (None, 14, 14, 288) 0 ['average_pooling2d_22[0][0]',te) 'conv2d_1152[0][0]'] batch_normalization_1159 ( (None, 14, 14, 288) 1152 ['concatenate_502[0][0]'] BatchNormalization) activation_1152 (Activatio (None, 14, 14, 288) 0 ['batch_normalization_1159[0][n) 0]'] conv2d_1153 (Conv2D) (None, 14, 14, 128) 36864 ['activation_1152[0][0]'] batch_normalization_1160 ( (None, 14, 14, 128) 512 ['conv2d_1153[0][0]'] BatchNormalization) activation_1153 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1160[0][n) 0]'] conv2d_1154 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1153[0][0]'] concatenate_503 (Concatena (None, 14, 14, 320) 0 ['concatenate_502[0][0]', te) 'conv2d_1154[0][0]'] batch_normalization_1161 ( (None, 14, 14, 320) 1280 ['concatenate_503[0][0]'] BatchNormalization) activation_1154 (Activatio (None, 14, 14, 320) 0 ['batch_normalization_1161[0][n) 0]'] conv2d_1155 (Conv2D) (None, 14, 14, 128) 40960 ['activation_1154[0][0]'] batch_normalization_1162 ( (None, 14, 14, 128) 512 ['conv2d_1155[0][0]'] BatchNormalization) activation_1155 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1162[0][n) 0]'] conv2d_1156 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1155[0][0]'] concatenate_504 (Concatena (None, 14, 14, 352) 0 ['concatenate_503[0][0]', te) 'conv2d_1156[0][0]'] batch_normalization_1163 ( (None, 14, 14, 352) 1408 ['concatenate_504[0][0]'] BatchNormalization) activation_1156 (Activatio (None, 14, 14, 352) 0 ['batch_normalization_1163[0][n) 0]'] conv2d_1157 (Conv2D) (None, 14, 14, 128) 45056 ['activation_1156[0][0]'] batch_normalization_1164 ( (None, 14, 14, 128) 512 ['conv2d_1157[0][0]'] BatchNormalization) activation_1157 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1164[0][n) 0]'] conv2d_1158 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1157[0][0]'] concatenate_505 (Concatena (None, 14, 14, 384) 0 ['concatenate_504[0][0]', te) 'conv2d_1158[0][0]'] batch_normalization_1165 ( (None, 14, 14, 384) 1536 ['concatenate_505[0][0]'] BatchNormalization) activation_1158 (Activatio (None, 14, 14, 384) 0 ['batch_normalization_1165[0][n) 0]'] conv2d_1159 (Conv2D) (None, 14, 14, 128) 49152 ['activation_1158[0][0]'] batch_normalization_1166 ( (None, 14, 14, 128) 512 ['conv2d_1159[0][0]'] BatchNormalization) activation_1159 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1166[0][n) 0]'] conv2d_1160 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1159[0][0]'] concatenate_506 (Concatena (None, 14, 14, 416) 0 ['concatenate_505[0][0]', te) 'conv2d_1160[0][0]'] batch_normalization_1167 ( (None, 14, 14, 416) 1664 ['concatenate_506[0][0]'] BatchNormalization) activation_1160 (Activatio (None, 14, 14, 416) 0 ['batch_normalization_1167[0][n) 0]'] conv2d_1161 (Conv2D) (None, 14, 14, 128) 53248 ['activation_1160[0][0]'] batch_normalization_1168 ( (None, 14, 14, 128) 512 ['conv2d_1161[0][0]'] BatchNormalization) activation_1161 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1168[0][n) 0]'] conv2d_1162 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1161[0][0]'] concatenate_507 (Concatena (None, 14, 14, 448) 0 ['concatenate_506[0][0]', te) 'conv2d_1162[0][0]'] batch_normalization_1169 ( (None, 14, 14, 448) 1792 ['concatenate_507[0][0]'] BatchNormalization) activation_1162 (Activatio (None, 14, 14, 448) 0 ['batch_normalization_1169[0][n) 0]'] conv2d_1163 (Conv2D) (None, 14, 14, 128) 57344 ['activation_1162[0][0]'] batch_normalization_1170 ( (None, 14, 14, 128) 512 ['conv2d_1163[0][0]'] BatchNormalization) activation_1163 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1170[0][n) 