深度学习 Day27——J6ResNeXt-50实战解析
- 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 | 接辅导、项目定制
- 文章来源：K同学的学习圈子
文章目录
- 前言
- 1 我的环境
- 2 pytorch实现DenseNet算法
-
- 2.1 前期准备
-
- 2.1.1 引入库
- 2.1.2 设置GPU（如果设备上支持GPU就使用GPU,否则使用CPU）
- 2.1.3 导入数据
- 2.1.4 可视化数据
- 2.1.4 图像数据变换
- 2.1.4 划分数据集
- 2.1.4 加载数据
- 2.1.4 查看数据
- 2.2 搭建ResNeXt50模型
- 2.3 训练模型
-
- 2.3.1 设置超参数
- 2.3.2 编写训练函数
- 2.3.3 编写测试函数
- 2.3.4 正式训练
- 2.4 结果可视化
- 2.4 指定图片进行预测
- 2.6 模型评估
- 3 tensorflow实现DenseNet算法
-
- 3.1.引入库
- 3.2.设置GPU（如果使用的是CPU可以忽略这步）
- 3.3.导入数据
- 3.4.查看数据
- 3.5.加载数据
- 3.6.再次检查数据
- 3.7.配置数据集
- 3.8.可视化数据
- 3.9.构建ResNeXt50网络
- 3.10.编译模型
- 3.11.训练模型
- 3.12.模型评估
- 3.13.图像预测
- 4 知识点详解
-
- 4.1ResNeXt50详解
- 4.2 ResNeXt50对比ResNet50V2、DenseNet
-
- 4.2.1 网络结构
- 4.2.2 精度和计算量
- 4.2.3 适用范围
- 4 总结
前言
关键字： pytorch实现ResNeXt50详解算法，tensorflow实现ResNeXt50详解算法，ResNeXt50详解
1 我的环境
- 电脑系统：Windows 11
- 语言环境：python 3.8.6
- 编译器：pycharm2020.2.3
- 深度学习环境：
torch == 1.9.1+cu111
torchvision == 0.10.1+cu111
TensorFlow 2.10.1 - 显卡：NVIDIA GeForce RTX 4070
2 pytorch实现DenseNet算法
2.1 前期准备
2.1.1 引入库
import torch
import torch.nn as nn
import time
import copy
from torchvision import transforms, datasets
from pathlib import Path
from PIL import Image
import torchsummary as summary
import torch.nn.functional as F
from collections import OrderedDict
import re
import torch.utils.model_zoo as model_zoo
import matplotlib.pyplot as plt
plt.rcParams['font.sans-serif'] = ['SimHei'] # 用来正常显示中文标签
plt.rcParams['axes.unicode_minus'] = False # 用来正常显示负号
plt.rcParams['figure.dpi'] = 100 # 分辨率
import warnings
warnings.filterwarnings('ignore') # 忽略一些warning内容，无需打印
2.1.2 设置GPU（如果设备上支持GPU就使用GPU,否则使用CPU）
"""前期准备-设置GPU"""
# 如果设备上支持GPU就使用GPU,否则使用CPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("Using {} device".format(device))
输出
Using cuda device
2.1.3 导入数据
'''前期工作-导入数据'''
data_dir = r"D:\DeepLearning\data\monkeypox_recognition"
data_dir = Path(data_dir)
data_paths = list(data_dir.glob('*'))
classeNames = [str(path).split("\\")[-1] for path in data_paths]
print(classeNames)
输出
['Monkeypox', 'Others']
2.1.4 可视化数据
'''前期工作-可视化数据'''
subfolder = Path(data_dir) / "Monkeypox"
image_files = list(p.resolve() for p in subfolder.glob('*') if p.suffix in [".jpg", ".png", ".jpeg"])
plt.figure(figsize=(10, 6))
for i in range(len(image_files[:12])):
image_file = image_files[i]
ax = plt.subplot(3, 4, i + 1)
img = Image.open(str(image_file))
plt.imshow(img)
plt.axis("off")
# 显示图片
plt.tight_layout()
plt.show()
2.1.4 图像数据变换
'''前期工作-图像数据变换'''
total_datadir = data_dir
# 关于transforms.Compose的更多介绍可以参考：https://blog.csdn.net/qq_38251616/article/details/124878863
train_transforms = transforms.Compose([
transforms.Resize([224, 224]), # 将输入图片resize成统一尺寸
transforms.ToTensor(), # 将PIL Image或numpy.ndarray转换为tensor，并归一化到[0,1]之间
transforms.Normalize( # 标准化处理-->转换为标准正太分布（高斯分布），使模型更容易收敛
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]) # 其中 mean=[0.485,0.456,0.406]与std=[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。
])
total_data = datasets.ImageFolder(total_datadir, transform=train_transforms)
print(total_data)
print(total_data.class_to_idx)
输出
Dataset ImageFolder
Number of datapoints: 2142
Root location: D:\DeepLearning\data\monkeypox_recognition
StandardTransform
Transform: Compose(
Resize(size=[224, 224], interpolation=bilinear, max_size=None, antialias=None)
ToTensor()
Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
)
{'Monkeypox': 0, 'Others': 1}
2.1.4 划分数据集
'''前期工作-划分数据集'''
train_size = int(0.8 * len(total_data)) # train_size表示训练集大小，通过将总体数据长度的80%转换为整数得到；
test_size = len(total_data) - train_size # test_size表示测试集大小，是总体数据长度减去训练集大小。
# 使用torch.utils.data.random_split()方法进行数据集划分。该方法将总体数据total_data按照指定的大小比例（[train_size, test_size]）随机划分为训练集和测试集，
# 并将划分结果分别赋值给train_dataset和test_dataset两个变量。
train_dataset, test_dataset = torch.utils.data.random_split(total_data, [train_size, test_size])
print("train_dataset={}\ntest_dataset={}".format(train_dataset, test_dataset))
print("train_size={}\ntest_size={}".format(train_size, test_size))
输出
train_dataset=
test_dataset=
train_size=1713
test_size=429
2.1.4 加载数据
'''前期工作-加载数据'''
batch_size = 32
train_dl = torch.utils.data.DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=1)
test_dl = torch.utils.data.DataLoader(test_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=1)
2.1.4 查看数据
'''前期工作-查看数据'''
for X, y in test_dl:
print("Shape of X [N, C, H, W]: ", X.shape)
print("Shape of y: ", y.shape, y.dtype)
break
输出
Shape of X [N, C, H, W]: torch.Size([32, 3, 224, 224])
Shape of y: torch.Size([32]) torch.int64
2.2 搭建ResNeXt50模型
"""构建ResNeXt50网络"""
class BN_Conv2d(nn.Module):
"""
BN_CONV_RELU
"""
def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, groups=1, bias=False):
super(BN_Conv2d, self).__init__()
self.seq = nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride,
padding=padding, dilation=dilation, groups=groups, bias=bias),
nn.BatchNorm2d(out_channels)
)
def forward(self, x):
return F.relu(self.seq(x))
class ResNeXt_Block(nn.Module):
"""
ResNeXt block with group convolutions
"""
def __init__(self, in_chnls, cardinality, group_depth, stride):
super(ResNeXt_Block, self).__init__()
self.group_chnls = cardinality * group_depth
self.conv1 = BN_Conv2d(in_chnls, self.group_chnls, 1, stride=1, padding=0)
self.conv2 = BN_Conv2d(self.group_chnls, self.group_chnls, 3, stride=stride, padding=1, groups=cardinality)
self.conv3 = nn.Conv2d(self.group_chnls, self.group_chnls * 2, 1, stride=1, padding=0)
self.bn = nn.BatchNorm2d(self.group_chnls * 2)
self.short_cut = nn.Sequential(
nn.Conv2d(in_chnls, self.group_chnls * 2, 1, stride, 0, bias=False),
nn.BatchNorm2d(self.group_chnls * 2)
)
def forward(self, x):
out = self.conv1(x)
out = self.conv2(out)
out = self.bn(self.conv3(out))
out += self.short_cut(x)
return F.relu(out)
class ResNeXt(nn.Module):
"""
ResNeXt builder
"""
def __init__(self, layers: object, cardinality, group_depth, num_classes) -> object:
super(ResNeXt, self).__init__()
self.cardinality = cardinality
self.channels = 64
self.conv1 = BN_Conv2d(3, self.channels, 7, stride=2, padding=3)
d1 = group_depth
self.conv2 = self.___make_layers(d1, layers[0], stride=1)
d2 = d1 * 2
self.conv3 = self.___make_layers(d2, layers[1], stride=2)
d3 = d2 * 2
self.conv4 = self.___make_layers(d3, layers[2], stride=2)
d4 = d3 * 2
self.conv5 = self.___make_layers(d4, layers[3], stride=2)
self.fc = nn.Linear(self.channels, num_classes) # 224x224 input size
def ___make_layers(self, d, blocks, stride):
strides = [stride] + [1] * (blocks - 1)
layers = []
for stride in strides:
layers.append(ResNeXt_Block(self.channels, self.cardinality, d, stride))
self.channels = self.cardinality * d * 2
return nn.Sequential(*layers)
def forward(self, x):
out = self.conv1(x)
out = F.max_pool2d(out, 3, 2, 1)
out = self.conv2(out)
out = self.conv3(out)
out = self.conv4(out)
out = self.conv5(out)
out = F.avg_pool2d(out, 7)
out = out.view(out.size(0), -1)
out = F.softmax(self.fc(out), dim=1)
return out
该模型相比DenseNet的区别是，在最后一个denseblock后增加SE_layer。
# SE_layer
self.features.add_module('SE-module', Squeeze_excitation_layer(num_features))
输出
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv2d-1 [-1, 64, 112, 112] 9,408
BatchNorm2d-2 [-1, 64, 112, 112] 128
BN_Conv2d-3 [-1, 64, 112, 112] 0
Conv2d-4 [-1, 128, 56, 56] 8,192
BatchNorm2d-5 [-1, 128, 56, 56] 256
BN_Conv2d-6 [-1, 128, 56, 56] 0
Conv2d-7 [-1, 128, 56, 56] 4,608
BatchNorm2d-8 [-1, 128, 56, 56] 256
BN_Conv2d-9 [-1, 128, 56, 56] 0
Conv2d-10 [-1, 256, 56, 56] 33,024
BatchNorm2d-11 [-1, 256, 56, 56] 512
Conv2d-12 [-1, 256, 56, 56] 16,384
BatchNorm2d-13 [-1, 256, 56, 56] 512
ResNeXt_Block-14 [-1, 256, 56, 56] 0
Conv2d-15 [-1, 128, 56, 56] 32,768
BatchNorm2d-16 [-1, 128, 56, 56] 256
BN_Conv2d-17 [-1, 128, 56, 56] 0
Conv2d-18 [-1, 128, 56, 56] 4,608
BatchNorm2d-19 [-1, 128, 56, 56] 256
BN_Conv2d-20 [-1, 128, 56, 56] 0
