🍨 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]
要求：
- 探索ResNet与DenseNet结合的可能性
- 根据模型特性构建新的模型框架
- 验证改进后模型的效果
一、 基础配置
- 语言环境：Python3.8
- 编译器选择：Pycharm
- 深度学习环境：
-
- torch==1.12.1+cu113
- torchvision==0.13.1+cu113
二、 前期准备
1.设置GPU
import pathlib
import torch
import torch.nn as nn
from torchvision import transforms, datasetsdevice = torch.device("cuda" if torch.cuda.is_available() else "cpu")print(device)2. 导入数据
本项目所采用的数据集未收录于公开数据中，故需要自己在文件目录中导入相应数据集合，并设置对应文件目录，以供后续学习过程中使用。
运行下述代码：
data_dir = './data/bird_photos'
data_dir = pathlib.Path(data_dir)data_paths = list(data_dir.glob('*'))
classeNames = [str(path).split("\\")[2] for path in data_paths]
print(classeNames)image_count = len(list(data_dir.glob('*/*')))
print("图片总数为：", image_count)得到如下输出：
['Bananaquit', 'Black Skimmer', 'Black Throated Bushtiti', 'Cockatoo']
图片总数为： 565接下来，我们通过transforms.Compose对整个数据集进行预处理：
train_transforms = transforms.Compose([transforms.Resize([224, 224]), # 将输入图片resize成统一尺寸# transforms.RandomHorizontalFlip(), # 随机水平翻转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。
])test_transform = 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("./data/bird_photos/", transform=train_transforms)
print(total_data.class_to_idx)得到如下输出：
{'Bananaquit': 0, 'Black Skimmer': 1, 'Black Throated Bushtiti': 2, 'Cockatoo': 3}3. 划分数据集
此处数据集需要做按比例划分的操作：
train_size = int(0.8 * len(total_data))
test_size = len(total_data) - train_size
train_dataset, test_dataset = torch.utils.data.random_split(total_data, [train_size, test_size])接下来，根据划分得到的训练集和验证集对数据集进行包装：
batch_size = 32
train_dl = torch.utils.data.DataLoader(train_dataset,batch_size=batch_size,shuffle=True,num_workers=0)
test_dl = torch.utils.data.DataLoader(test_dataset,batch_size=batch_size,shuffle=True,num_workers=0)并通过：
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输出测试数据集的数据分布情况：
Shape of X [N, C, H, W]: torch.Size([32, 3, 224, 224])
Shape of y: torch.Size([32]) torch.int644.搭建模型
DPN网络通过High Order RNN（HORNN）将ResNet和DenseNet进行了融合，实现了ResNet特征复用及DenseNet特征生成，在保持了二者复用特征+挖掘特征能力的同时，避免了像原始DenseNet那样臃肿的结构。
1.模型搭建
class Block(nn.Module):def __init__(self, in_channel, mid_channel, out_channel, dense_channel, stride, groups, is_shortcut=False):# in_channel,是输入通道数,mid_channel是中间经历的通道数，out_channels是经过一次板块之后的输出通道数。# dense_channels设置这个参数的原因就是一边进行着resnet方式的卷积运算，另一边也同时进行着dense的卷积计算，之后特征图融合形成新的特征图super().