0]'] conv2d_1164 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1163[0][0]'] concatenate_508 (Concatena (None, 14, 14, 480) 0 ['concatenate_507[0][0]', te) 'conv2d_1164[0][0]'] batch_normalization_1171 ( (None, 14, 14, 480) 1920 ['concatenate_508[0][0]'] BatchNormalization) activation_1164 (Activatio (None, 14, 14, 480) 0 ['batch_normalization_1171[0][n) 0]'] conv2d_1165 (Conv2D) (None, 14, 14, 128) 61440 ['activation_1164[0][0]'] batch_normalization_1172 ( (None, 14, 14, 128) 512 ['conv2d_1165[0][0]'] BatchNormalization) activation_1165 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1172[0][n) 0]'] conv2d_1166 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1165[0][0]'] concatenate_509 (Concatena (None, 14, 14, 512) 0 ['concatenate_508[0][0]', te) 'conv2d_1166[0][0]'] batch_normalization_1173 ( (None, 14, 14, 512) 2048 ['concatenate_509[0][0]'] BatchNormalization) activation_1166 (Activatio (None, 14, 14, 512) 0 ['batch_normalization_1173[0][n) 0]'] conv2d_1167 (Conv2D) (None, 14, 14, 128) 65536 ['activation_1166[0][0]'] batch_normalization_1174 ( (None, 14, 14, 128) 512 ['conv2d_1167[0][0]'] BatchNormalization) activation_1167 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1174[0][n) 0]'] conv2d_1168 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1167[0][0]'] concatenate_510 (Concatena (None, 14, 14, 544) 0 ['concatenate_509[0][0]', te) 'conv2d_1168[0][0]'] batch_normalization_1175 ( (None, 14, 14, 544) 2176 ['concatenate_510[0][0]'] BatchNormalization) activation_1168 (Activatio (None, 14, 14, 544) 0 ['batch_normalization_1175[0][n) 0]'] conv2d_1169 (Conv2D) (None, 14, 14, 128) 69632 ['activation_1168[0][0]'] batch_normalization_1176 ( (None, 14, 14, 128) 512 ['conv2d_1169[0][0]'] BatchNormalization) activation_1169 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1176[0][n) 0]'] conv2d_1170 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1169[0][0]'] concatenate_511 (Concatena (None, 14, 14, 576) 0 ['concatenate_510[0][0]', te) 'conv2d_1170[0][0]'] batch_normalization_1177 ( (None, 14, 14, 576) 2304 ['concatenate_511[0][0]'] BatchNormalization) activation_1170 (Activatio (None, 14, 14, 576) 0 ['batch_normalization_1177[0][n) 0]'] conv2d_1171 (Conv2D) (None, 14, 14, 128) 73728 ['activation_1170[0][0]'] batch_normalization_1178 ( (None, 14, 14, 128) 512 ['conv2d_1171[0][0]'] BatchNormalization) activation_1171 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1178[0][n) 0]'] conv2d_1172 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1171[0][0]'] concatenate_512 (Concatena (None, 14, 14, 608) 0 ['concatenate_511[0][0]', te) 'conv2d_1172[0][0]'] batch_normalization_1179 ( (None, 14, 14, 608) 2432 ['concatenate_512[0][0]'] BatchNormalization) activation_1172 (Activatio (None, 14, 14, 608) 0 ['batch_normalization_1179[0][n) 0]'] conv2d_1173 (Conv2D) (None, 14, 14, 128) 77824 ['activation_1172[0][0]'] batch_normalization_1180 ( (None, 14, 14, 128) 512 ['conv2d_1173[0][0]'] BatchNormalization) activation_1173 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1180[0][n) 0]'] conv2d_1174 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1173[0][0]'] concatenate_513 (Concatena (None, 14, 14, 640) 0 ['concatenate_512[0][0]', te) 'conv2d_1174[0][0]'] batch_normalization_1181 ( (None, 14, 14, 640) 2560 ['concatenate_513[0][0]'] BatchNormalization) activation_1174 (Activatio (None, 14, 14, 640) 0 ['batch_normalization_1181[0][n) 0]'] conv2d_1175 (Conv2D) (None, 14, 14, 128) 81920 ['activation_1174[0][0]'] batch_normalization_1182 ( (None, 14, 14, 128) 512 ['conv2d_1175[0][0]'] BatchNormalization) activation_1175 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1182[0][n) 0]'] conv2d_1176 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1175[0][0]'] concatenate_514 (Concatena (None, 14, 14, 672) 0 ['concatenate_513[0][0]', te) 'conv2d_1176[0][0]'] batch_normalization_1183 ( (None, 14, 14, 672) 2688 ['concatenate_514[0][0]'] BatchNormalization) activation_1176 (Activatio (None, 14, 14, 672) 0 ['batch_normalization_1183[0][n) 0]'] conv2d_1177 (Conv2D) (None, 14, 14, 128) 86016 ['activation_1176[0][0]'] batch_normalization_1184 ( (None, 14, 14, 128) 512 ['conv2d_1177[0][0]'] BatchNormalization) activation_1177 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1184[0][n) 0]'] conv2d_1178 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1177[0][0]'] concatenate_515 (Concatena (None, 14, 14, 704) 0 ['concatenate_514[0][0]', te) 'conv2d_1178[0][0]'] batch_normalization_1185 ( (None, 14, 14, 704) 2816 ['concatenate_515[0][0]'] BatchNormalization) activation_1178 (Activatio (None, 14, 14, 704) 0 ['batch_normalization_1185[0][n) 0]'] conv2d_1179 (Conv2D) (None, 14, 14, 128) 90112 ['activation_1178[0][0]'] batch_normalization_1186 ( (None, 14, 14, 128) 512 ['conv2d_1179[0][0]'] BatchNormalization) activation_1179 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1186[0][n) 0]'] conv2d_1180 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1179[0][0]'] concatenate_516 (Concatena (None, 14, 14, 736) 0 ['concatenate_515[0][0]', te) 'conv2d_1180[0][0]'] batch_normalization_1187 ( (None, 14, 14, 736) 2944 ['concatenate_516[0][0]'] BatchNormalization) activation_1180 (Activatio (None, 14, 14, 736) 0 ['batch_normalization_1187[0][n) 0]'] conv2d_1181 (Conv2D) (None, 14, 14, 128) 94208 ['activation_1180[0][0]'] batch_normalization_1188 ( (None, 14, 14, 128) 512 ['conv2d_1181[0][0]'] BatchNormalization) activation_1181 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1188[0][n) 0]'] conv2d_1182 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1181[0][0]'] concatenate_517 (Concatena (None, 14, 14, 768) 0 ['concatenate_516[0][0]', te) 'conv2d_1182[0][0]'] batch_normalization_1189 ( (None, 14, 14, 768) 3072 ['concatenate_517[0][0]'] BatchNormalization) activation_1182 (Activatio (None, 14, 14, 768) 0 ['batch_normalization_1189[0][n) 0]'] conv2d_1183 (Conv2D) (None, 14, 14, 128) 98304 ['activation_1182[0][0]'] batch_normalization_1190 ( (None, 14, 14, 128) 512 ['conv2d_1183[0][0]'] BatchNormalization) activation_1183 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1190[0][n) 0]'] conv2d_1184 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1183[0][0]'] concatenate_518 (Concatena (None, 14, 14, 800) 0 ['concatenate_517[0][0]', te) 'conv2d_1184[0][0]'] batch_normalization_1191 ( (None, 14, 14, 800) 3200 ['concatenate_518[0][0]'] BatchNormalization) activation_1184 (Activatio (None, 14, 14, 800) 0 ['batch_normalization_1191[0][n) 0]'] conv2d_1185 (Conv2D) (None, 14, 14, 128) 102400 ['activation_1184[0][0]'] batch_normalization_1192 ( (None, 14, 14, 128) 512 ['conv2d_1185[0][0]'] BatchNormalization) activation_1185 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1192[0][n) 0]'] conv2d_1186 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1185[0][0]'] concatenate_519 (Concatena (None, 14, 14, 832) 0 ['concatenate_518[0][0]', te) 'conv2d_1186[0][0]'] batch_normalization_1193 ( (None, 14, 14, 832) 3328 ['concatenate_519[0][0]'] BatchNormalization) activation_1186 (Activatio (None, 14, 14, 832) 0 ['batch_normalization_1193[0][n) 0]'] conv2d_1187 (Conv2D) (None, 14, 14, 