Conv2d-21 [-1, 256, 56, 56] 33,024
BatchNorm2d-22 [-1, 256, 56, 56] 512
Conv2d-23 [-1, 256, 56, 56] 65,536
BatchNorm2d-24 [-1, 256, 56, 56] 512
ResNeXt_Block-25 [-1, 256, 56, 56] 0
Conv2d-26 [-1, 128, 56, 56] 32,768
BatchNorm2d-27 [-1, 128, 56, 56] 256
BN_Conv2d-28 [-1, 128, 56, 56] 0
Conv2d-29 [-1, 128, 56, 56] 4,608
BatchNorm2d-30 [-1, 128, 56, 56] 256
BN_Conv2d-31 [-1, 128, 56, 56] 0
Conv2d-32 [-1, 256, 56, 56] 33,024
BatchNorm2d-33 [-1, 256, 56, 56] 512
Conv2d-34 [-1, 256, 56, 56] 65,536
BatchNorm2d-35 [-1, 256, 56, 56] 512
ResNeXt_Block-36 [-1, 256, 56, 56] 0
Conv2d-37 [-1, 256, 56, 56] 65,536
BatchNorm2d-38 [-1, 256, 56, 56] 512
BN_Conv2d-39 [-1, 256, 56, 56] 0
Conv2d-40 [-1, 256, 28, 28] 18,432
BatchNorm2d-41 [-1, 256, 28, 28] 512
BN_Conv2d-42 [-1, 256, 28, 28] 0
Conv2d-43 [-1, 512, 28, 28] 131,584
BatchNorm2d-44 [-1, 512, 28, 28] 1,024
Conv2d-45 [-1, 512, 28, 28] 131,072
BatchNorm2d-46 [-1, 512, 28, 28] 1,024
ResNeXt_Block-47 [-1, 512, 28, 28] 0
Conv2d-48 [-1, 256, 28, 28] 131,072
BatchNorm2d-49 [-1, 256, 28, 28] 512
BN_Conv2d-50 [-1, 256, 28, 28] 0
Conv2d-51 [-1, 256, 28, 28] 18,432
BatchNorm2d-52 [-1, 256, 28, 28] 512
BN_Conv2d-53 [-1, 256, 28, 28] 0
Conv2d-54 [-1, 512, 28, 28] 131,584
BatchNorm2d-55 [-1, 512, 28, 28] 1,024
Conv2d-56 [-1, 512, 28, 28] 262,144
BatchNorm2d-57 [-1, 512, 28, 28] 1,024
ResNeXt_Block-58 [-1, 512, 28, 28] 0
Conv2d-59 [-1, 256, 28, 28] 131,072
BatchNorm2d-60 [-1, 256, 28, 28] 512
BN_Conv2d-61 [-1, 256, 28, 28] 0
Conv2d-62 [-1, 256, 28, 28] 18,432
BatchNorm2d-63 [-1, 256, 28, 28] 512
BN_Conv2d-64 [-1, 256, 28, 28] 0
Conv2d-65 [-1, 512, 28, 28] 131,584
BatchNorm2d-66 [-1, 512, 28, 28] 1,024
Conv2d-67 [-1, 512, 28, 28] 262,144
BatchNorm2d-68 [-1, 512, 28, 28] 1,024
ResNeXt_Block-69 [-1, 512, 28, 28] 0
Conv2d-70 [-1, 256, 28, 28] 131,072
BatchNorm2d-71 [-1, 256, 28, 28] 512
BN_Conv2d-72 [-1, 256, 28, 28] 0
Conv2d-73 [-1, 256, 28, 28] 18,432
BatchNorm2d-74 [-1, 256, 28, 28] 512
BN_Conv2d-75 [-1, 256, 28, 28] 0
Conv2d-76 [-1, 512, 28, 28] 131,584
BatchNorm2d-77 [-1, 512, 28, 28] 1,024
Conv2d-78 [-1, 512, 28, 28] 262,144
BatchNorm2d-79 [-1, 512, 28, 28] 1,024
ResNeXt_Block-80 [-1, 512, 28, 28] 0
Conv2d-81 [-1, 512, 28, 28] 262,144
BatchNorm2d-82 [-1, 512, 28, 28] 1,024
BN_Conv2d-83 [-1, 512, 28, 28] 0
Conv2d-84 [-1, 512, 14, 14] 73,728
BatchNorm2d-85 [-1, 512, 14, 14] 1,024
BN_Conv2d-86 [-1, 512, 14, 14] 0
Conv2d-87 [-1, 1024, 14, 14] 525,312
BatchNorm2d-88 [-1, 1024, 14, 14] 2,048
Conv2d-89 [-1, 1024, 14, 14] 524,288
BatchNorm2d-90 [-1, 1024, 14, 14] 2,048
ResNeXt_Block-91 [-1, 1024, 14, 14] 0
Conv2d-92 [-1, 512, 14, 14] 524,288
BatchNorm2d-93 [-1, 512, 14, 14] 1,024
BN_Conv2d-94 [-1, 512, 14, 14] 0
Conv2d-95 [-1, 512, 14, 14] 73,728
BatchNorm2d-96 [-1, 512, 14, 14] 1,024
BN_Conv2d-97 [-1, 512, 14, 14] 0
Conv2d-98 [-1, 1024, 14, 14] 525,312
BatchNorm2d-99 [-1, 1024, 14, 14] 2,048
Conv2d-100 [-1, 1024, 14, 14] 1,048,576
BatchNorm2d-101 [-1, 1024, 14, 14] 2,048
ResNeXt_Block-102 [-1, 1024, 14, 14] 0
Conv2d-103 [-1, 512, 14, 14] 524,288
BatchNorm2d-104 [-1, 512, 14, 14] 1,024
BN_Conv2d-105 [-1, 512, 14, 14] 0
Conv2d-106 [-1, 512, 14, 14] 73,728
BatchNorm2d-107 [-1, 512, 14, 14] 1,024
BN_Conv2d-108 [-1, 512, 14, 14] 0
Conv2d-109 [-1, 1024, 14, 14] 525,312
BatchNorm2d-110 [-1, 1024, 14, 14] 2,048
Conv2d-111 [-1, 1024, 14, 14] 1,048,576
BatchNorm2d-112 [-1, 1024, 14, 14] 2,048
ResNeXt_Block-113 [-1, 1024, 14, 14] 0
Conv2d-114 [-1, 512, 14, 14] 524,288
BatchNorm2d-115 [-1, 512, 14, 14] 1,024
BN_Conv2d-116 [-1, 512, 14, 14] 0
Conv2d-117 [-1, 512, 14, 14] 73,728
BatchNorm2d-118 [-1, 512, 14, 14] 1,024
BN_Conv2d-119 [-1, 512, 14, 14] 0
Conv2d-120 [-1, 1024, 14, 14] 525,312
BatchNorm2d-121 [-1, 1024, 14, 14] 2,048
Conv2d-122 [-1, 1024, 14, 14] 1,048,576
BatchNorm2d-123 [-1, 1024, 14, 14] 2,048
ResNeXt_Block-124 [-1, 1024, 14, 14] 0
Conv2d-125 [-1, 512, 14, 14] 524,288
BatchNorm2d-126 [-1, 512, 14, 14] 1,024
BN_Conv2d-127 [-1, 512, 14, 14] 0
Conv2d-128 [-1, 512, 14, 14] 73,728
BatchNorm2d-129 [-1, 512, 14, 14] 1,024
BN_Conv2d-130 [-1, 512, 14, 14] 0
Conv2d-131 [-1, 1024, 14, 14] 525,312
BatchNorm2d-132 [-1, 1024, 14, 14] 2,048
Conv2d-133 [-1, 1024, 14, 14] 1,048,576
BatchNorm2d-134 [-1, 1024, 14, 14] 2,048
ResNeXt_Block-135 [-1, 1024, 14, 14] 0
Conv2d-136 [-1, 512, 14, 14] 524,288
BatchNorm2d-137 [-1, 512, 14, 14] 1,024
BN_Conv2d-138 [-1, 512, 14, 14] 0
Conv2d-139 [-1, 512, 14, 14] 73,728
BatchNorm2d-140 [-1, 512, 14, 14] 1,024
BN_Conv2d-141 [-1, 512, 14, 14] 0
Conv2d-142 [-1, 1024, 14, 14] 525,312
BatchNorm2d-143 [-1, 1024, 14, 14] 2,048
Conv2d-144 [-1, 1024, 14, 14] 1,048,576
BatchNorm2d-145 [-1, 1024, 14, 14] 2,048
ResNeXt_Block-146 [-1, 1024, 14, 14] 0
Conv2d-147 [-1, 1024, 14, 14] 1,048,576
BatchNorm2d-148 [-1, 1024, 14, 14] 2,048
BN_Conv2d-149 [-1, 1024, 14, 14] 0
Conv2d-150 [-1, 1024, 7, 7] 294,912
BatchNorm2d-151 [-1, 1024, 7, 7] 2,048
BN_Conv2d-152 [-1, 1024, 7, 7] 0
Conv2d-153 [-1, 2048, 7, 7] 2,099,200
BatchNorm2d-154 [-1, 2048, 7, 7] 4,096
Conv2d-155 [-1, 2048, 7, 7] 2,097,152
BatchNorm2d-156 [-1, 2048, 7, 7] 4,096
ResNeXt_Block-157 [-1, 2048, 7, 7] 0
Conv2d-158 [-1, 1024, 7, 7] 2,097,152
BatchNorm2d-159 [-1, 1024, 7, 7] 2,048
BN_Conv2d-160 [-1, 1024, 7, 7] 0
Conv2d-161 [-1, 1024, 7, 7] 294,912
BatchNorm2d-162 [-1, 1024, 7, 7] 2,048
BN_Conv2d-163 [-1, 1024, 7, 7] 0
Conv2d-164 [-1, 2048, 7, 7] 2,099,200
BatchNorm2d-165 [-1, 2048, 7, 7] 4,096
Conv2d-166 [-1, 2048, 7, 7] 4,194,304
BatchNorm2d-167 [-1, 2048, 7, 7] 4,096
ResNeXt_Block-168 [-1, 2048, 7, 7] 0
Conv2d-169 [-1, 1024, 7, 7] 2,097,152
BatchNorm2d-170 [-1, 1024, 7, 7] 2,048
BN_Conv2d-171 [-1, 1024, 7, 7] 0
Conv2d-172 [-1, 1024, 7, 7] 294,912
BatchNorm2d-173 [-1, 1024, 7, 7] 2,048
BN_Conv2d-174 [-1, 1024, 7, 7] 0
Conv2d-175 [-1, 2048, 7, 7] 2,099,200
BatchNorm2d-176 [-1, 2048, 7, 7] 4,096
Conv2d-177 [-1, 2048, 7, 7] 4,194,304
BatchNorm2d-178 [-1, 2048, 7, 7] 4,096
ResNeXt_Block-179 [-1, 2048, 7, 7] 0
Linear-180 [-1, 4] 8,196
================================================================
Total params: 37,574,724
Trainable params: 37,574,724
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 379.37
Params size (MB): 143.34
Estimated Total Size (MB): 523.28
----------------------------------------------------------------
None
2.3 训练模型
2.3.1 设置超参数
"""训练模型--设置超参数"""
loss_fn = nn.CrossEntropyLoss() # 创建损失函数，计算实际输出和真实相差多少，交叉熵损失函数，事实上，它就是做图片分类任务时常用的损失函数
learn_rate = 1e-4 # 学习率
optimizer1 = torch.optim.SGD(model.parameters(), lr=learn_rate)# 作用是定义优化器，用来训练时候优化模型参数；其中，SGD表示随机梯度下降，用于控制实际输出y与真实y之间的相差有多大
optimizer2 = torch.optim.Adam(model.parameters(), lr=learn_rate)
lr_opt = optimizer2
model_opt = optimizer2
# 调用官方动态学习率接口时使用2
lambda1 = lambda epoch : 0.92 ** (epoch // 4)
# optimizer = torch.optim.SGD(model.parameters(), lr=learn_rate)
scheduler = torch.optim.lr_scheduler.LambdaLR(lr_opt, lr_lambda=lambda1) #选定调整方法
2.3.2 编写训练函数
"""训练模型--编写训练函数"""
# 训练循环
def train(dataloader, model, loss_fn, optimizer):
size = len(dataloader.dataset) # 训练集的大小，一共60000张图片
num_batches = len(dataloader) # 批次数目，1875（60000/32）
train_loss, train_acc = 0, 0 # 初始化训练损失和正确率
for X, y in dataloader: # 加载数据加载器，得到里面的 X（图片数据）和 y（真实标签）
X, y = X.to(device), y.to(device) # 用于将数据存到显卡
# 计算预测误差
pred = model(X) # 网络输出
loss = loss_fn(pred, y) # 计算网络输出和真实值之间的差距，targets为真实值，计算二者差值即为损失
# 反向传播
optimizer.zero_grad() # 清空过往梯度
loss.backward() # 反向传播，计算当前梯度
optimizer.step() # 根据梯度更新网络参数
# 记录acc与loss
train_acc += (pred.argmax(1) == y).type(torch.float).sum().item()
train_loss += loss.item()
train_acc /= size
train_loss /= num_batches
return train_acc, train_loss
2.3.3 编写测试函数
"""训练模型--编写测试函数"""
# 测试函数和训练函数大致相同，但是由于不进行梯度下降对网络权重进行更新，所以不需要传入优化器
def test(dataloader, model, loss_fn):
size = len(dataloader.dataset) # 测试集的大小，一共10000张图片
num_batches = len(dataloader) # 批次数目，313（10000/32=312.5，向上取整）
test_loss, test_acc = 0, 0