__init__()self.is_shortcut = is_shortcutself.out_channel = out_channelself.conv1 = nn.Sequential(nn.Conv2d(in_channel, mid_channel, kernel_size=1, bias=False),nn.BatchNorm2d(mid_channel),nn.ReLU())self.conv2 = nn.Sequential(nn.Conv2d(mid_channel, mid_channel, kernel_size=3, stride=stride, padding=1, groups=groups, bias=False),nn.BatchNorm2d(mid_channel),nn.ReLU())self.conv3 = nn.Sequential(nn.Conv2d(mid_channel, out_channel + dense_channel, kernel_size=1, bias=False),nn.BatchNorm2d(out_channel + dense_channel))if self.is_shortcut:self.shortcut = nn.Sequential(nn.Conv2d(in_channel, out_channel + dense_channel, kernel_size=3, padding=1, stride=stride, bias=False),nn.BatchNorm2d(out_channel + dense_channel))self.relu = nn.ReLU(inplace=True)def forward(self, x):a = xx = self.conv1(x)x = self.conv2(x)x = self.conv3(x)if self.is_shortcut:a = self.shortcut(a)d = self.out_channelx = torch.cat([a[:, :d, :, :] + x[:, :d, :, :], a[:, d:, :, :], x[:, d:, :, :]], dim=1)x = self.relu(x)return xclass DPN(nn.Module):def __init__(self, cfg):super(DPN, self).__init__()self.group = cfg['group']self.in_channel = cfg['in_channel']mid_channels = cfg['mid_channels']out_channels = cfg['out_channels']dense_channels = cfg['dense_channels']num = cfg['num']self.conv1 = nn.Sequential(nn.Conv2d(3, self.in_channel, 7, stride=2, padding=3, bias=False, padding_mode='zeros'),nn.BatchNorm2d(self.in_channel),nn.ReLU(),nn.MaxPool2d(kernel_size=3, stride=2, padding=0))self.conv2 = self._make_layers(mid_channels[0], out_channels[0], dense_channels[0], num[0], stride=1)self.conv3 = self._make_layers(mid_channels[1], out_channels[1], dense_channels[1], num[1], stride=2)self.conv4 = self._make_layers(mid_channels[2], out_channels[2], dense_channels[2], num[2], stride=2)self.conv5 = self._make_layers(mid_channels[3], out_channels[3], dense_channels[3], num[3], stride=2)self.pool = nn.AdaptiveAvgPool2d((1, 1))self.fc = nn.Linear(cfg['out_channels'][3] + (num[3] + 1) * cfg['dense_channels'][3], cfg['classes']) # fc层需要计算def _make_layers(self, mid_channel, out_channel, dense_channel, num, stride=2):layers = []layers.append(Block(self.in_channel, mid_channel, out_channel, dense_channel, stride=stride, groups=self.group,is_shortcut=True))# block_1里面is_shortcut=True就是resnet中的shortcut连接，将浅层的特征进行一次卷积之后与进行三次卷积的特征图相加# 后面几次相同的板块is_shortcut=False简单的理解就是一个多次重复的板块，第一次利用就可以满足浅层特征的利用，后面重复的不在需要self.in_channel = out_channel + dense_channel * 2# 由于里面包含dense这种一直在叠加的特征图计算，# 所以第一次是2倍的dense_channel,后面每次一都会多出1倍，所以有(i+2)*dense_channelfor i in range(1, num):layers.append(Block(self.in_channel, mid_channel, out_channel, dense_channel, stride=1, groups=self.group))self.in_channel = self.in_channel + dense_channel# self.in_channel = out_channel + (i+2)*dense_channelreturn nn.Sequential(*layers)def forward(self, x):x = self.conv1(x)x = self.conv2(x)x = self.conv3(x)x = self.conv4(x)x = self.conv5(x)x = self.pool(x)x = torch.flatten(x, start_dim=1)x = self.fc(x)return xdef DPN92(n_class=10):cfg = {'group': 32,'in_channel': 64,'mid_channels': (96, 192, 384, 768),'out_channels': (256, 512, 1024, 2048),'dense_channels': (16, 32, 24, 128),'num': (3, 4, 20, 3),'classes': (n_class)}return DPN(cfg)def DPN98(n_class=10):cfg = {'group': 40,'in_channel': 96,'mid_channels': (160, 320, 640, 1280),'out_channels': (256, 512, 1024, 2048),'dense_channels': (16, 32, 32, 128),'num': (3, 6, 20, 3),'classes': (n_class)}return DPN(cfg)2.查看模型信息
x = torch.randn(2, 3, 224, 224)
model = DPN98(4)
model.to(device)
import torchsummary as summary
summary.summary(model, (3, 224, 224))得到如下输出：
----------------------------------------------------------------Layer (type) Output Shape Param #
================================================================Conv2d-1 [-1, 96, 112, 112] 14,112BatchNorm2d-2 [-1, 96, 112, 112] 192ReLU-3 [-1, 96, 112, 112] 0MaxPool2d-4 [-1, 96, 55, 55] 0Conv2d-5 [-1, 160, 55, 55] 15,360BatchNorm2d-6 [-1, 160, 55, 55] 320ReLU-7 [-1, 160, 55, 55] 0Conv2d-8 [-1, 160, 55, 55] 5,760BatchNorm2d-9 [-1, 160, 55, 55] 320ReLU-10 [-1, 160, 55, 55] 0Conv2d-11 [-1, 272, 55, 55] 43,520BatchNorm2d-12 [-1, 272, 55, 55] 544Conv2d-13 [-1, 272, 55, 55] 235,008BatchNorm2d-14 [-1, 272, 55, 55] 544ReLU-15 [-1, 288, 55, 55] 0Block-16 [-1, 288, 55, 55] 0Conv2d-17 [-1, 160, 55, 55] 46,080BatchNorm2d-18 [-1, 160, 55, 55] 320ReLU-19 [-1, 160, 55, 55] 0Conv2d-20 [-1, 160, 55, 55] 5,760BatchNorm2d-21 [-1, 160, 55, 55] 320ReLU-22 [-1, 160, 55, 55] 0Conv2d-23 [-1, 272, 55, 55] 43,520BatchNorm2d-24 [-1, 272, 55, 55] 544ReLU-25 [-1, 304, 55, 55] 0Block-26 [-1, 304, 55, 55] 0Conv2d-27 [-1, 160, 55, 55] 48,640BatchNorm2d-28 [-1, 160, 55, 55] 320ReLU-29 [-1, 160, 55, 55] 0Conv2d-30 [-1, 160, 55, 55] 5,760BatchNorm2d-31 [-1, 160, 55, 55] 320ReLU-32 [-1, 160, 55, 55] 0Conv2d-33 [-1, 272, 55, 55] 43,520BatchNorm2d-34 [-1, 272, 55, 55] 544ReLU-35 [-1, 320, 55, 55] 0Block-36 [-1, 320, 55, 55] 0Conv2d-37 [-1, 320, 55, 55] 102,400BatchNorm2d-38 [-1, 320, 55, 55] 640ReLU-39 [-1, 320, 55, 55] 0Conv2d-40 [-1, 320, 28, 28] 23,040BatchNorm2d-41 [-1, 320, 28, 28] 640ReLU-42 [-1, 320, 28, 28] 0Conv2d-43 [-1, 544, 28, 28] 174,080BatchNorm2d-44 [-1, 544, 28, 28] 1,088Conv2d-45 [-1, 544, 28, 28] 1,566,720BatchNorm2d-46 [-1, 544, 28, 28] 1,088ReLU-47 [-1, 576, 28, 28] 0Block-48 [-1, 576, 28, 28] 0Conv2d-49 [-1, 320, 28, 28] 184,320BatchNorm2d-50 [-1, 320, 28, 28] 640ReLU-51 [-1, 320, 28, 28] 0Conv2d-52 [-1, 320, 28, 28] 23,040BatchNorm2d-53 [-1, 320, 28, 28] 640ReLU-54 [-1, 320, 28, 28] 0Conv2d-55 [-1, 544, 28, 28] 174,080BatchNorm2d-56 [-1, 544, 28, 28] 1,088ReLU-57 [-1, 608, 28, 28] 0Block-58 [-1, 608, 28, 28] 0Conv2d-59 [-1, 320, 28, 28] 194,560BatchNorm2d-60 [-1, 320, 28, 28] 640ReLU-61 [-1, 320, 28, 28] 0Conv2d-62 [-1, 320, 28, 28] 23,040BatchNorm2d-63 [-1, 320, 28, 28] 640ReLU-64 [-1, 320, 28, 28] 0Conv2d-65 [-1, 544, 28, 28] 174,080BatchNorm2d-66 [-1, 544, 28, 28] 1,088ReLU-67 [-1, 640, 28, 28] 0Block-68 [-1, 640, 28, 28] 0Conv2d-69 [-1, 320, 28, 28] 204,800BatchNorm2d-70 [-1, 320, 28, 28] 640ReLU-71 [-1, 320, 28, 28] 0Conv2d-72 [-1, 320, 28, 28] 23,040BatchNorm2d-73 [-1, 320, 28, 28] 640ReLU-74 [-1, 320, 28, 28] 0Conv2d-75 [-1, 544, 28, 28] 174,080BatchNorm2d-76 [-1, 544, 28, 28] 1,088ReLU-77 [-1, 672, 28, 28] 0Block-78 [-1, 672, 28, 28] 0Conv2d-79 [-1, 320, 28, 28] 215,040BatchNorm2d-80 [-1, 320, 28, 28] 640ReLU-81 [-1, 320, 28, 28] 0Conv2d-82 [-1, 320, 28, 28] 23,040BatchNorm2d-83 [-1, 320, 28, 28] 640ReLU-84 [-1, 320, 28, 28] 0Conv2d-85 [-1, 544, 28, 28] 174,080BatchNorm2d-86 [-1, 544, 28, 28] 1,088ReLU-87 [-1, 704, 28, 28] 0Block-88 [-1, 704, 28, 28] 0Conv2d-89 [-1, 320, 28, 28] 225,280BatchNorm2d-90 [-1, 320, 28, 28] 640ReLU-91 [-1, 320, 28, 28] 0Conv2d-92 [-1, 320, 28, 28] 23,040BatchNorm2d-93 [-1, 320, 28, 28] 640ReLU-94 [-1, 320, 28, 28] 0Conv2d-95 [-1, 544, 28, 28] 174,080BatchNorm2d-96 [-1, 544, 28, 28] 1,088ReLU-97 [-1, 736, 28, 28] 0Block-98 [-1, 736, 28, 28] 0Conv2d-99 [-1, 640, 28, 28] 471,040BatchNorm2d-100 [-1, 640, 28, 28] 1,280ReLU-101 [-1, 640, 28, 28] 0Conv2d-102 [-1, 640, 14, 14] 92,160BatchNorm2d-103 [-1, 640, 14, 14] 1,280ReLU-104 [-1, 640, 14, 14] 0Conv2d-105 [-1, 1056, 14, 14] 675,840BatchNorm2d-106 [-1, 1056, 14, 14] 2,112Conv2d-107 [-1, 1056, 14, 14] 6,994,944BatchNorm2d-108 [-1, 1056, 14, 14] 2,112ReLU-109 [-1, 1088, 14, 14] 0Block-110 [-1, 1088, 14, 14] 0Conv2d-111 [-1, 640, 14, 14] 696,320BatchNorm2d-112 [-1, 640, 14, 14] 1,280ReLU-113 [-1, 640, 14, 14] 0Conv2d-114 [-1, 640, 14, 14] 92,160BatchNorm2d-115 [-1, 640, 14, 14] 1,280ReLU-116 [-1, 640, 14, 14] 0Conv2d-117 [-1, 1056, 14, 14] 675,840BatchNorm2d-118 [-1, 1056, 14, 14] 2,112ReLU-119 [-1, 1120, 14, 14] 0Block-120 [-1, 1120, 14, 14] 0Conv2d-121 [-1, 640, 14, 14] 716,800BatchNorm2d-122 [-1, 640, 14, 14] 1,280ReLU-123 [-1, 640, 14, 14] 0Conv2d-124 [-1, 640, 14, 14] 92,160BatchNorm2d-125 [-1, 640, 14, 14] 1,280ReLU-126 [-1, 640, 14, 14] 0Conv2d-127 [-1, 1056, 14, 14] 675,840BatchNorm2d-128 [-1, 1056, 14, 14] 2,112ReLU-129 [-1, 1152, 14, 14] 0Block-130 [-1, 1152, 14, 14] 0Conv2d-131 [-1, 640, 14, 14] 737,280BatchNorm2d-132 [-1, 640, 14, 14] 1,280ReLU-133 [-1, 640, 14, 14] 0Conv2d-134 [-1, 640, 14, 14] 