128) 106496 ['activation_1186[0][0]'] batch_normalization_1194 ( (None, 14, 14, 128) 512 ['conv2d_1187[0][0]'] BatchNormalization) activation_1187 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1194[0][n) 0]'] conv2d_1188 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1187[0][0]'] concatenate_520 (Concatena (None, 14, 14, 864) 0 ['concatenate_519[0][0]', te) 'conv2d_1188[0][0]'] batch_normalization_1195 ( (None, 14, 14, 864) 3456 ['concatenate_520[0][0]'] BatchNormalization) activation_1188 (Activatio (None, 14, 14, 864) 0 ['batch_normalization_1195[0][n) 0]'] conv2d_1189 (Conv2D) (None, 14, 14, 128) 110592 ['activation_1188[0][0]'] batch_normalization_1196 ( (None, 14, 14, 128) 512 ['conv2d_1189[0][0]'] BatchNormalization) activation_1189 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1196[0][n) 0]'] conv2d_1190 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1189[0][0]'] concatenate_521 (Concatena (None, 14, 14, 896) 0 ['concatenate_520[0][0]', te) 'conv2d_1190[0][0]'] batch_normalization_1197 ( (None, 14, 14, 896) 3584 ['concatenate_521[0][0]'] BatchNormalization) activation_1190 (Activatio (None, 14, 14, 896) 0 ['batch_normalization_1197[0][n) 0]'] conv2d_1191 (Conv2D) (None, 14, 14, 128) 114688 ['activation_1190[0][0]'] batch_normalization_1198 ( (None, 14, 14, 128) 512 ['conv2d_1191[0][0]'] BatchNormalization) activation_1191 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1198[0][n) 0]'] conv2d_1192 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1191[0][0]'] concatenate_522 (Concatena (None, 14, 14, 928) 0 ['concatenate_521[0][0]', te) 'conv2d_1192[0][0]'] batch_normalization_1199 ( (None, 14, 14, 928) 3712 ['concatenate_522[0][0]'] BatchNormalization) activation_1192 (Activatio (None, 14, 14, 928) 0 ['batch_normalization_1199[0][n) 0]'] conv2d_1193 (Conv2D) (None, 14, 14, 128) 118784 ['activation_1192[0][0]'] batch_normalization_1200 ( (None, 14, 14, 128) 512 ['conv2d_1193[0][0]'] BatchNormalization) activation_1193 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1200[0][n) 0]'] conv2d_1194 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1193[0][0]'] concatenate_523 (Concatena (None, 14, 14, 960) 0 ['concatenate_522[0][0]', te) 'conv2d_1194[0][0]'] batch_normalization_1201 ( (None, 14, 14, 960) 3840 ['concatenate_523[0][0]'] BatchNormalization) activation_1194 (Activatio (None, 14, 14, 960) 0 ['batch_normalization_1201[0][n) 0]'] conv2d_1195 (Conv2D) (None, 14, 14, 128) 122880 ['activation_1194[0][0]'] batch_normalization_1202 ( (None, 14, 14, 128) 512 ['conv2d_1195[0][0]'] BatchNormalization) activation_1195 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1202[0][n) 0]'] conv2d_1196 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1195[0][0]'] concatenate_524 (Concatena (None, 14, 14, 992) 0 ['concatenate_523[0][0]', te) 'conv2d_1196[0][0]'] batch_normalization_1203 ( (None, 14, 14, 992) 3968 ['concatenate_524[0][0]'] BatchNormalization) activation_1196 (Activatio (None, 14, 14, 992) 0 ['batch_normalization_1203[0][n) 0]'] conv2d_1197 (Conv2D) (None, 14, 14, 128) 126976 ['activation_1196[0][0]'] batch_normalization_1204 ( (None, 14, 14, 128) 512 ['conv2d_1197[0][0]'] BatchNormalization) activation_1197 (Activatio (None, 14, 14, 128) 0 ['batch_normalization_1204[0][n) 0]'] conv2d_1198 (Conv2D) (None, 14, 14, 32) 36864 ['activation_1197[0][0]'] concatenate_525 (Concatena (None, 14, 14, 1024) 0 ['concatenate_524[0][0]', te) 'conv2d_1198[0][0]'] batch_normalization_1205 ( (None, 14, 14, 1024) 4096 ['concatenate_525[0][0]'] BatchNormalization) activation_1198 (Activatio (None, 14, 14, 1024) 0 ['batch_normalization_1205[0][n) 0]'] conv2d_1199 (Conv2D) (None, 14, 14, 512) 524288 ['activation_1198[0][0]'] average_pooling2d_23 (Aver (None, 7, 7, 512) 0 ['conv2d_1199[0][0]'] agePooling2D) batch_normalization_1206 ( (None, 7, 7, 512) 2048 ['average_pooling2d_23[0][0]']BatchNormalization) activation_1199 (Activatio (None, 7, 7, 512) 0 ['batch_normalization_1206[0][n) 0]'] conv2d_1200 (Conv2D) (None, 7, 7, 128) 65536 ['activation_1199[0][0]'] batch_normalization_1207 ( (None, 7, 7, 128) 512 ['conv2d_1200[0][0]'] BatchNormalization) activation_1200 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1207[0][n) 0]'] conv2d_1201 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1200[0][0]'] concatenate_526 (Concatena (None, 7, 7, 544) 0 ['average_pooling2d_23[0][0]',te) 'conv2d_1201[0][0]'] batch_normalization_1208 ( (None, 7, 7, 544) 2176 ['concatenate_526[0][0]'] BatchNormalization) activation_1201 (Activatio (None, 7, 7, 544) 0 ['batch_normalization_1208[0][n) 0]'] conv2d_1202 (Conv2D) (None, 7, 7, 128) 69632 ['activation_1201[0][0]'] batch_normalization_1209 ( (None, 7, 7, 128) 512 ['conv2d_1202[0][0]'] BatchNormalization) activation_1202 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1209[0][n) 0]'] conv2d_1203 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1202[0][0]'] concatenate_527 (Concatena (None, 7, 7, 576) 0 ['concatenate_526[0][0]', te) 'conv2d_1203[0][0]'] batch_normalization_1210 ( (None, 7, 7, 576) 2304 ['concatenate_527[0][0]'] BatchNormalization) activation_1203 (Activatio (None, 7, 7, 576) 0 ['batch_normalization_1210[0][n) 0]'] conv2d_1204 (Conv2D) (None, 7, 7, 128) 73728 ['activation_1203[0][0]'] batch_normalization_1211 ( (None, 7, 7, 128) 512 ['conv2d_1204[0][0]'] BatchNormalization) activation_1204 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1211[0][n) 0]'] conv2d_1205 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1204[0][0]'] concatenate_528 (Concatena (None, 7, 7, 608) 0 ['concatenate_527[0][0]', te) 'conv2d_1205[0][0]'] batch_normalization_1212 ( (None, 7, 7, 608) 2432 ['concatenate_528[0][0]'] BatchNormalization) activation_1205 (Activatio (None, 7, 7, 608) 0 ['batch_normalization_1212[0][n) 0]'] conv2d_1206 (Conv2D) (None, 7, 7, 128) 77824 ['activation_1205[0][0]'] batch_normalization_1213 ( (None, 7, 7, 128) 512 ['conv2d_1206[0][0]'] BatchNormalization) activation_1206 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1213[0][n) 0]'] conv2d_1207 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1206[0][0]'] concatenate_529 (Concatena (None, 7, 7, 640) 0 ['concatenate_528[0][0]', te) 'conv2d_1207[0][0]'] batch_normalization_1214 ( (None, 7, 7, 640) 2560 ['concatenate_529[0][0]'] BatchNormalization) activation_1207 (Activatio (None, 7, 7, 640) 0 ['batch_normalization_1214[0][n) 0]'] conv2d_1208 (Conv2D) (None, 7, 7, 128) 81920 ['activation_1207[0][0]'] batch_normalization_1215 ( (None, 7, 7, 128) 512 ['conv2d_1208[0][0]'] BatchNormalization) activation_1208 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1215[0][n) 0]'] conv2d_1209 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1208[0][0]'] concatenate_530 (Concatena (None, 7, 7, 672) 0 ['concatenate_529[0][0]', te) 'conv2d_1209[0][0]'] batch_normalization_1216 ( (None, 7, 7, 672) 2688 ['concatenate_530[0][0]'] BatchNormalization) activation_1209 (Activatio (None, 7, 7, 672) 0 ['batch_normalization_1216[0][n) 0]'] conv2d_1210 (Conv2D) (None, 7, 7, 128) 86016 ['activation_1209[0][0]'] batch_normalization_1217 ( (None, 7, 7, 128) 512 ['conv2d_1210[0][0]'] BatchNormalization) activation_1210 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1217[0][n) 0]'] conv2d_1211 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1210[0][0]'] concatenate_531 (Concatena (None, 7, 7, 704) 0 ['concatenate_530[0][0]', te) 'conv2d_1211[0][0]'] batch_normalization_1218 ( (None, 7, 7, 704) 2816 ['concatenate_531[0][0]'] BatchNormalization) activation_1211 (Activatio (None, 7, 7, 704) 0 ['batch_normalization_1218[0][n) 0]'] conv2d_1212 (Conv2D) (None, 7, 7, 128) 90112 ['activation_1211[0][0]'] batch_normalization_1219 ( (None, 7, 7, 128) 512 ['conv2d_1212[0][0]'] BatchNormalization) activation_1212 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1219[0][n) 0]'] conv2d_1213 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1212[0][0]'] concatenate_532 (Concatena (None, 7, 7, 736) 0 ['concatenate_531[0][0]', te) 'conv2d_1213[0][0]'] batch_normalization_1220 ( (None, 7, 7, 736) 2944 ['concatenate_532[0][0]'] BatchNormalization) activation_1213 (Activatio (None, 7, 7, 736) 0 ['batch_normalization_1220[0][n) 0]'] conv2d_1214 (Conv2D) (None, 7, 7, 128) 94208 ['activation_1213[0][0]'] batch_normalization_1221 ( (None, 7, 7, 128) 512 ['conv2d_1214[0][0]'] BatchNormalization) activation_1214 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1221[0][n) 0]'] conv2d_1215 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1214[0][0]'] concatenate_533 (Concatena (None, 7, 7, 768) 0 ['concatenate_532[0][0]', te) 'conv2d_1215[0][0]'] batch_normalization_1222 ( (None, 7, 7, 768) 3072 ['concatenate_533[0][0]'] BatchNormalization) activation_1215 (Activatio (None, 7, 7, 768) 0 ['batch_normalization_1222[0][n) 0]'] conv2d_1216 (Conv2D) (None, 7, 7, 128) 98304 ['activation_1215[0][0]'] batch_normalization_1223 ( (None, 7, 7, 128) 512 ['conv2d_1216[0][0]'] BatchNormalization) activation_1216 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1223[0][n) 0]'] conv2d_1217 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1216[0][0]'] concatenate_534 (Concatena (None, 7, 7, 800) 0 ['concatenate_533[0][0]', te) 'conv2d_1217[0][0]'] batch_normalization_1224 ( (None, 7, 7, 800) 3200 ['concatenate_534[0][0]'] BatchNormalization) activation_1217 (Activatio (None, 7, 7, 800) 0 ['batch_normalization_1224[0][n) 0]'] conv2d_1218 (Conv2D) (None, 7, 7, 128) 102400 ['activation_1217[0][0]'] batch_normalization_1225 ( (None, 7, 7, 128) 512 ['conv2d_1218[0][0]'] BatchNormalization) activation_1218 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1225[0][n) 0]'] conv2d_1219 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1218[0][0]'] concatenate_535 (Concatena (None, 7, 7, 832) 0 ['concatenate_534[0][0]', te) 'conv2d_1219[0][0]'] batch_normalization_1226 ( (None, 7, 7, 832) 3328 ['concatenate_535[0][0]'] BatchNormalization) activation_1219 (Activatio (None, 7, 7, 832) 0 ['batch_normalization_1226[0][n) 0]'] conv2d_1220 (Conv2D) (None, 7, 7, 128) 106496 ['activation_1219[0][0]'] batch_normalization_1227 ( (None, 7, 7, 128) 512 ['conv2d_1220[0][0]'] BatchNormalization) activation_1220 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1227[0][n) 0]'] conv2d_1221 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1220[0][0]'] concatenate_536 (Concatena (None, 7, 7, 864) 0 ['concatenate_535[0][0]', te) 'conv2d_1221[0][0]'] batch_normalization_1228 ( (None, 7, 7, 864) 3456 ['concatenate_536[0][0]'] BatchNormalization) activation_1221 (Activatio (None, 7, 7, 864) 0 ['batch_normalization_1228[0][n) 0]'] conv2d_1222 (Conv2D) (None, 7, 7, 128) 110592 ['activation_1221[0][0]'] batch_normalization_1229 ( (None, 7, 7, 128) 512 ['conv2d_1222[0][0]'] BatchNormalization) activation_1222 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1229[0][n) 0]'] conv2d_1223 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1222[0][0]'] concatenate_537 (Concatena (None, 7, 7, 896) 0 ['concatenate_536[0][0]', te) 'conv2d_1223[0][0]'] batch_normalization_1230 ( (None, 7, 7, 896) 3584 ['concatenate_537[0][0]'] BatchNormalization) activation_1223 (Activatio (None, 7, 7, 896) 0 ['batch_normalization_1230[0][n) 0]'] conv2d_1224 (Conv2D) (None, 7, 7, 128) 114688 ['activation_1223[0][0]'] batch_normalization_1231 ( (None, 7, 7, 128) 512 ['conv2d_1224[0][0]'] BatchNormalization) activation_1224 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1231[0][n) 0]'] conv2d_1225 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1224[0][0]'] concatenate_538 (Concatena (None, 7, 7, 928) 0 ['concatenate_537[0][0]', te) 'conv2d_1225[0][0]'] batch_normalization_1232 ( (None, 7, 7, 928) 3712 ['concatenate_538[0][0]'] BatchNormalization) activation_1225 (Activatio (None, 7, 7, 928) 0 ['batch_normalization_1232[0][n) 0]'] conv2d_1226 (Conv2D) (None, 7, 7, 128) 118784 ['activation_1225[0][0]'] batch_normalization_1233 ( (None, 7, 7, 128) 512 ['conv2d_1226[0][0]'] BatchNormalization) activation_1226 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1233[0][n) 0]'] conv2d_1227 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1226[0][0]'] concatenate_539 (Concatena (None, 7, 7, 960) 0 ['concatenate_538[0][0]', te) 'conv2d_1227[0][0]'] batch_normalization_1234 ( (None, 7, 7, 960) 3840 ['concatenate_539[0][0]'] BatchNormalization) activation_1227 (Activatio (None, 7, 7, 960) 0 ['batch_normalization_1234[0][n) 0]'] conv2d_1228 (Conv2D) (None, 7, 7, 128) 122880 ['activation_1227[0][0]'] batch_normalization_1235 ( (None, 7, 7, 128) 512 ['conv2d_1228[0][0]'] BatchNormalization) activation_1228 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1235[0][n) 0]'] conv2d_1229 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1228[0][0]'] concatenate_540 (Concatena (None, 7, 7, 992) 0 ['concatenate_539[0][0]', te) 'conv2d_1229[0][0]'] batch_normalization_1236 ( (None, 7, 7, 992) 3968 ['concatenate_540[0][0]'] BatchNormalization) activation_1229 (Activatio (None, 7, 7, 992) 0 ['batch_normalization_1236[0][n) 0]'] conv2d_1230 (Conv2D) (None, 7, 7, 128) 126976 ['activation_1229[0][0]'] batch_normalization_1237 ( (None, 7, 7, 128) 512 ['conv2d_1230[0][0]'] BatchNormalization) activation_1230 (Activatio (None, 7, 7, 128) 0 ['batch_normalization_1237[0][n) 0]'] conv2d_1231 (Conv2D) (None, 7, 7, 32) 36864 ['activation_1230[0][0]'] concatenate_541 (Concatena (None, 7, 7, 1024) 0 ['concatenate_540[0][0]', te) 'conv2d_1231[0][0]'] batch_normalization_1238 ( (None, 7, 7, 1024) 4096 ['concatenate_541[0][0]'] BatchNormalization) activation_1231 (Activatio (None, 7, 7, 1024) 0 ['batch_normalization_1238[0][n) 0]'] global_average_pooling2d_7 (None, 1024) 0 ['activation_1231[0][0]'] (GlobalAveragePooling2D) dense_7 (Dense) (None, 4) 4100 ['global_average_pooling2d_7[0][0]'] ==================================================================================================
Total params: 7041604 (26.86 MB)
Trainable params: 6957956 (26.54 MB)
Non-trainable params: 83648 (326.75 KB)
__________________________________________________________________________________________________四、编译
在准备对模型进行训练之前，还需要再对其进行一些设置。以下内容是在模型的编译步骤中添加的：
损失函数（loss）：用于衡量模型在训练期间的准确率。
优化器（optimizer）：决定模型如何根据其看到的数据和自身的损失函数进行更新。
指标（metrics）：用于监控训练和测试步骤。以下示例使用了准确率，即被正确分类的图像的比率。
#设置优化器