# 当不进行训练时，停止梯度更新，节省计算内存消耗
with torch.no_grad(): # 测试时模型参数不用更新，所以 no_grad，整个模型参数正向推就ok，不反向更新参数
for imgs, target in dataloader:
imgs, target = imgs.to(device), target.to(device)
# 计算loss
target_pred = model(imgs)
loss = loss_fn(target_pred, target)
test_loss += loss.item()
test_acc += (target_pred.argmax(1) == target).type(torch.float).sum().item()#统计预测正确的个数
test_acc /= size
test_loss /= num_batches
return test_acc, test_loss
2.3.4 正式训练
"""训练模型--正式训练"""
epochs = 20
train_loss = []
train_acc = []
test_loss = []
test_acc = []
best_test_acc=0
for epoch in range(epochs):
milliseconds_t1 = int(time.time() * 1000)
# 更新学习率（使用自定义学习率时使用）
# adjust_learning_rate(lr_opt, epoch, learn_rate)
model.train()
epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, model_opt)
scheduler.step() # 更新学习率（调用官方动态学习率接口时使用）
model.eval()
epoch_test_acc, epoch_test_loss = test(test_dl, model, loss_fn)
train_acc.append(epoch_train_acc)
train_loss.append(epoch_train_loss)
test_acc.append(epoch_test_acc)
test_loss.append(epoch_test_loss)
# 获取当前的学习率
lr = lr_opt.state_dict()['param_groups'][0]['lr']
milliseconds_t2 = int(time.time() * 1000)
template = ('Epoch:{:2d}, duration:{}ms, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%，Test_loss:{:.3f}, Lr:{:.2E}')
if best_test_acc < epoch_test_acc:
best_test_acc = epoch_test_acc
#备份最好的模型
best_model = copy.deepcopy(model)
template = (
'Epoch:{:2d}, duration:{}ms, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%，Test_loss:{:.3f}, Lr:{:.2E},Update the best model')
print(
template.format(epoch + 1, milliseconds_t2-milliseconds_t1, epoch_train_acc * 100, epoch_train_loss, epoch_test_acc * 100, epoch_test_loss, lr))
# 保存最佳模型到文件中
PATH = './best_model.pth' # 保存的参数文件名
torch.save(model.state_dict(), PATH)
print('Done')
Epoch: 1, duration:15650ms, Train_acc:54.8%, Train_loss:1.187, Test_acc:59.9%，Test_loss:1.147, Lr:1.00E-04,Update the best model
Epoch: 2, duration:15311ms, Train_acc:62.2%, Train_loss:1.112, Test_acc:58.7%，Test_loss:1.150, Lr:1.00E-04
Epoch: 3, duration:15336ms, Train_acc:67.3%, Train_loss:1.067, Test_acc:62.9%，Test_loss:1.117, Lr:1.00E-04,Update the best model
Epoch: 4, duration:14853ms, Train_acc:68.0%, Train_loss:1.061, Test_acc:65.0%，Test_loss:1.093, Lr:1.00E-04,Update the best model
Epoch: 5, duration:14930ms, Train_acc:68.0%, Train_loss:1.059, Test_acc:64.6%，Test_loss:1.087, Lr:1.00E-04
Epoch: 6, duration:15118ms, Train_acc:67.2%, Train_loss:1.067, Test_acc:60.1%，Test_loss:1.126, Lr:1.00E-04
Epoch: 7, duration:15024ms, Train_acc:67.8%, Train_loss:1.059, Test_acc:68.5%，Test_loss:1.050, Lr:1.00E-04,Update the best model
Epoch: 8, duration:14973ms, Train_acc:66.9%, Train_loss:1.065, Test_acc:67.6%，Test_loss:1.074, Lr:1.00E-04
Epoch: 9, duration:14902ms, Train_acc:69.3%, Train_loss:1.049, Test_acc:64.1%，Test_loss:1.099, Lr:1.00E-04
Epoch:10, duration:15237ms, Train_acc:70.2%, Train_loss:1.035, Test_acc:71.6%，Test_loss:1.024, Lr:1.00E-04,Update the best model
Epoch:11, duration:14890ms, Train_acc:71.0%, Train_loss:1.029, Test_acc:73.4%，Test_loss:1.010, Lr:1.00E-04,Update the best model
Epoch:12, duration:14951ms, Train_acc:70.5%, Train_loss:1.034, Test_acc:70.2%，Test_loss:1.043, Lr:1.00E-04
Epoch:13, duration:14967ms, Train_acc:72.3%, Train_loss:1.020, Test_acc:71.8%，Test_loss:1.022, Lr:1.00E-04
Epoch:14, duration:14966ms, Train_acc:73.8%, Train_loss:1.004, Test_acc:72.5%，Test_loss:1.017, Lr:1.00E-04
Epoch:15, duration:14886ms, Train_acc:75.5%, Train_loss:0.987, Test_acc:72.3%，Test_loss:1.015, Lr:1.00E-04
Epoch:16, duration:14895ms, Train_acc:72.6%, Train_loss:1.012, Test_acc:72.5%，Test_loss:1.025, Lr:1.00E-04
Epoch:17, duration:15037ms, Train_acc:74.3%, Train_loss:0.994, Test_acc:73.2%，Test_loss:1.016, Lr:1.00E-04
Epoch:18, duration:14797ms, Train_acc:76.5%, Train_loss:0.976, Test_acc:70.6%，Test_loss:1.026, Lr:1.00E-04
Epoch:19, duration:15157ms, Train_acc:72.6%, Train_loss:1.018, Test_acc:72.0%，Test_loss:1.018, Lr:1.00E-04
Epoch:20, duration:14767ms, Train_acc:73.1%, Train_loss:1.009, Test_acc:74.4%，Test_loss:1.003, Lr:1.00E-04,Update the best model
2.4 结果可视化
"""训练模型--结果可视化"""
epochs_range = range(epochs)
plt.figure(figsize=(12, 3))
plt.subplot(1, 2, 1)
plt.plot(epochs_range, train_acc, label='Training Accuracy')
plt.plot(epochs_range, test_acc, label='Test Accuracy')
plt.legend(loc='lower right')
plt.title('Training and Validation Accuracy')
plt.subplot(1, 2, 2)
plt.plot(epochs_range, train_loss, label='Training Loss')
plt.plot(epochs_range, test_loss, label='Test Loss')
plt.legend(loc='upper right')
plt.title('Training and Validation Loss')
plt.show()
2.4 指定图片进行预测
def predict_one_image(image_path, model, transform, classes):
test_img = Image.open(image_path).convert('RGB')
plt.imshow(test_img) # 展示预测的图片
plt.show()
test_img = transform(test_img)
img = test_img.to(device).unsqueeze(0)
model.eval()
output = model(img)
_, pred = torch.max(output, 1)
pred_class = classes[pred]
print(f'预测结果是：{pred_class}')
# 将参数加载到model当中
model.load_state_dict(torch.load(PATH, map_location=device))
"""指定图片进行预测"""
classes = list(total_data.class_to_idx)
# 预测训练集中的某张照片
predict_one_image(image_path=str(Path(data_dir) / "Monkeypox/M01_01_01.jpg"),
model=model,
transform=train_transforms,
classes=classes)
输出
预测结果是：Monkeypox
2.6 模型评估
"""模型评估"""
best_model.eval()
epoch_test_acc, epoch_test_loss = test(test_dl, best_model, loss_fn)
# 查看是否与我们记录的最高准确率一致
print(epoch_test_acc, epoch_test_loss)
输出
0.7435897435897436 0.9976991329874311
3 tensorflow实现DenseNet算法
3.1.引入库
from PIL import Image
import numpy as np
from pathlib import Path
import matplotlib.pyplot as plt
# 支持中文
plt.rcParams['font.sans-serif'] = ['SimHei'] # 用来正常显示中文标签
plt.rcParams['axes.unicode_minus'] = False # 用来正常显示负号
import tensorflow as tf
from keras import layers, models, Input
from keras.layers import Input, Activation, BatchNormalization, Flatten
from keras.layers import Dense, Conv2D, MaxPooling2D, ZeroPadding2D, GlobalMaxPooling2D, AveragePooling2D, Flatten, \
Dropout, BatchNormalization, GlobalAveragePooling2D
from keras.models import Model
from keras import regularizers
from tensorflow import keras
from keras.callbacks import ModelCheckpoint
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings('ignore') # 忽略一些warning内容，无需打印
3.2.设置GPU（如果使用的是CPU可以忽略这步）
'''前期工作-设置GPU（如果使用的是CPU可以忽略这步）'''
# 检查GPU是否可用
print(tf.test.is_built_with_cuda())
gpus = tf.config.list_physical_devices("GPU")
print(gpus)
if gpus:
gpu0 = gpus[0] # 如果有多个GPU，仅使用第0个GPU
tf.config.experimental.set_memory_growth(gpu0, True) # 设置GPU显存用量按需使用
tf.config.set_visible_devices([gpu0], "GPU")
执行结果
True
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]
3.3.导入数据
'''前期工作-导入数据'''
data_dir = r"D:\DeepLearning\data\monkeypox_recognition"
data_dir = Path(data_dir)
3.4.查看数据
'''前期工作-查看数据'''
image_count = len(list(data_dir.glob('*/*.jpg')))
print("图片总数为：", image_count)
image_list = list(data_dir.glob('Monkeypox/*.jpg'))
image = Image.open(str(image_list[1]))
# 查看图像实例的属性
print(image.format, image.size, image.mode)
plt.imshow(image)
plt.axis("off")
plt.show()
执行结果：
图片总数为： 2142
JPEG (224, 224) RGB
3.5.加载数据
'''数据预处理-加载数据'''
batch_size = 32
img_height = 224
img_width = 224
"""
关于image_dataset_from_directory()的详细介绍可以参考文章：https://mtyjkh.blog.csdn.net/article/details/117018789
"""
train_ds = tf.keras.preprocessing.image_dataset_from_directory(
data_dir,
validation_split=0.2,
subset="training",
seed=123,
image_size=(img_height, img_width),
batch_size=batch_size)
val_ds = tf.keras.preprocessing.image_dataset_from_directory(
data_dir,
validation_split=0.2,
subset="validation",
seed=123,
image_size=(img_height, img_width),
batch_size=batch_size)
class_names = train_ds.class_names
print(class_names)
运行结果：
Found 2142 files belonging to 2 classes.
Using 1714 files for training.
Found 2142 files belonging to 2 classes.
Using 428 files for validation.