92,160BatchNorm2d-135 [-1, 640, 14, 14] 1,280ReLU-136 [-1, 640, 14, 14] 0Conv2d-137 [-1, 1056, 14, 14] 675,840BatchNorm2d-138 [-1, 1056, 14, 14] 2,112ReLU-139 [-1, 1184, 14, 14] 0Block-140 [-1, 1184, 14, 14] 0Conv2d-141 [-1, 640, 14, 14] 757,760BatchNorm2d-142 [-1, 640, 14, 14] 1,280ReLU-143 [-1, 640, 14, 14] 0Conv2d-144 [-1, 640, 14, 14] 92,160BatchNorm2d-145 [-1, 640, 14, 14] 1,280ReLU-146 [-1, 640, 14, 14] 0Conv2d-147 [-1, 1056, 14, 14] 675,840BatchNorm2d-148 [-1, 1056, 14, 14] 2,112ReLU-149 [-1, 1216, 14, 14] 0Block-150 [-1, 1216, 14, 14] 0Conv2d-151 [-1, 640, 14, 14] 778,240BatchNorm2d-152 [-1, 640, 14, 14] 1,280ReLU-153 [-1, 640, 14, 14] 0Conv2d-154 [-1, 640, 14, 14] 92,160BatchNorm2d-155 [-1, 640, 14, 14] 1,280ReLU-156 [-1, 640, 14, 14] 0Conv2d-157 [-1, 1056, 14, 14] 675,840BatchNorm2d-158 [-1, 1056, 14, 14] 2,112ReLU-159 [-1, 1248, 14, 14] 0Block-160 [-1, 1248, 14, 14] 0Conv2d-161 [-1, 640, 14, 14] 798,720BatchNorm2d-162 [-1, 640, 14, 14] 1,280ReLU-163 [-1, 640, 14, 14] 0Conv2d-164 [-1, 640, 14, 14] 92,160BatchNorm2d-165 [-1, 640, 14, 14] 1,280ReLU-166 [-1, 640, 14, 14] 0Conv2d-167 [-1, 1056, 14, 14] 675,840BatchNorm2d-168 [-1, 1056, 14, 14] 2,112ReLU-169 [-1, 1280, 14, 14] 0Block-170 [-1, 1280, 14, 14] 0Conv2d-171 [-1, 640, 14, 14] 819,200BatchNorm2d-172 [-1, 640, 14, 14] 1,280ReLU-173 [-1, 640, 14, 14] 0Conv2d-174 [-1, 640, 14, 14] 92,160BatchNorm2d-175 [-1, 640, 14, 14] 1,280ReLU-176 [-1, 640, 14, 14] 0Conv2d-177 [-1, 1056, 14, 14] 675,840BatchNorm2d-178 [-1, 1056, 14, 14] 2,112ReLU-179 [-1, 1312, 14, 14] 0Block-180 [-1, 1312, 14, 14] 0Conv2d-181 [-1, 640, 14, 14] 839,680BatchNorm2d-182 [-1, 640, 14, 14] 1,280ReLU-183 [-1, 640, 14, 14] 0Conv2d-184 [-1, 640, 14, 14] 92,160BatchNorm2d-185 [-1, 640, 14, 14] 1,280ReLU-186 [-1, 640, 14, 14] 0Conv2d-187 [-1, 1056, 14, 14] 675,840BatchNorm2d-188 [-1, 1056, 14, 14] 2,112ReLU-189 [-1, 1344, 14, 14] 0Block-190 [-1, 1344, 14, 14] 0Conv2d-191 [-1, 640, 14, 14] 860,160BatchNorm2d-192 [-1, 640, 14, 14] 1,280ReLU-193 [-1, 640, 14, 14] 0Conv2d-194 [-1, 640, 14, 14] 92,160BatchNorm2d-195 [-1, 640, 14, 14] 1,280ReLU-196 [-1, 640, 14, 14] 0Conv2d-197 [-1, 1056, 14, 14] 675,840BatchNorm2d-198 [-1, 1056, 14, 14] 2,112ReLU-199 [-1, 1376, 14, 14] 0Block-200 [-1, 1376, 14, 14] 0Conv2d-201 [-1, 640, 14, 14] 880,640BatchNorm2d-202 [-1, 640, 14, 14] 1,280ReLU-203 [-1, 640, 14, 14] 0Conv2d-204 [-1, 640, 14, 14] 92,160BatchNorm2d-205 [-1, 640, 14, 14] 1,280ReLU-206 [-1, 640, 14, 14] 0Conv2d-207 [-1, 1056, 14, 14] 675,840BatchNorm2d-208 [-1, 1056, 14, 14] 2,112ReLU-209 [-1, 1408, 14, 14] 0Block-210 [-1, 1408, 14, 14] 0Conv2d-211 [-1, 640, 14, 