opt=tf.keras.optimizers.Adam(learning_rate=1e-3)model.compile(optimizer=opt,loss=tf.keras.losses.SparseCategoricalCrossentropy(),metrics=['accuracy'])五、训练模型
epochs=10history=model.fit(train_ds,validation_data=val_ds,epochs=epochs
)运行结果：
Epoch 1/10
57/57 [==============================] - 176s 3s/step - loss: 1.1100 - accuracy: 0.6018 - val_loss: 65.5068 - val_accuracy: 0.3186
Epoch 2/10
57/57 [==============================] - 144s 3s/step - loss: 0.8184 - accuracy: 0.7102 - val_loss: 12.4867 - val_accuracy: 0.2743
Epoch 3/10
57/57 [==============================] - 145s 3s/step - loss: 0.7348 - accuracy: 0.7345 - val_loss: 13.2388 - val_accuracy: 0.3274
Epoch 4/10
57/57 [==============================] - 145s 3s/step - loss: 0.7246 - accuracy: 0.7500 - val_loss: 0.9510 - val_accuracy: 0.7257
Epoch 5/10
57/57 [==============================] - 146s 3s/step - loss: 0.6159 - accuracy: 0.7588 - val_loss: 0.6772 - val_accuracy: 0.7965
Epoch 6/10
57/57 [==============================] - 146s 3s/step - loss: 0.4837 - accuracy: 0.8363 - val_loss: 1.9455 - val_accuracy: 0.5221
Epoch 7/10
57/57 [==============================] - 146s 3s/step - loss: 0.5053 - accuracy: 0.8252 - val_loss: 1.6885 - val_accuracy: 0.4159
Epoch 8/10
57/57 [==============================] - 146s 3s/step - loss: 0.4130 - accuracy: 0.8518 - val_loss: 1.9283 - val_accuracy: 0.6726
Epoch 9/10
57/57 [==============================] - 146s 3s/step - loss: 0.4273 - accuracy: 0.8429 - val_loss: 2.2898 - val_accuracy: 0.4867
Epoch 10/10
57/57 [==============================] - 152s 3s/step - loss: 0.3546 - accuracy: 0.8695 - val_loss: 0.9306 - val_accuracy: 0.7345六、模型评估
acc=history.history['accuracy']
val_acc=history.history['val_accuracy']loss=history.history['loss']
val_loss=history.history['val_loss']epochs_range=range(epochs)plt.figure(figsize=(12,4))
plt.suptitle("OreoCC")plt.subplot(1,2,1)
plt.plot(epochs_range,acc,label='Training Accuracy')
plt.plot(epochs_range,val_acc,label='Validation Accuracy')
plt.legend(loc='lower right')
plt.title('Training and Validation Accuracy')plt.subplot(1,2,2)
plt.plot(epochs_range,loss,label='Training Loss')
plt.plot(epochs_range,val_loss,label='Validation Loss')
plt.legend(loc='upper right')
plt.title('Training and Validation Loss')
plt.show()运行结果：
七、预测
import numpy as np
#采用加载的模型（new_model）来看预测结果
plt.figure(figsize=(10,5))
plt.suptitle("OreoCC")for images,labels in val_ds.take(1):for i in range(8):ax=plt.subplot(2,4,i+1)#显示图片plt.imshow(images[i].numpy().astype("uint8"))#需要给图片增加一个维度img_array=tf.expand_dims(images[i],0)#使用模型预测图片中的人物predictions=model.predict(img_array)plt.title(classNames[np.argmax(predictions)])plt.axis("off")运行结果：
八、心得体会
本次项目中，体会了再TensorFlow环境下建立DenseNet模型的过程。深入了解了模型密集连接模式。这种设计促进了特征的重用，并鼓励梯度流动，有助于缓解深度学习中的梯度消失问题。
下面是DenseNet结构的关键组成部分：
初始卷积层：网络通常以一个标准的卷积层开始，用于初步提取输入图像的特征，并可能伴随有池化层来缩小输入尺寸。
Dense Blocks（密集块）：DenseNet的主要构建模块。每个密集块内，每新增一个层，都会将其输出特征图与之前所有层的输出特征图进行拼接（concatenation），作为下一个层的输入。这保证了信息流的高效传递和特征的复用。为了控制模型复杂度，每个层通过较小的增长率（growth rate）来增加特征图的数量，即每个层产生的新特征图数量。
Bottleneck Layers（瓶颈层）：为了减少计算成本，实际应用中的DenseNet常采用Bottleneck层设计。这些层首先使用1x1卷积来减少输入特征图的数量，然后是BN（Batch Normalization）和ReLU激活函数，接着是3x3卷积来提取特征。这样的设计保持了模型的效率，同时维持了特征的丰富性。
Transition Layers（过渡层）：位于Dense Blocks之间，用于过渡并控制模型的复杂度。过渡层通常包含1x1的卷积用于压缩特征图的通道数（使用压缩因子θ），以及可选的平均池化（Average Pooling）来进一步减小空间尺寸，帮助减少计算负担和过拟合风险。
分类层：网络的尾部通常包括全局平均池化（Global Average Pooling）层，用于将每个特征图的 spatial 维度压缩为一个值，随后连接一个或多个全连接层用于最终的分类或回归任务。
但模型的预测结果波动较大，初始loss也过于大，我想可能是由于模型层数较多，而数据集较少的原因造成的，故更换了大约有2000多张图片的数据集重新利用模型进行预测，结果如下：
可以看出，模型的准确率有所上升，但依然有些许波动，而loss也依然有所波动，考虑今后采用更大型数据集对模型进行验证。