['Monkeypox', 'Others']
3.6.再次检查数据
'''数据预处理-再次检查数据'''
# Image_batch是形状的张量(16, 336, 336, 3)。这是一批形状336x336x3的16张图片（最后一维指的是彩色通道RGB）。
# Label_batch是形状（16，）的张量，这些标签对应16张图片
for image_batch, labels_batch in train_ds:
print(image_batch.shape)
print(labels_batch.shape)
break
运行结果
(32, 224, 224, 3)
(32,)
3.7.配置数据集
'''数据预处理-配置数据集'''
AUTOTUNE = tf.data.AUTOTUNE
train_ds = train_ds.cache().shuffle(1000).prefetch(buffer_size=AUTOTUNE)
val_ds = val_ds.cache().prefetch(buffer_size=AUTOTUNE)
3.8.可视化数据
'''数据预处理-可视化数据'''
plt.figure(figsize=(10, 5))
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(class_names[labels[i]], fontsize=10)
plt.axis("off")
# 显示图片
plt.show()
3.9.构建ResNeXt50网络
"""构建ResNeXt50网络"""
# ----------------------- #
# groups代表多少组
# g_channels代表每组的特征图数量
# ----------------------- #
def group_conv2_block(x_0, strides, groups, g_channels):
g_list = []
for i in range(groups):
x = Lambda(lambda x: x[:, :, :, i * g_channels: (i + 1) * g_channels])(x_0)
x = Conv2D(filters=g_channels, kernel_size=3, strides=strides, padding='same', use_bias=False)(x)
g_list.append(x)
x = concatenate(g_list, axis=3)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = Activation('relu')(x)
return x
# 结构快
def block(x, filters, strides=1, groups=32, conv_short=True):
if conv_short:
short_cut = Conv2D(filters=filters * 2, kernel_size=1, strides=strides, padding='same')(x)
short_cut = BatchNormalization(epsilon=1.001e-5)(short_cut)
else:
short_cut = x
# 三层卷积
x = Conv2D(filters=filters, kernel_size=1, strides=1, padding='same')(x)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = Activation('relu')(x)
g_channels = int(filters / groups)
x = group_conv2_block(x, strides=strides, groups=groups, g_channels=g_channels)
x = Conv2D(filters=filters * 2, kernel_size=1, strides=1, padding='same')(x)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = Add()([x, short_cut])
x = Activation('relu')(x)
return x
def Resnext(inputs, classes):
x_input = keras.layers.Input(shape=inputs)
x = ZeroPadding2D((3, 3))(x_input)
x = Conv2D(filters=64, kernel_size=7, strides=2, padding='valid')(x)
x = BatchNormalization(epsilon=1.001e-5)(x)
x = Activation('relu')(x)
x = ZeroPadding2D((1, 1))(x)
x = MaxPool2D(pool_size=3, strides=2, padding='valid')(x)
x = block(x, filters=128, strides=1, conv_short=True)
x = block(x, filters=128, conv_short=False)
x = block(x, filters=128, conv_short=False)
x = block(x, filters=256, strides=2, conv_short=True)
x = block(x, filters=256, conv_short=False)
x = block(x, filters=256, conv_short=False)
x = block(x, filters=256, conv_short=False)
x = block(x, filters=512, strides=2, conv_short=True)
x = block(x, filters=512, conv_short=False)
x = block(x, filters=512, conv_short=False)
x = block(x, filters=512, conv_short=False)
x = block(x, filters=512, conv_short=False)
x = block(x, filters=512, conv_short=False)
x = block(x, filters=1024, strides=2, conv_short=True)
x = block(x, filters=1024, conv_short=False)
x = block(x, filters=1024, conv_short=False)
x = GlobalAvgPool2D()(x)
x = Dense(classes, activation='softmax')(x)
model = keras.models.Model(inputs=[x_input], outputs=[x])
return model
model = Resnext(inputs= (img_width, img_height, 3),classes=4)
model.summary()
网络结构结果如下：
Model: "model"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 224, 224, 3 0 []
)]
conv2d (Conv2D) (None, 112, 112, 64 9408 ['input_1[0][0]']
)
batch_normalization (BatchNorm (None, 112, 112, 64 256 ['conv2d[0][0]']
alization) )
max_pooling2d (MaxPooling2D) (None, 56, 56, 64) 0 ['batch_normalization[0][0]']
batch_normalization_1 (BatchNo (None, 56, 56, 64) 256 ['max_pooling2d[0][0]']
rmalization)
activation (Activation) (None, 56, 56, 64) 0 ['batch_normalization_1[0][0]']
conv2d_1 (Conv2D) (None, 56, 56, 128) 8192 ['activation[0][0]']
batch_normalization_2 (BatchNo (None, 56, 56, 128) 512 ['conv2d_1[0][0]']
rmalization)
activation_1 (Activation) (None, 56, 56, 128) 0 ['batch_normalization_2[0][0]']
conv2d_2 (Conv2D) (None, 56, 56, 32) 36864 ['activation_1[0][0]']
concatenate (Concatenate) (None, 56, 56, 96) 0 ['max_pooling2d[0][0]',
'conv2d_2[0][0]']
batch_normalization_3 (BatchNo (None, 56, 56, 96) 384 ['concatenate[0][0]']
rmalization)
activation_2 (Activation) (None, 56, 56, 96) 0 ['batch_normalization_3[0][0]']
conv2d_3 (Conv2D) (None, 56, 56, 128) 12288 ['activation_2[0][0]']
batch_normalization_4 (BatchNo (None, 56, 56, 128) 512 ['conv2d_3[0][0]']
rmalization)
activation_3 (Activation) (None, 56, 56, 128) 0 ['batch_normalization_4[0][0]']
conv2d_4 (Conv2D) (None, 56, 56, 32) 36864 ['activation_3[0][0]']
concatenate_1 (Concatenate) (None, 56, 56, 128) 0 ['concatenate[0][0]',
'conv2d_4[0][0]']
batch_normalization_5 (BatchNo (None, 56, 56, 128) 512 ['concatenate_1[0][0]']
rmalization)
activation_4 (Activation) (None, 56, 56, 128) 0 ['batch_normalization_5[0][0]']
conv2d_5 (Conv2D) (None, 56, 56, 128) 16384 ['activation_4[0][0]']
batch_normalization_6 (BatchNo (None, 56, 56, 128) 512 ['conv2d_5[0][0]']
rmalization)
activation_5 (Activation) (None, 56, 56, 128) 0 ['batch_normalization_6[0][0]']
conv2d_6 (Conv2D) (None, 56, 56, 32) 36864 ['activation_5[0][0]']
concatenate_2 (Concatenate) (None, 56, 56, 160) 0 ['concatenate_1[0][0]',
'conv2d_6[0][0]']
batch_normalization_7 (BatchNo (None, 56, 56, 160) 640 ['concatenate_2[0][0]']
rmalization)
activation_6 (Activation) (None, 56, 56, 160) 0 ['batch_normalization_7[0][0]']
conv2d_7 (Conv2D) (None, 56, 56, 128) 20480 ['activation_6[0][0]']
batch_normalization_8 (BatchNo (None, 56, 56, 128) 512 ['conv2d_7[0][0]']
rmalization)
activation_7 (Activation) (None, 56, 56, 128) 0 ['batch_normalization_8[0][0]']
conv2d_8 (Conv2D) (None, 56, 56, 32) 36864 ['activation_7[0][0]']
concatenate_3 (Concatenate) (None, 56, 56, 192) 0 ['concatenate_2[0][0]',
'conv2d_8[0][0]']
batch_normalization_9 (BatchNo (None, 56, 56, 192) 768 ['concatenate_3[0][0]']
rmalization)
activation_8 (Activation) (None, 56, 56, 192) 0 ['batch_normalization_9[0][0]']
conv2d_9 (Conv2D) (None, 56, 56, 128) 24576 ['activation_8[0][0]']
batch_normalization_10 (BatchN (None, 56, 56, 128) 512 ['conv2d_9[0][0]']
ormalization)
activation_9 (Activation) (None, 56, 56, 128) 0 ['batch_normalization_10[0][0]']
conv2d_10 (Conv2D) (None, 56, 56, 32) 36864 ['activation_9[0][0]']
concatenate_4 (Concatenate) (None, 56, 56, 224) 0 ['concatenate_3[0][0]',
'conv2d_10[0][0]']
batch_normalization_11 (BatchN (None, 56, 56, 224) 896 ['concatenate_4[0][0]']
ormalization)
activation_10 (Activation) (None, 56, 56, 224) 0 ['batch_normalization_11[0][0]']
conv2d_11 (Conv2D) (None, 56, 56, 128) 28672 ['activation_10[0][0]']
batch_normalization_12 (BatchN (None, 56, 56, 128) 512 ['conv2d_11[0][0]']
ormalization)
activation_11 (Activation) (None, 56, 56, 128) 0 ['batch_normalization_12[0][0]']
conv2d_12 (Conv2D) (None, 56, 56, 32) 36864 ['activation_11[0][0]']
concatenate_5 (Concatenate) (None, 56, 56, 256) 0 ['concatenate_4[0][0]',
'conv2d_12[0][0]']
batch_normalization_13 (BatchN (None, 56, 56, 256) 1024 ['concatenate_5[0][0]']
ormalization)
activation_12 (Activation) (None, 56, 56, 256) 0 ['batch_normalization_13[0][0]']
conv2d_13 (Conv2D) (None, 56, 56, 128) 32768 ['activation_12[0][0]']
average_pooling2d (AveragePool (None, 28, 28, 128) 0 ['conv2d_13[0][0]']
ing2D)
batch_normalization_14 (BatchN (None, 28, 28, 128) 512 ['average_pooling2d[0][0]']
ormalization)
activation_13 (Activation) (None, 28, 28, 128) 0 ['batch_normalization_14[0][0]']
conv2d_14 (Conv2D) (None, 28, 28, 128) 16384 ['activation_13[0][0]']
batch_normalization_15 (BatchN (None, 28, 28, 128) 512 ['conv2d_14[0][0]']
ormalization)
activation_14 (Activation) (None, 28, 28, 128) 0 ['batch_normalization_15[0][0]']
conv2d_15 (Conv2D) (None, 28, 28, 32) 36864 ['activation_14[0][0]']
concatenate_6 (Concatenate) (None, 28, 28, 160) 0 ['average_pooling2d[0][0]',
'conv2d_15[0][0]']
batch_normalization_16 (BatchN (None, 28, 28, 160) 640 ['concatenate_6[0][0]']
ormalization)
activation_15 (Activation) (None, 28, 28, 160) 0 ['batch_normalization_16[0][0]']
conv2d_16 (Conv2D) (None, 28, 28, 128) 20480 ['activation_15[0][0]']
batch_normalization_17 (BatchN (None, 28, 28, 128) 512 ['conv2d_16[0][0]']
ormalization)
activation_16 (Activation) (None, 28, 28, 128) 0 ['batch_normalization_17[0][0]']
conv2d_17 (Conv2D) (None, 28, 28, 32) 36864 ['activation_16[0][0]']
concatenate_7 (Concatenate) (None, 28, 28, 192) 0 ['concatenate_6[0][0]',
'conv2d_17[0][0]']
batch_normalization_18 (BatchN (None, 28, 28, 192) 768 ['concatenate_7[0][0]']
ormalization)
activation_17 (Activation) (None, 28, 28, 192) 0 ['batch_normalization_18[0][0]']
conv2d_18 (Conv2D) (None, 28, 28, 128) 24576 ['activation_17[0][0]']
batch_normalization_19 (BatchN (None, 28, 28, 128) 512 ['conv2d_18[0][0]']
ormalization)
activation_18 (Activation) (None, 28, 28, 128) 0 ['batch_normalization_19[0][0]']
conv2d_19 (Conv2D) (None, 28, 28, 32) 36864 ['activation_18[0][0]']
concatenate_8 (Concatenate) (None, 28, 28, 224) 0 ['concatenate_7[0][0]',
'conv2d_19[0][0]']
batch_normalization_20 (BatchN (None, 28, 28, 224) 896 ['concatenate_8[0][0]']
ormalization)
activation_19 (Activation) (None, 28, 28, 224) 0 ['batch_normalization_20[0][0]']
conv2d_20 (Conv2D) (None, 28, 28, 128) 28672 ['activation_19[0][0]']
batch_normalization_21 (BatchN (None, 28, 28, 128) 512 ['conv2d_20[0][0]']
ormalization)
activation_20 (Activation) (None, 28, 28, 128) 0 ['batch_normalization_21[0][0]']
conv2d_21 (Conv2D) (None, 28, 28, 32) 36864 ['activation_20[0][0]']