14] 901,120BatchNorm2d-212 [-1, 640, 14, 14] 1,280ReLU-213 [-1, 640, 14, 14] 0Conv2d-214 [-1, 640, 14, 14] 92,160BatchNorm2d-215 [-1, 640, 14, 14] 1,280ReLU-216 [-1, 640, 14, 14] 0Conv2d-217 [-1, 1056, 14, 14] 675,840BatchNorm2d-218 [-1, 1056, 14, 14] 2,112ReLU-219 [-1, 1440, 14, 14] 0Block-220 [-1, 1440, 14, 14] 0Conv2d-221 [-1, 640, 14, 14] 921,600BatchNorm2d-222 [-1, 640, 14, 14] 1,280ReLU-223 [-1, 640, 14, 14] 0Conv2d-224 [-1, 640, 14, 14] 92,160BatchNorm2d-225 [-1, 640, 14, 14] 1,280ReLU-226 [-1, 640, 14, 14] 0Conv2d-227 [-1, 1056, 14, 14] 675,840BatchNorm2d-228 [-1, 1056, 14, 14] 2,112ReLU-229 [-1, 1472, 14, 14] 0Block-230 [-1, 1472, 14, 14] 0Conv2d-231 [-1, 640, 14, 14] 942,080BatchNorm2d-232 [-1, 640, 14, 14] 1,280ReLU-233 [-1, 640, 14, 14] 0Conv2d-234 [-1, 640, 14, 14] 92,160BatchNorm2d-235 [-1, 640, 14, 14] 1,280ReLU-236 [-1, 640, 14, 14] 0Conv2d-237 [-1, 1056, 14, 14] 675,840BatchNorm2d-238 [-1, 1056, 14, 14] 2,112ReLU-239 [-1, 1504, 14, 14] 0Block-240 [-1, 1504, 14, 14] 0Conv2d-241 [-1, 640, 14, 14] 962,560BatchNorm2d-242 [-1, 640, 14, 14] 1,280ReLU-243 [-1, 640, 14, 14] 0Conv2d-244 [-1, 640, 14, 14] 92,160BatchNorm2d-245 [-1, 640, 14, 14] 1,280ReLU-246 [-1, 640, 14, 14] 0Conv2d-247 [-1, 1056, 14, 14] 675,840BatchNorm2d-248 [-1, 1056, 14, 14] 2,112ReLU-249 [-1, 1536, 14, 14] 0Block-250 [-1, 1536, 14, 14] 0Conv2d-251 [-1, 640, 14, 14] 983,040BatchNorm2d-252 [-1, 640, 14, 14] 1,280ReLU-253 [-1, 640, 14, 14] 0Conv2d-254 [-1, 640, 14, 14] 92,160BatchNorm2d-255 [-1, 640, 14, 14] 1,280ReLU-256 [-1, 640, 14, 14] 0Conv2d-257 [-1, 1056, 14, 14] 675,840BatchNorm2d-258 [-1, 1056, 14, 14] 2,112ReLU-259 [-1, 1568, 14, 14] 0Block-260 [-1, 1568, 14, 14] 0Conv2d-261 [-1, 640, 14, 14] 1,003,520BatchNorm2d-262 [-1, 640, 14, 14] 1,280ReLU-263 [-1, 640, 14, 14] 0Conv2d-264 [-1, 640, 14, 14] 92,160BatchNorm2d-265 [-1, 640, 14, 14] 1,280ReLU-266 [-1, 640, 14, 14] 0Conv2d-267 [-1, 1056, 14, 14] 675,840BatchNorm2d-268 [-1, 1056, 14, 14] 2,112ReLU-269 [-1, 1600, 14, 14] 0Block-270 [-1, 1600, 14, 14] 0Conv2d-271 [-1, 640, 14, 14] 1,024,000BatchNorm2d-272 [-1, 640, 14, 14] 1,280ReLU-273 [-1, 640, 14, 14] 0Conv2d-274 [-1, 640, 14, 14] 92,160BatchNorm2d-275 [-1, 640, 14, 14] 1,280ReLU-276 [-1, 640, 14, 14] 0Conv2d-277 [-1, 1056, 14, 14] 675,840BatchNorm2d-278 [-1, 1056, 14, 14] 2,112ReLU-279 [-1, 1632, 14, 14] 0Block-280 [-1, 1632, 14, 14] 0Conv2d-281 [-1, 640, 14, 14] 1,044,480BatchNorm2d-282 [-1, 640, 14, 14] 1,280ReLU-283 [-1, 640, 14, 14] 0Conv2d-284 [-1, 640, 14, 14] 92,160BatchNorm2d-285 [-1, 640, 14, 14] 1,280ReLU-286 [-1, 640, 14, 14] 0Conv2d-287 [-1, 1056, 14, 14] 675,840BatchNorm2d-288 [-1, 1056, 14, 14] 2,112ReLU-289 [-1, 1664, 14, 14] 0Block-290 [-1, 1664, 14, 14] 0Conv2d-291 [-1, 640, 14, 14] 1,064,960BatchNorm2d-292 [-1, 640, 14, 14] 1,280ReLU-293 [-1, 640, 14, 14] 0Conv2d-294 [-1, 640, 14, 14] 92,160BatchNorm2d-295 [-1, 640, 14, 14] 1,280ReLU-296 [-1, 640, 14, 14] 0Conv2d-297 [-1, 1056, 14, 14] 675,840BatchNorm2d-298 [-1, 1056, 14, 14] 2,112ReLU-299 [-1, 1696, 14, 14] 0Block-300 [-1, 1696, 14, 14] 0Conv2d-301 [-1, 1280, 14, 14] 2,170,880BatchNorm2d-302 [-1, 1280, 14, 14] 2,560ReLU-303 [-1, 1280, 14, 14] 0Conv2d-304 [-1, 1280, 7, 7] 368,640BatchNorm2d-305 [-1, 1280, 7, 7] 2,560ReLU-306 [-1, 1280, 7, 7] 0Conv2d-307 [-1, 2176, 7, 7] 2,785,280BatchNorm2d-308 [-1, 2176, 7, 7] 4,352Conv2d-309 [-1, 2176, 7, 7] 33,214,464BatchNorm2d-310 [-1, 2176, 7, 7] 4,352ReLU-311 [-1, 2304, 7, 7] 0Block-312 [-1, 2304, 7, 7] 0Conv2d-313 [-1, 1280, 7, 7] 2,949,120BatchNorm2d-314 [-1, 1280, 7, 7] 2,560ReLU-315 [-1, 1280, 7, 7] 0Conv2d-316 [-1, 1280, 7, 7] 368,640BatchNorm2d-317 [-1, 1280, 7, 7] 2,560ReLU-318 [-1, 1280, 7, 7] 0Conv2d-319 [-1, 2176, 7, 7] 2,785,280BatchNorm2d-320 [-1, 2176, 7, 7] 4,352ReLU-321 [-1, 2432, 7, 7] 0Block-322 [-1, 2432, 7, 7] 0Conv2d-323 [-1, 1280, 7, 7] 3,112,960BatchNorm2d-324 [-1, 1280, 7, 7] 2,560ReLU-325 [-1, 1280, 7, 7] 0Conv2d-326 [-1, 1280, 7, 7] 368,640BatchNorm2d-327 [-1, 1280, 7, 7] 2,560ReLU-328 [-1, 1280, 7, 7] 0Conv2d-329 [-1, 2176, 7, 7] 2,785,280BatchNorm2d-330 [-1, 2176, 7, 7] 4,352ReLU-331 [-1, 2560, 7, 7] 0Block-332 [-1, 2560, 7, 7] 0
AdaptiveAvgPool2d-333 [-1, 2560, 1, 1] 0Linear-334 [-1, 4] 10,244
================================================================
Total params: 95,008,356
Trainable params: 95,008,356
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 664.46
Params size (MB): 362.43
Estimated Total Size (MB): 1027.47
----------------------------------------------------------------三、 训练模型
1. 编写训练函数
def train(dataloader, model, loss_fn, optimizer):size = len(dataloader.dataset) # 训练集的大小num_batches = len(dataloader) # 批次数目, (size/batch_size，向上取整)train_loss, train_acc = 0, 0 # 初始化训练损失和正确率for X, y in dataloader: # 获取图片及其标签X, y = X.to(device), y.to(device)# 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() # grad属性归零loss.backward() # 反向传播optimizer.step() # 每一步自动更新# 记录acc与losstrain_acc += (pred.argmax(1) == y).type(torch.float).sum().item()train_loss += loss.item()train_acc /= sizetrain_loss /= num_batchesreturn train_acc, train_loss2. 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num_batches = len(dataloader) # 批次数目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():for imgs, target in dataloader:imgs, target = imgs.to(device), target.to(device)# 计算losstarget_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()test_acc /= sizetest_loss /= num_batchesreturn test_acc, test_loss3.正式训练
import copyoptimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
loss_fn = nn.CrossEntropyLoss() # 创建损失函数epochs = 10train_loss = []
train_acc = []
test_loss = []
test_acc = []best_acc = 0 # 设置一个最佳准确率，作为最佳模型的判别指标for epoch in range(epochs):# 更新学习率（使用自定义学习率时使用）# adjust_learning_rate(optimizer, epoch, learn_rate)model.train()epoch_train_acc, epoch_train_loss = train(train_dl, model, loss_fn, optimizer)# 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)# 保存最佳模型到 best_modelif epoch_test_acc > best_acc:best_acc = epoch_test_accbest_model = copy.deepcopy(model)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 = optimizer.state_dict()['param_groups'][0]['lr']template = ('Epoch:{:2d}, Train_acc:{:.1f}%, Train_loss:{:.3f}, Test_acc:{:.1f}%, Test_loss:{:.3f}, Lr:{:.2E}')print(template.format(epoch + 1, 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')
得到如下输出：
Epoch: 1, Train_acc:35.0%, Train_loss:1.512, Test_acc:15.0%, Test_loss:2.101, Lr:1.00E-04
Epoch: 2, Train_acc:55.1%, Train_loss:1.088, Test_acc:15.9%, Test_loss:5.737, Lr:1.00E-04
Epoch: 3, Train_acc:71.0%, Train_loss:0.773, Test_acc:39.8%, Test_loss:2.180, Lr:1.00E-04
Epoch: 4, Train_acc:76.3%, Train_loss:0.616, Test_acc:62.8%, Test_loss:1.222, Lr:1.00E-04
Epoch: 5, Train_acc:79.4%, Train_loss:0.565, Test_acc:61.1%, Test_loss:2.034, Lr:1.00E-04
Epoch: 6, Train_acc:79.6%, Train_loss:0.492, Test_acc:61.9%, Test_loss:1.497, Lr:1.00E-04
Epoch: 7, Train_acc:83.2%, Train_loss:0.480, Test_acc:69.0%, Test_loss:1.305, Lr:1.00E-04
Epoch: 8, Train_acc:84.1%, Train_loss:0.403, Test_acc:56.6%, Test_loss:2.690, Lr:1.00E-04
Epoch: 9, Train_acc:90.0%, Train_loss:0.304, Test_acc:71.7%, Test_loss:1.104, Lr:1.00E-04
Epoch:10, Train_acc:93.8%, Train_loss:0.190, Test_acc:55.8%, Test_loss:2.481, Lr:1.00E-04
Done
预测结果是：CockatooProcess finished with exit code 0
四、 结果可视化
1. Loss&Accuracy
import matplotlib.pyplot as plt
# 隐藏警告
import warningswarnings.filterwarnings("ignore") # 忽略警告信息
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()得到的可视化结果：
2. 指定图片进行预测
首先，先定义出一个用于预测的函数：
from PIL import Imageclasses = list(total_data.class_to_idx)from PIL import Imageclasses = list(total_data.class_to_idx)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}')接着调用函数对指定图片进行预测：
# 预测训练集中的某张照片
predict_one_image(image_path='./data/bird_photos/Cockatoo/011.jpg',model=model,transform=train_transforms,classes=classes)得到如下结果：
预测结果是：Cockatoo五、网络架构及参数