concatenate_9 (Concatenate) (None, 28, 28, 256) 0 ['concatenate_8[0][0]',
'conv2d_21[0][0]']
batch_normalization_22 (BatchN (None, 28, 28, 256) 1024 ['concatenate_9[0][0]']
ormalization)
activation_21 (Activation) (None, 28, 28, 256) 0 ['batch_normalization_22[0][0]']
conv2d_22 (Conv2D) (None, 28, 28, 128) 32768 ['activation_21[0][0]']
batch_normalization_23 (BatchN (None, 28, 28, 128) 512 ['conv2d_22[0][0]']
ormalization)
activation_22 (Activation) (None, 28, 28, 128) 0 ['batch_normalization_23[0][0]']
conv2d_23 (Conv2D) (None, 28, 28, 32) 36864 ['activation_22[0][0]']
concatenate_10 (Concatenate) (None, 28, 28, 288) 0 ['concatenate_9[0][0]',
'conv2d_23[0][0]']
batch_normalization_24 (BatchN (None, 28, 28, 288) 1152 ['concatenate_10[0][0]']
ormalization)
activation_23 (Activation) (None, 28, 28, 288) 0 ['batch_normalization_24[0][0]']
conv2d_24 (Conv2D) (None, 28, 28, 128) 36864 ['activation_23[0][0]']
batch_normalization_25 (BatchN (None, 28, 28, 128) 512 ['conv2d_24[0][0]']
ormalization)
activation_24 (Activation) (None, 28, 28, 128) 0 ['batch_normalization_25[0][0]']
conv2d_25 (Conv2D) (None, 28, 28, 32) 36864 ['activation_24[0][0]']
concatenate_11 (Concatenate) (None, 28, 28, 320) 0 ['concatenate_10[0][0]',
'conv2d_25[0][0]']
batch_normalization_26 (BatchN (None, 28, 28, 320) 1280 ['concatenate_11[0][0]']
ormalization)
activation_25 (Activation) (None, 28, 28, 320) 0 ['batch_normalization_26[0][0]']
conv2d_26 (Conv2D) (None, 28, 28, 128) 40960 ['activation_25[0][0]']
batch_normalization_27 (BatchN (None, 28, 28, 128) 512 ['conv2d_26[0][0]']
ormalization)
activation_26 (Activation) (None, 28, 28, 128) 0 ['batch_normalization_27[0][0]']
conv2d_27 (Conv2D) (None, 28, 28, 32) 36864 ['activation_26[0][0]']
concatenate_12 (Concatenate) (None, 28, 28, 352) 0 ['concatenate_11[0][0]',
'conv2d_27[0][0]']
batch_normalization_28 (BatchN (None, 28, 28, 352) 1408 ['concatenate_12[0][0]']
ormalization)
activation_27 (Activation) (None, 28, 28, 352) 0 ['batch_normalization_28[0][0]']
conv2d_28 (Conv2D) (None, 28, 28, 128) 45056 ['activation_27[0][0]']
batch_normalization_29 (BatchN (None, 28, 28, 128) 512 ['conv2d_28[0][0]']
ormalization)
activation_28 (Activation) (None, 28, 28, 128) 0 ['batch_normalization_29[0][0]']
conv2d_29 (Conv2D) (None, 28, 28, 32) 36864 ['activation_28[0][0]']
concatenate_13 (Concatenate) (None, 28, 28, 384) 0 ['concatenate_12[0][0]',
'conv2d_29[0][0]']
batch_normalization_30 (BatchN (None, 28, 28, 384) 1536 ['concatenate_13[0][0]']
ormalization)
activation_29 (Activation) (None, 28, 28, 384) 0 ['batch_normalization_30[0][0]']
conv2d_30 (Conv2D) (None, 28, 28, 128) 49152 ['activation_29[0][0]']
batch_normalization_31 (BatchN (None, 28, 28, 128) 512 ['conv2d_30[0][0]']
ormalization)
activation_30 (Activation) (None, 28, 28, 128) 0 ['batch_normalization_31[0][0]']
conv2d_31 (Conv2D) (None, 28, 28, 32) 36864 ['activation_30[0][0]']
concatenate_14 (Concatenate) (None, 28, 28, 416) 0 ['concatenate_13[0][0]',
'conv2d_31[0][0]']
batch_normalization_32 (BatchN (None, 28, 28, 416) 1664 ['concatenate_14[0][0]']
ormalization)
activation_31 (Activation) (None, 28, 28, 416) 0 ['batch_normalization_32[0][0]']
conv2d_32 (Conv2D) (None, 28, 28, 128) 53248 ['activation_31[0][0]']
batch_normalization_33 (BatchN (None, 28, 28, 128) 512 ['conv2d_32[0][0]']
ormalization)
activation_32 (Activation) (None, 28, 28, 128) 0 ['batch_normalization_33[0][0]']
conv2d_33 (Conv2D) (None, 28, 28, 32) 36864 ['activation_32[0][0]']
concatenate_15 (Concatenate) (None, 28, 28, 448) 0 ['concatenate_14[0][0]',
'conv2d_33[0][0]']
batch_normalization_34 (BatchN (None, 28, 28, 448) 1792 ['concatenate_15[0][0]']
ormalization)
activation_33 (Activation) (None, 28, 28, 448) 0 ['batch_normalization_34[0][0]']
conv2d_34 (Conv2D) (None, 28, 28, 128) 57344 ['activation_33[0][0]']
batch_normalization_35 (BatchN (None, 28, 28, 128) 512 ['conv2d_34[0][0]']
ormalization)
activation_34 (Activation) (None, 28, 28, 128) 0 ['batch_normalization_35[0][0]']
conv2d_35 (Conv2D) (None, 28, 28, 32) 36864 ['activation_34[0][0]']
concatenate_16 (Concatenate) (None, 28, 28, 480) 0 ['concatenate_15[0][0]',
'conv2d_35[0][0]']
batch_normalization_36 (BatchN (None, 28, 28, 480) 1920 ['concatenate_16[0][0]']
ormalization)
activation_35 (Activation) (None, 28, 28, 480) 0 ['batch_normalization_36[0][0]']
conv2d_36 (Conv2D) (None, 28, 28, 128) 61440 ['activation_35[0][0]']
batch_normalization_37 (BatchN (None, 28, 28, 128) 512 ['conv2d_36[0][0]']
ormalization)
activation_36 (Activation) (None, 28, 28, 128) 0 ['batch_normalization_37[0][0]']
conv2d_37 (Conv2D) (None, 28, 28, 32) 36864 ['activation_36[0][0]']
concatenate_17 (Concatenate) (None, 28, 28, 512) 0 ['concatenate_16[0][0]',
'conv2d_37[0][0]']
batch_normalization_38 (BatchN (None, 28, 28, 512) 2048 ['concatenate_17[0][0]']
ormalization)
activation_37 (Activation) (None, 28, 28, 512) 0 ['batch_normalization_38[0][0]']
conv2d_38 (Conv2D) (None, 28, 28, 256) 131072 ['activation_37[0][0]']
average_pooling2d_1 (AveragePo (None, 14, 14, 256) 0 ['conv2d_38[0][0]']
oling2D)
batch_normalization_39 (BatchN (None, 14, 14, 256) 1024 ['average_pooling2d_1[0][0]']
ormalization)
activation_38 (Activation) (None, 14, 14, 256) 0 ['batch_normalization_39[0][0]']
conv2d_39 (Conv2D) (None, 14, 14, 128) 32768 ['activation_38[0][0]']
batch_normalization_40 (BatchN (None, 14, 14, 128) 512 ['conv2d_39[0][0]']
ormalization)
activation_39 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_40[0][0]']
conv2d_40 (Conv2D) (None, 14, 14, 32) 36864 ['activation_39[0][0]']
concatenate_18 (Concatenate) (None, 14, 14, 288) 0 ['average_pooling2d_1[0][0]',
'conv2d_40[0][0]']
batch_normalization_41 (BatchN (None, 14, 14, 288) 1152 ['concatenate_18[0][0]']
ormalization)
activation_40 (Activation) (None, 14, 14, 288) 0 ['batch_normalization_41[0][0]']
conv2d_41 (Conv2D) (None, 14, 14, 128) 36864 ['activation_40[0][0]']
batch_normalization_42 (BatchN (None, 14, 14, 128) 512 ['conv2d_41[0][0]']
ormalization)
activation_41 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_42[0][0]']
conv2d_42 (Conv2D) (None, 14, 14, 32) 36864 ['activation_41[0][0]']
concatenate_19 (Concatenate) (None, 14, 14, 320) 0 ['concatenate_18[0][0]',
'conv2d_42[0][0]']
batch_normalization_43 (BatchN (None, 14, 14, 320) 1280 ['concatenate_19[0][0]']
ormalization)
activation_42 (Activation) (None, 14, 14, 320) 0 ['batch_normalization_43[0][0]']
conv2d_43 (Conv2D) (None, 14, 14, 128) 40960 ['activation_42[0][0]']
batch_normalization_44 (BatchN (None, 14, 14, 128) 512 ['conv2d_43[0][0]']
ormalization)
activation_43 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_44[0][0]']
conv2d_44 (Conv2D) (None, 14, 14, 32) 36864 ['activation_43[0][0]']
concatenate_20 (Concatenate) (None, 14, 14, 352) 0 ['concatenate_19[0][0]',
'conv2d_44[0][0]']
batch_normalization_45 (BatchN (None, 14, 14, 352) 1408 ['concatenate_20[0][0]']
ormalization)
activation_44 (Activation) (None, 14, 14, 352) 0 ['batch_normalization_45[0][0]']
conv2d_45 (Conv2D) (None, 14, 14, 128) 45056 ['activation_44[0][0]']
batch_normalization_46 (BatchN (None, 14, 14, 128) 512 ['conv2d_45[0][0]']
ormalization)
activation_45 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_46[0][0]']
conv2d_46 (Conv2D) (None, 14, 14, 32) 36864 ['activation_45[0][0]']
concatenate_21 (Concatenate) (None, 14, 14, 384) 0 ['concatenate_20[0][0]',
'conv2d_46[0][0]']
batch_normalization_47 (BatchN (None, 14, 14, 384) 1536 ['concatenate_21[0][0]']
ormalization)
activation_46 (Activation) (None, 14, 14, 384) 0 ['batch_normalization_47[0][0]']
conv2d_47 (Conv2D) (None, 14, 14, 128) 49152 ['activation_46[0][0]']
batch_normalization_48 (BatchN (None, 14, 14, 128) 512 ['conv2d_47[0][0]']
ormalization)
activation_47 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_48[0][0]']
conv2d_48 (Conv2D) (None, 14, 14, 32) 36864 ['activation_47[0][0]']
concatenate_22 (Concatenate) (None, 14, 14, 416) 0 ['concatenate_21[0][0]',
'conv2d_48[0][0]']
batch_normalization_49 (BatchN (None, 14, 14, 416) 1664 ['concatenate_22[0][0]']
ormalization)
activation_48 (Activation) (None, 14, 14, 416) 0 ['batch_normalization_49[0][0]']
conv2d_49 (Conv2D) (None, 14, 14, 128) 53248 ['activation_48[0][0]']
batch_normalization_50 (BatchN (None, 14, 14, 128) 512 ['conv2d_49[0][0]']
ormalization)
activation_49 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_50[0][0]']
conv2d_50 (Conv2D) (None, 14, 14, 32) 36864 ['activation_49[0][0]']
concatenate_23 (Concatenate) (None, 14, 14, 448) 0 ['concatenate_22[0][0]',
'conv2d_50[0][0]']
batch_normalization_51 (BatchN (None, 14, 14, 448) 1792 ['concatenate_23[0][0]']
ormalization)
activation_50 (Activation) (None, 14, 14, 448) 0 ['batch_normalization_51[0][0]']
conv2d_51 (Conv2D) (None, 14, 14, 128) 57344 ['activation_50[0][0]']
batch_normalization_52 (BatchN (None, 14, 14, 128) 512 ['conv2d_51[0][0]']
ormalization)
activation_51 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_52[0][0]']
conv2d_52 (Conv2D) (None, 14, 14, 32) 36864 ['activation_51[0][0]']
concatenate_24 (Concatenate) (None, 14, 14, 480) 0 ['concatenate_23[0][0]',
'conv2d_52[0][0]']
batch_normalization_53 (BatchN (None, 14, 14, 480) 1920 ['concatenate_24[0][0]']
ormalization)
activation_52 (Activation) (None, 14, 14, 480) 0 ['batch_normalization_53[0][0]']
conv2d_53 (Conv2D) (None, 14, 14, 128) 61440 ['activation_52[0][0]']
batch_normalization_54 (BatchN (None, 14, 14, 128) 512 ['conv2d_53[0][0]']
ormalization)
activation_53 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_54[0][0]']
conv2d_54 (Conv2D) (None, 14, 14, 32) 36864 ['activation_53[0][0]']
concatenate_25 (Concatenate) (None, 14, 14, 512) 0 ['concatenate_24[0][0]',
'conv2d_54[0][0]']
batch_normalization_55 (BatchN (None, 14, 14, 512) 2048 ['concatenate_25[0][0]']
ormalization)
activation_54 (Activation) (None, 14, 14, 512) 0 ['batch_normalization_55[0][0]']
conv2d_55 (Conv2D) (None, 14, 14, 128) 65536 ['activation_54[0][0]']
batch_normalization_56 (BatchN (None, 14, 14, 128) 512 ['conv2d_55[0][0]']
ormalization)
activation_55 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_56[0][0]']
conv2d_56 (Conv2D) (None, 14, 14, 32) 36864 ['activation_55[0][0]']
concatenate_26 (Concatenate) (None, 14, 14, 544) 0 ['concatenate_25[0][0]',
'conv2d_56[0][0]']
batch_normalization_57 (BatchN (None, 14, 14, 544) 2176 ['concatenate_26[0][0]']
ormalization)
activation_56 (Activation) (None, 14, 14, 544) 0 ['batch_normalization_57[0][0]']
conv2d_57 (Conv2D) (None, 14, 14, 128) 69632 ['activation_56[0][0]']
batch_normalization_58 (BatchN (None, 14, 14, 128) 512 ['conv2d_57[0][0]']
ormalization)
activation_57 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_58[0][0]']
conv2d_58 (Conv2D) (None, 14, 14, 32) 36864 ['activation_57[0][0]']
concatenate_27 (Concatenate) (None, 14, 14, 576) 0 ['concatenate_26[0][0]',
'conv2d_58[0][0]']
batch_normalization_59 (BatchN (None, 14, 14, 576) 2304 ['concatenate_27[0][0]']
ormalization)
activation_58 (Activation) (None, 14, 14, 576) 0 ['batch_normalization_59[0][0]']
conv2d_59 (Conv2D) (None, 14, 14, 128) 73728 ['activation_58[0][0]']
batch_normalization_60 (BatchN (None, 14, 14, 128) 512 ['conv2d_59[0][0]']
ormalization)
activation_59 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_60[0][0]']
conv2d_60 (Conv2D) (None, 14, 14, 32) 36864 ['activation_59[0][0]']
concatenate_28 (Concatenate) (None, 14, 14, 608) 0 ['concatenate_27[0][0]',
'conv2d_60[0][0]']
batch_normalization_61 (BatchN (None, 14, 14, 608) 2432 ['concatenate_28[0][0]']
ormalization)
activation_60 (Activation) (None, 14, 14, 608) 0 ['batch_normalization_61[0][0]']
conv2d_61 (Conv2D) (None, 14, 14, 128) 77824 ['activation_60[0][0]']
batch_normalization_62 (BatchN (None, 14, 14, 128) 512 ['conv2d_61[0][0]']
ormalization)
activation_61 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_62[0][0]']
conv2d_62 (Conv2D) (None, 14, 14, 32) 36864 ['activation_61[0][0]']
concatenate_29 (Concatenate) (None, 14, 14, 640) 0 ['concatenate_28[0][0]',
'conv2d_62[0][0]']
batch_normalization_63 (BatchN (None, 14, 14, 640) 2560 ['concatenate_29[0][0]']
ormalization)
activation_62 (Activation) (None, 14, 14, 640) 0 ['batch_normalization_63[0][0]']
conv2d_63 (Conv2D) (None, 14, 14, 128) 81920 ['activation_62[0][0]']
batch_normalization_64 (BatchN (None, 14, 14, 128) 512 ['conv2d_63[0][0]']
ormalization)
activation_63 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_64[0][0]']
conv2d_64 (Conv2D) (None, 14, 14, 32) 36864 ['activation_63[0][0]']
concatenate_30 (Concatenate) (None, 14, 14, 672) 0 ['concatenate_29[0][0]',
'conv2d_64[0][0]']
batch_normalization_65 (BatchN (None, 14, 14, 672) 2688 ['concatenate_30[0][0]']
ormalization)
activation_64 (Activation) (None, 14, 14, 672) 0 ['batch_normalization_65[0][0]']
conv2d_65 (Conv2D) (None, 14, 14, 128) 86016 ['activation_64[0][0]']
batch_normalization_66 (BatchN (None, 14, 14, 128) 512 ['conv2d_65[0][0]']
ormalization)
activation_65 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_66[0][0]']
conv2d_66 (Conv2D) (None, 14, 14, 32) 36864 ['activation_65[0][0]']
concatenate_31 (Concatenate) (None, 14, 14, 704) 0 ['concatenate_30[0][0]',
'conv2d_66[0][0]']
batch_normalization_67 (BatchN (None, 14, 14, 704) 2816 ['concatenate_31[0][0]']
ormalization)
activation_66 (Activation) (None, 14, 14, 704) 0 ['batch_normalization_67[0][0]']
conv2d_67 (Conv2D) (None, 14, 14, 128) 90112 ['activation_66[0][0]']
batch_normalization_68 (BatchN (None, 14, 14, 128) 512 ['conv2d_67[0][0]']
ormalization)
activation_67 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_68[0][0]']
conv2d_68 (Conv2D) (None, 14, 14, 32) 36864 ['activation_67[0][0]']
concatenate_32 (Concatenate) (None, 14, 14, 736) 0 ['concatenate_31[0][0]',
'conv2d_68[0][0]']
batch_normalization_69 (BatchN (None, 14, 14, 736) 2944 ['concatenate_32[0][0]']
ormalization)
activation_68 (Activation) (None, 14, 14, 736) 0 ['batch_normalization_69[0][0]']
conv2d_69 (Conv2D) (None, 14, 14, 128) 94208 ['activation_68[0][0]']
batch_normalization_70 (BatchN (None, 14, 14, 128) 512 ['conv2d_69[0][0]']
ormalization)
activation_69 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_70[0][0]']
conv2d_70 (Conv2D) (None, 14, 14, 32) 36864 ['activation_69[0][0]']
concatenate_33 (Concatenate) (None, 14, 14, 768) 0 ['concatenate_32[0][0]',
'conv2d_70[0][0]']
batch_normalization_71 (BatchN (None, 14, 14, 768) 3072 ['concatenate_33[0][0]']
ormalization)
activation_70 (Activation) (None, 14, 14, 768) 0 ['batch_normalization_71[0][0]']
conv2d_71 (Conv2D) (None, 14, 14, 128) 98304 ['activation_70[0][0]']
batch_normalization_72 (BatchN (None, 14, 14, 128) 512 ['conv2d_71[0][0]']
ormalization)
activation_71 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_72[0][0]']
conv2d_72 (Conv2D) (None, 14, 14, 32) 36864 ['activation_71[0][0]']
concatenate_34 (Concatenate) (None, 14, 14, 800) 0 ['concatenate_33[0][0]',
'conv2d_72[0][0]']
batch_normalization_73 (BatchN (None, 14, 14, 800) 3200 ['concatenate_34[0][0]']
ormalization)
activation_72 (Activation) (None, 14, 14, 800) 0 ['batch_normalization_73[0][0]']
conv2d_73 (Conv2D) (None, 14, 14, 128) 102400 ['activation_72[0][0]']
batch_normalization_74 (BatchN (None, 14, 14, 128) 512 ['conv2d_73[0][0]']
ormalization)
activation_73 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_74[0][0]']
conv2d_74 (Conv2D) (None, 14, 14, 32) 36864 ['activation_73[0][0]']
concatenate_35 (Concatenate) (None, 14, 14, 832) 0 ['concatenate_34[0][0]',
'conv2d_74[0][0]']
batch_normalization_75 (BatchN (None, 14, 14, 832) 3328 ['concatenate_35[0][0]']
ormalization)
activation_74 (Activation) (None, 14, 14, 832) 0 ['batch_normalization_75[0][0]']
conv2d_75 (Conv2D) (None, 14, 14, 128) 106496 ['activation_74[0][0]']
batch_normalization_76 (BatchN (None, 14, 14, 128) 512 ['conv2d_75[0][0]']
ormalization)
activation_75 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_76[0][0]']
conv2d_76 (Conv2D) (None, 14, 14, 32) 36864 ['activation_75[0][0]']
concatenate_36 (Concatenate) (None, 14, 14, 864) 0 ['concatenate_35[0][0]',
'conv2d_76[0][0]']
batch_normalization_77 (BatchN (None, 14, 14, 864) 3456 ['concatenate_36[0][0]']
ormalization)
activation_76 (Activation) (None, 14, 14, 864) 0 ['batch_normalization_77[0][0]']
conv2d_77 (Conv2D) (None, 14, 14, 128) 110592 ['activation_76[0][0]']
batch_normalization_78 (BatchN (None, 14, 14, 128) 512 ['conv2d_77[0][0]']
ormalization)
activation_77 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_78[0][0]']
conv2d_78 (Conv2D) (None, 14, 14, 32) 36864 ['activation_77[0][0]']
concatenate_37 (Concatenate) (None, 14, 14, 896) 0 ['concatenate_36[0][0]',
'conv2d_78[0][0]']
batch_normalization_79 (BatchN (None, 14, 14, 896) 3584 ['concatenate_37[0][0]']
ormalization)
activation_78 (Activation) (None, 14, 14, 896) 0 ['batch_normalization_79[0][0]']
conv2d_79 (Conv2D) (None, 14, 14, 128) 114688 ['activation_78[0][0]']
batch_normalization_80 (BatchN (None, 14, 14, 128) 512 ['conv2d_79[0][0]']
ormalization)
activation_79 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_80[0][0]']
conv2d_80 (Conv2D) (None, 14, 14, 32) 36864 ['activation_79[0][0]']
concatenate_38 (Concatenate) (None, 14, 14, 928) 0 ['concatenate_37[0][0]',
'conv2d_80[0][0]']
batch_normalization_81 (BatchN (None, 14, 14, 928) 3712 ['concatenate_38[0][0]']
ormalization)
activation_80 (Activation) (None, 14, 14, 928) 0 ['batch_normalization_81[0][0]']
conv2d_81 (Conv2D) (None, 14, 14, 128) 118784 ['activation_80[0][0]']
batch_normalization_82 (BatchN (None, 14, 14, 128) 512 ['conv2d_81[0][0]']
ormalization)
activation_81 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_82[0][0]']
conv2d_82 (Conv2D) (None, 14, 14, 32) 36864 ['activation_81[0][0]']
concatenate_39 (Concatenate) (None, 14, 14, 960) 0 ['concatenate_38[0][0]',
'conv2d_82[0][0]']
batch_normalization_83 (BatchN (None, 14, 14, 960) 3840 ['concatenate_39[0][0]']
ormalization)
activation_82 (Activation) (None, 14, 14, 960) 0 ['batch_normalization_83[0][0]']
conv2d_83 (Conv2D) (None, 14, 14, 128) 122880 ['activation_82[0][0]']
batch_normalization_84 (BatchN (None, 14, 14, 128) 512 ['conv2d_83[0][0]']
ormalization)
activation_83 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_84[0][0]']
conv2d_84 (Conv2D) (None, 14, 14, 32) 36864 ['activation_83[0][0]']
concatenate_40 (Concatenate) (None, 14, 14, 992) 0 ['concatenate_39[0][0]',
'conv2d_84[0][0]']
batch_normalization_85 (BatchN (None, 14, 14, 992) 3968 ['concatenate_40[0][0]']
ormalization)
activation_84 (Activation) (None, 14, 14, 992) 0 ['batch_normalization_85[0][0]']
conv2d_85 (Conv2D) (None, 14, 14, 128) 126976 ['activation_84[0][0]']
batch_normalization_86 (BatchN (None, 14, 14, 128) 512 ['conv2d_85[0][0]']
ormalization)
activation_85 (Activation) (None, 14, 14, 128) 0 ['batch_normalization_86[0][0]']
conv2d_86 (Conv2D) (None, 14, 14, 32) 36864 ['activation_85[0][0]']
concatenate_41 (Concatenate) (None, 14, 14, 1024 0 ['concatenate_40[0][0]',
) 'conv2d_86[0][0]']
batch_normalization_87 (BatchN (None, 14, 14, 1024 4096 ['concatenate_41[0][0]']
ormalization) )
activation_86 (Activation) (None, 14, 14, 1024 0 ['batch_normalization_87[0][0]']
)
conv2d_87 (Conv2D) (None, 14, 14, 512) 524288 ['activation_86[0][0]']
average_pooling2d_2 (AveragePo (None, 7, 7, 512) 0 ['conv2d_87[0][0]']
oling2D)
batch_normalization_88 (BatchN (None, 7, 7, 512) 2048 ['average_pooling2d_2[0][0]']
ormalization)
activation_87 (Activation) (None, 7, 7, 512) 0 ['batch_normalization_88[0][0]']
conv2d_88 (Conv2D) (None, 7, 7, 128) 65536 ['activation_87[0][0]']
batch_normalization_89 (BatchN (None, 7, 7, 128) 512 ['conv2d_88[0][0]']
ormalization)
activation_88 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_89[0][0]']
conv2d_89 (Conv2D) (None, 7, 7, 32) 36864 ['activation_88[0][0]']
concatenate_42 (Concatenate) (None, 7, 7, 544) 0 ['average_pooling2d_2[0][0]',
'conv2d_89[0][0]']
batch_normalization_90 (BatchN (None, 7, 7, 544) 2176 ['concatenate_42[0][0]']
ormalization)
activation_89 (Activation) (None, 7, 7, 544) 0 ['batch_normalization_90[0][0]']
conv2d_90 (Conv2D) (None, 7, 7, 128) 69632 ['activation_89[0][0]']
batch_normalization_91 (BatchN (None, 7, 7, 128) 512 ['conv2d_90[0][0]']
ormalization)
activation_90 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_91[0][0]']
conv2d_91 (Conv2D) (None, 7, 7, 32) 36864 ['activation_90[0][0]']
concatenate_43 (Concatenate) (None, 7, 7, 576) 0 ['concatenate_42[0][0]',
'conv2d_91[0][0]']
batch_normalization_92 (BatchN (None, 7, 7, 576) 2304 ['concatenate_43[0][0]']
ormalization)
activation_91 (Activation) (None, 7, 7, 576) 0 ['batch_normalization_92[0][0]']
conv2d_92 (Conv2D) (None, 7, 7, 128) 73728 ['activation_91[0][0]']
batch_normalization_93 (BatchN (None, 7, 7, 128) 512 ['conv2d_92[0][0]']
ormalization)
activation_92 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_93[0][0]']
conv2d_93 (Conv2D) (None, 7, 7, 32) 36864 ['activation_92[0][0]']
concatenate_44 (Concatenate) (None, 7, 7, 608) 0 ['concatenate_43[0][0]',
'conv2d_93[0][0]']
batch_normalization_94 (BatchN (None, 7, 7, 608) 2432 ['concatenate_44[0][0]']
ormalization)
activation_93 (Activation) (None, 7, 7, 608) 0 ['batch_normalization_94[0][0]']
conv2d_94 (Conv2D) (None, 7, 7, 128) 77824 ['activation_93[0][0]']
batch_normalization_95 (BatchN (None, 7, 7, 128) 512 ['conv2d_94[0][0]']
ormalization)
activation_94 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_95[0][0]']
conv2d_95 (Conv2D) (None, 7, 7, 32) 36864 ['activation_94[0][0]']
concatenate_45 (Concatenate) (None, 7, 7, 640) 0 ['concatenate_44[0][0]',
'conv2d_95[0][0]']
batch_normalization_96 (BatchN (None, 7, 7, 640) 2560 ['concatenate_45[0][0]']
ormalization)
activation_95 (Activation) (None, 7, 7, 640) 0 ['batch_normalization_96[0][0]']
conv2d_96 (Conv2D) (None, 7, 7, 128) 81920 ['activation_95[0][0]']
batch_normalization_97 (BatchN (None, 7, 7, 128) 512 ['conv2d_96[0][0]']
ormalization)
activation_96 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_97[0][0]']
conv2d_97 (Conv2D) (None, 7, 7, 32) 36864 ['activation_96[0][0]']
concatenate_46 (Concatenate) (None, 7, 7, 672) 0 ['concatenate_45[0][0]',
'conv2d_97[0][0]']
batch_normalization_98 (BatchN (None, 7, 7, 672) 2688 ['concatenate_46[0][0]']
ormalization)
activation_97 (Activation) (None, 7, 7, 672) 0 ['batch_normalization_98[0][0]']
conv2d_98 (Conv2D) (None, 7, 7, 128) 86016 ['activation_97[0][0]']
batch_normalization_99 (BatchN (None, 7, 7, 128) 512 ['conv2d_98[0][0]']
ormalization)
activation_98 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_99[0][0]']
conv2d_99 (Conv2D) (None, 7, 7, 32) 36864 ['activation_98[0][0]']
concatenate_47 (Concatenate) (None, 7, 7, 704) 0 ['concatenate_46[0][0]',
'conv2d_99[0][0]']
batch_normalization_100 (Batch (None, 7, 7, 704) 2816 ['concatenate_47[0][0]']
Normalization)
activation_99 (Activation) (None, 7, 7, 704) 0 ['batch_normalization_100[0][0]']
conv2d_100 (Conv2D) (None, 7, 7, 128) 90112 ['activation_99[0][0]']
batch_normalization_101 (Batch (None, 7, 7, 128) 512 ['conv2d_100[0][0]']
Normalization)
activation_100 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_101[0][0]']
conv2d_101 (Conv2D) (None, 7, 7, 32) 36864 ['activation_100[0][0]']
concatenate_48 (Concatenate) (None, 7, 7, 736) 0 ['concatenate_47[0][0]',
'conv2d_101[0][0]']
batch_normalization_102 (Batch (None, 7, 7, 736) 2944 ['concatenate_48[0][0]']
Normalization)
activation_101 (Activation) (None, 7, 7, 736) 0 ['batch_normalization_102[0][0]']
conv2d_102 (Conv2D) (None, 7, 7, 128) 94208 ['activation_101[0][0]']
batch_normalization_103 (Batch (None, 7, 7, 128) 512 ['conv2d_102[0][0]']
Normalization)
activation_102 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_103[0][0]']
conv2d_103 (Conv2D) (None, 7, 7, 32) 36864 ['activation_102[0][0]']
concatenate_49 (Concatenate) (None, 7, 7, 768) 0 ['concatenate_48[0][0]',
'conv2d_103[0][0]']
batch_normalization_104 (Batch (None, 7, 7, 768) 3072 ['concatenate_49[0][0]']
Normalization)
activation_103 (Activation) (None, 7, 7, 768) 0 ['batch_normalization_104[0][0]']
conv2d_104 (Conv2D) (None, 7, 7, 128) 98304 ['activation_103[0][0]']
batch_normalization_105 (Batch (None, 7, 7, 128) 512 ['conv2d_104[0][0]']
Normalization)
activation_104 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_105[0][0]']
conv2d_105 (Conv2D) (None, 7, 7, 32) 36864 ['activation_104[0][0]']
concatenate_50 (Concatenate) (None, 7, 7, 800) 0 ['concatenate_49[0][0]',
'conv2d_105[0][0]']
batch_normalization_106 (Batch (None, 7, 7, 800) 3200 ['concatenate_50[0][0]']
Normalization)
activation_105 (Activation) (None, 7, 7, 800) 0 ['batch_normalization_106[0][0]']
conv2d_106 (Conv2D) (None, 7, 7, 128) 102400 ['activation_105[0][0]']
batch_normalization_107 (Batch (None, 7, 7, 128) 512 ['conv2d_106[0][0]']
Normalization)
activation_106 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_107[0][0]']
conv2d_107 (Conv2D) (None, 7, 7, 32) 36864 ['activation_106[0][0]']
concatenate_51 (Concatenate) (None, 7, 7, 832) 0 ['concatenate_50[0][0]',
'conv2d_107[0][0]']
batch_normalization_108 (Batch (None, 7, 7, 832) 3328 ['concatenate_51[0][0]']
Normalization)
activation_107 (Activation) (None, 7, 7, 832) 0 ['batch_normalization_108[0][0]']
conv2d_108 (Conv2D) (None, 7, 7, 128) 106496 ['activation_107[0][0]']
batch_normalization_109 (Batch (None, 7, 7, 128) 512 ['conv2d_108[0][0]']
Normalization)
activation_108 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_109[0][0]']
conv2d_109 (Conv2D) (None, 7, 7, 32) 36864 ['activation_108[0][0]']
concatenate_52 (Concatenate) (None, 7, 7, 864) 0 ['concatenate_51[0][0]',
'conv2d_109[0][0]']
batch_normalization_110 (Batch (None, 7, 7, 864) 3456 ['concatenate_52[0][0]']
Normalization)
activation_109 (Activation) (None, 7, 7, 864) 0 ['batch_normalization_110[0][0]']
conv2d_110 (Conv2D) (None, 7, 7, 128) 110592 ['activation_109[0][0]']
batch_normalization_111 (Batch (None, 7, 7, 128) 512 ['conv2d_110[0][0]']
Normalization)
activation_110 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_111[0][0]']
conv2d_111 (Conv2D) (None, 7, 7, 32) 36864 ['activation_110[0][0]']
concatenate_53 (Concatenate) (None, 7, 7, 896) 0 ['concatenate_52[0][0]',
'conv2d_111[0][0]']
batch_normalization_112 (Batch (None, 7, 7, 896) 3584 ['concatenate_53[0][0]']
Normalization)
activation_111 (Activation) (None, 7, 7, 896) 0 ['batch_normalization_112[0][0]']
conv2d_112 (Conv2D) (None, 7, 7, 128) 114688 ['activation_111[0][0]']
batch_normalization_113 (Batch (None, 7, 7, 128) 512 ['conv2d_112[0][0]']
Normalization)
activation_112 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_113[0][0]']
conv2d_113 (Conv2D) (None, 7, 7, 32) 36864 ['activation_112[0][0]']
concatenate_54 (Concatenate) (None, 7, 7, 928) 0 ['concatenate_53[0][0]',
'conv2d_113[0][0]']
batch_normalization_114 (Batch (None, 7, 7, 928) 3712 ['concatenate_54[0][0]']
Normalization)
activation_113 (Activation) (None, 7, 7, 928) 0 ['batch_normalization_114[0][0]']
conv2d_114 (Conv2D) (None, 7, 7, 128) 118784 ['activation_113[0][0]']
batch_normalization_115 (Batch (None, 7, 7, 128) 512 ['conv2d_114[0][0]']
Normalization)
activation_114 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_115[0][0]']
conv2d_115 (Conv2D) (None, 7, 7, 32) 36864 ['activation_114[0][0]']
concatenate_55 (Concatenate) (None, 7, 7, 960) 0 ['concatenate_54[0][0]',
'conv2d_115[0][0]']
batch_normalization_116 (Batch (None, 7, 7, 960) 3840 ['concatenate_55[0][0]']
Normalization)
activation_115 (Activation) (None, 7, 7, 960) 0 ['batch_normalization_116[0][0]']
conv2d_116 (Conv2D) (None, 7, 7, 128) 122880 ['activation_115[0][0]']
batch_normalization_117 (Batch (None, 7, 7, 128) 512 ['conv2d_116[0][0]']
Normalization)
activation_116 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_117[0][0]']
conv2d_117 (Conv2D) (None, 7, 7, 32) 36864 ['activation_116[0][0]']
concatenate_56 (Concatenate) (None, 7, 7, 992) 0 ['concatenate_55[0][0]',
'conv2d_117[0][0]']
batch_normalization_118 (Batch (None, 7, 7, 992) 3968 ['concatenate_56[0][0]']
Normalization)
activation_117 (Activation) (None, 7, 7, 992) 0 ['batch_normalization_118[0][0]']
conv2d_118 (Conv2D) (None, 7, 7, 128) 126976 ['activation_117[0][0]']
batch_normalization_119 (Batch (None, 7, 7, 128) 512 ['conv2d_118[0][0]']
Normalization)
activation_118 (Activation) (None, 7, 7, 128) 0 ['batch_normalization_119[0][0]']
conv2d_119 (Conv2D) (None, 7, 7, 32) 36864 ['activation_118[0][0]']
concatenate_57 (Concatenate) (None, 7, 7, 1024) 0 ['concatenate_56[0][0]',
'conv2d_119[0][0]']
global_average_pooling2d (Glob (None, 1024) 0 ['concatenate_57[0][0]']
alAveragePooling2D)
dense (Dense) (None, 16) 16400 ['global_average_pooling2d[0][0]'
]
activation_119 (Activation) (None, 16) 0 ['dense[0][0]']
dense_1 (Dense) (None, 1024) 17408 ['activation_119[0][0]']
activation_120 (Activation) (None, 1024) 0 ['dense_1[0][0]']
reshape (Reshape) (None, 1, 1, 1024) 0 ['activation_120[0][0]']
tf.math.multiply (TFOpLambda) (None, 7, 7, 1024) 0 ['concatenate_57[0][0]',
'reshape[0][0]']
batch_normalization_120 (Batch (None, 7, 7, 1024) 4096 ['tf.math.multiply[0][0]']
Normalization)
activation_121 (Activation) (None, 7, 7, 1024) 0 ['batch_normalization_120[0][0]']
global_average_pooling2d_1 (Gl (None, 1024) 0 ['activation_121[0][0]']
obalAveragePooling2D)
dense_2 (Dense) (None, 1000) 1025000 ['global_average_pooling2d_1[0][0
]']
==================================================================================================
Total params: 8,096,312
Trainable params: 8,012,664
Non-trainable params: 83,648
__________________________________________________________________________________________________
3.10.编译模型
#设置初始学习率
initial_learning_rate = 1e-4
opt = tf.keras.optimizers.Adam(learning_rate=initial_learning_rate)
model.compile(optimizer=opt,
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
3.11.训练模型
'''训练模型'''
epochs = 20
history = model.fit(
train_ds,
validation_data=val_ds,
epochs=epochs
)
训练记录如下：
Epoch 1/20
54/54 [==============================] - ETA: 0s - loss: 4.1244 - accuracy: 0.5560
Epoch 1: val_accuracy improved from -inf to 0.07818, saving model to best_model.h5
54/54 [==============================] - 25s 236ms/step - loss: 4.1244 - accuracy: 0.5560 - val_loss: 8.7794 - val_accuracy: 0.0782
Epoch 2/20
54/54 [==============================] - ETA: 0s - loss: 1.3264 - accuracy: 0.6972
Epoch 2: val_accuracy improved from 0.07818 to 0.63477, saving model to best_model.h5
54/54 [==============================] - 12s 214ms/step - loss: 1.3264 - accuracy: 0.6972 - val_loss: 4.7183 - val_accuracy: 0.6348
Epoch 3/20
54/54 [==============================] - ETA: 0s - loss: 0.6500 - accuracy: 0.7515
Epoch 3: val_accuracy did not improve from 0.63477
54/54 [==============================] - 11s 210ms/step - loss: 0.6500 - accuracy: 0.7515 - val_loss: 3.0509 - val_accuracy: 0.5828
Epoch 4/20
54/54 [==============================] - ETA: 0s - loss: 0.4991 - accuracy: 0.8028
Epoch 4: val_accuracy improved from 0.63477 to 0.65811, saving model to best_model.h5
54/54 [==============================] - 12s 217ms/step - loss: 0.4991 - accuracy: 0.8028 - val_loss: 1.5565 - val_accuracy: 0.6581
Epoch 5/20
54/54 [==============================] - ETA: 0s - loss: 0.3937 - accuracy: 0.8448
Epoch 5: val_accuracy did not improve from 0.65811
54/54 [==============================] - 11s 211ms/step - loss: 0.3937 - accuracy: 0.8448 - val_loss: 1.0164 - val_accuracy: 0.6540
Epoch 6/20
54/54 [==============================] - ETA: 0s - loss: 0.3300 - accuracy: 0.8716
Epoch 6: val_accuracy did not improve from 0.65811
54/54 [==============================] - 11s 210ms/step - loss: 0.3300 - accuracy: 0.8716 - val_loss: 0.8846 - val_accuracy: 0.5799
Epoch 7/20
54/54 [==============================] - ETA: 0s - loss: 0.2872 - accuracy: 0.8839
Epoch 7: val_accuracy improved from 0.65811 to 0.66219, saving model to best_model.h5
54/54 [==============================] - 12s 221ms/step - loss: 0.2872 - accuracy: 0.8839 - val_loss: 0.8020 - val_accuracy: 0.6622
Epoch 8/20
54/54 [==============================] - ETA: 0s - loss: 0.2339 - accuracy: 0.9090
Epoch 8: val_accuracy improved from 0.66219 to 0.81855, saving model to best_model.h5
54/54 [==============================] - 12s 220ms/step - loss: 0.2339 - accuracy: 0.9090 - val_loss: 0.4418 - val_accuracy: 0.8186
Epoch 9/20
54/54 [==============================] - ETA: 0s - loss: 0.2030 - accuracy: 0.9247
Epoch 9: val_accuracy improved from 0.81855 to 0.82555, saving model to best_model.h5
54/54 [==============================] - 12s 222ms/step - loss: 0.2030 - accuracy: 0.9247 - val_loss: 0.4440 - val_accuracy: 0.8256
Epoch 10/20
54/54 [==============================] - ETA: 0s - loss: 0.1891 - accuracy: 0.9259
Epoch 10: val_accuracy did not improve from 0.82555
54/54 [==============================] - 12s 215ms/step - loss: 0.1891 - accuracy: 0.9259 - val_loss: 1.0064 - val_accuracy: 0.7421
Epoch 11/20
54/54 [==============================] - ETA: 0s - loss: 0.1565 - accuracy: 0.9440
Epoch 11: val_accuracy did not improve from 0.82555
54/54 [==============================] - 11s 211ms/step - loss: 0.1565 - accuracy: 0.9440 - val_loss: 0.5438 - val_accuracy: 0.8121
Epoch 12/20
54/54 [==============================] - ETA: 0s - loss: 0.1580 - accuracy: 0.9422
Epoch 12: val_accuracy did not improve from 0.82555
54/54 [==============================] - 11s 210ms/step - loss: 0.1580 - accuracy: 0.9422 - val_loss: 1.1685 - val_accuracy: 0.7100
Epoch 13/20
54/54 [==============================] - ETA: 0s - loss: 0.1100 - accuracy: 0.9650
Epoch 13: val_accuracy improved from 0.82555 to 0.89148, saving model to best_model.h5
54/54 [==============================] - 12s 218ms/step - loss: 0.1100 - accuracy: 0.9650 - val_loss: 0.3890 - val_accuracy: 0.8915
Epoch 14/20
54/54 [==============================] - ETA: 0s - loss: 0.0560 - accuracy: 0.9848
Epoch 14: val_accuracy did not improve from 0.89148
54/54 [==============================] - 11s 213ms/step - loss: 0.0560 - accuracy: 0.9848 - val_loss: 0.4445 - val_accuracy: 0.8676
Epoch 15/20
54/54 [==============================] - ETA: 0s - loss: 0.0700 - accuracy: 0.9772
Epoch 15: val_accuracy did not improve from 0.89148
54/54 [==============================] - 11s 212ms/step - loss: 0.0700 - accuracy: 0.9772 - val_loss: 0.4124 - val_accuracy: 0.8839
Epoch 16/20
54/54 [==============================] - ETA: 0s - loss: 0.0994 - accuracy: 0.9638
Epoch 16: val_accuracy did not improve from 0.89148
54/54 [==============================] - 11s 211ms/step - loss: 0.0994 - accuracy: 0.9638 - val_loss: 0.5568 - val_accuracy: 0.8261
Epoch 17/20
54/54 [==============================] - ETA: 0s - loss: 0.0689 - accuracy: 0.9743
Epoch 17: val_accuracy did not improve from 0.89148
54/54 [==============================] - 11s 214ms/step - loss: 0.0689 - accuracy: 0.9743 - val_loss: 0.5721 - val_accuracy: 0.8436
Epoch 18/20
54/54 [==============================] - ETA: 0s - loss: 0.0355 - accuracy: 0.9924
Epoch 18: val_accuracy improved from 0.89148 to 0.91832, saving model to best_model.h5
54/54 [==============================] - 12s 219ms/step - loss: 0.0355 - accuracy: 0.9924 - val_loss: 0.3478 - val_accuracy: 0.9183
Epoch 19/20
54/54 [==============================] - ETA: 0s - loss: 0.0100 - accuracy: 0.9994
Epoch 19: val_accuracy improved from 0.91832 to 0.94516, saving model to best_model.h5
54/54 [==============================] - 12s 217ms/step - loss: 0.0100 - accuracy: 0.9994 - val_loss: 0.1933 - val_accuracy: 0.9452
Epoch 20/20
54/54 [==============================] - ETA: 0s - loss: 0.0208 - accuracy: 0.9947
Epoch 20: val_accuracy did not improve from 0.94516
54/54 [==============================] - 11s 211ms/step - loss: 0.0208 - accuracy: 0.9947 - val_loss: 0.6098 - val_accuracy: 0.8460
3.12.模型评估
'''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(len(loss))
plt.figure(figsize=(12, 4))
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()
3.13.图像预测
'''指定图片进行预测'''
# 采用加载的模型（new_model）来看预测结果
plt.figure(figsize=(10, 5)) # 图形的宽为10高为5
plt.suptitle("预测结果展示", fontsize=10)
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(class_names[np.argmax(predictions)], fontsize=10)
plt.axis("off")
plt.show()
4 知识点详解
4.1ResNeXt50详解
论文：Aggregated Residual Transformations for Deep Neural Networks.pdf
ResNeXt是由何凯明团队在2017年CVPR会议上提出来的新型图像分类网络。ResNeXt是ResNet的升级版，在ResNet的基础上，引入了cardinality的概念，类似于ResNet，ResNeXt也有ResNeXt-50，ResNeXt-101的版本。
这篇文章介绍了一种用于图像分类的简单而有效的网络架构，称为Aggregated Residual Transformations for Deep Neural Networks。该网络采用了VGG/ResNets的策略，通过重复层来增加深度和宽度，并利用分裂-变换-合并策略以易于扩展的方式进行转换。文章还提出了一个新的维度——“基数”，它是指转换集合的大小，可以在保持复杂性不变的情况下提高分类准确性。作者在ImageNet-1K数据集上进行了实证研究，证明了这种方法的有效性。
下图是ResNet（左）与ResNeXt（右）block的差异。在ResNet中，输入的具有256个通道的特征经过1×1卷积压缩4倍到64个通道，之后3×3的卷积核用于处理特征，经1×1卷积扩大通道数与原特征残差连接后输出。
ResNeXt也是相同的处理策略，但在ResNeXt中，输入的具有256个通道的特征被分为32个组，每组被压缩64倍到4个通道后进行处理。32个组相加后与原特征残差连接后输出。这里cardinatity指的是一个block中所具有的相同分支的数目。下图为等效模型。
下图为ResNet50和ResNeXt50(32x4d)的结构对比图。
分组卷积
ResNeXt中采用的分组卷机简单来说就是将特征图分为不同的组，再对每组特征图分别进行卷积，这个操作可以有效的降低计算量。
在分组卷积中，每个卷积核只处理部分通道，比如下图中，红色卷积核只处理红色的通道，绿色卷积核只处理绿色通道，黄色卷积核只处理黄色通道。此时每个卷积核有2个通道，每个卷积核生成一张特征图。
4.2 ResNeXt50对比ResNet50V2、DenseNet
4.2.1 网络结构
ResNet-50v2是ResNet系列中的一个经典模型，由50层卷积层、批量归一化、激活函数和池化层构成。它引入了一种全新的残差块结构，即bottleneck结构，使得网络参数量大幅度降低，同时精度也有所提升。
DenseNet是一种全新的网络结构，其特点是不同于传统的网络结构，DenseNet中每一层的输出不仅和前一层的输出有关，还和之前所有层的输出有关，这种密集连接的结构可以有效地缓解梯度消失和参数稀疏问题，提高了模型的泛化能力和精度。
AggResNet（ResNeXt50）则是基于ResNet结构改进而来的新型深度神经网络结构，其特点是采用了聚合残差结构和局部连接结构，同时引入了Random Erasing和Mixup等数据增强和正则化方法，可以进一步提高网络的精度和鲁棒性。
4.2.2 精度和计算量
在ImageNet数据集上，ResNet-50v2和DenseNet在Top-1和Top-5指标上都取得了优异的性能。与之相比，AggResNet在相同的深度下具有更高的精度，并且在参数量和计算量上都显著降低。同时，在较深的网络结构下，AggResNet的优势更加明显，可以达到更高的精度，而ResNet-50v2和DenseNet则难以继续提高精度。
4.2.3 适用范围
ResNet-50v2适用于各种图像分类任务，但在一些特定的视觉任务，如目标检测、语义分割等方面的表现可能不如其他模型。
DenseNet则在各种任务中都具有优异的性能，尤其在目标检测和语义分割等像素级别的任务中表现突出。
AggResNet则不仅适用于图像分类任务，同时也可以应用于目标检测、语义分割和行人重识别等视觉任务中，并且在这些任务中具有优异的性能。
4 总结
ResNet-50v2、DenseNet和AggResNet都是非常优秀的深度神经网络结构，它们在不同的任务和场景中都具有不同的优势和适用性。