- 🍨 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densenet121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 知识点详解
- 3.1 nn.Sequential和nn.Module区别与选择
- 3.1.1 nn.Sequential
- 3.1.2 nn.Module
- 3.1.3 对比
- 3.1.4 总结
- 3.2 python中OrderedDict的使用
- 总结
前言
关键字： pytorch实现DenseNet算法，nn.Sequential和nn.Module区别与选择，python中OrderedDict的使用
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 warningswarnings.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CPUdevice = 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\BreastCancer"
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
['.DS_Store', '0', '1']2.1.4 可视化数据
'''前期工作-可视化数据'''
subfolder = Path(data_dir) / "1"
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 ImageFolderNumber of datapoints: 13403Root location: D:\DeepLearning\data\BreastCancerStandardTransform
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]))
{'0': 0, '1': 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=<torch.utils.data.dataset.Subset object at 0x000001AB3AD06BE0>
test_dataset=<torch.utils.data.dataset.Subset object at 0x000001AB3AD06B20>
train_size=10722
test_size=2681
2.1.4 加载数据
'''前期工作-加载数据'''
batch_size = 32train_dl = torch.utils.data.DataLoader(train_dataset,batch_size=batch_size,shuffle=True,num_workers=4)
test_dl = torch.utils.data.DataLoader(test_dataset,batch_size=batch_size,shuffle=True,num_workers=4)
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densenet121模型
"""构建DenseNet网络"""
# 这里我们采用了Pytorch的框架来实现DenseNet，
# 首先实现DenseBlock中的内部结构，这里是BN+ReLU+1×1Conv+BN+ReLU+3×3Conv结构，最后也加入dropout层用于训练过程。
class _DenseLayer(nn.Sequential):"""Basic unit of DenseBlock (using bottleneck layer) """def __init__(self, num_input_features, growth_rate, bn_size, drop_rate):super(_DenseLayer, self).__init__()self.add_module('norm1', nn.BatchNorm2d(num_input_features)),self.add_module('relu1', nn.ReLU(inplace=True)),self.add_module('conv1', nn.Conv2d(num_input_features, bn_size * growth_rate,kernel_size=1, stride=1, bias=False)),self.add_module('norm2', nn.BatchNorm2d(bn_size * growth_rate)),self.add_module('relu2', nn.ReLU(inplace=True)),self.add_module('conv2', nn.Conv2d(bn_size * growth_rate, growth_rate,kernel_size=3, stride=1, padding=1, bias=False)),self.drop_rate = drop_ratedef forward(self, x):new_features = super(_DenseLayer, self).forward(x)if self.drop_rate > 0:new_features = F.dropout(new_features, p=self.drop_rate, training=self.training)return torch.cat([x, new_features], 1)# 实现DenseBlock模块，内部是密集连接方式（输入特征数线性增长）：
class _DenseBlock(nn.Sequential):"""DenseBlock """def __init__(self, num_layers, num_input_features, bn_size, growth_rate, drop_rate):super(_DenseBlock, self).__init__()for i in range(num_layers):layer = _DenseLayer(num_input_features + i * growth_rate, growth_rate, bn_size, drop_rate)self.add_module('denselayer%d' % (i + 1), layer)# 实现Transition层，它主要是一个卷积层和一个池化层：
class _Transition(nn.Sequential):def __init__(self, num_input_features, num_output_features):super(_Transition, self).__init__()self.add_module('norm', nn.BatchNorm2d(num_input_features))self.add_module('relu', nn.ReLU(inplace=True))self.add_module('conv', nn.Conv2d(num_input_features, num_output_features,kernel_size=1, stride=1, bias=False))self.add_module('pool', nn.AvgPool2d(kernel_size=2, stride=2))# 最后我们实现DenseNet网络：
class DenseNet(nn.Module):r"""Densenet-BC model class, based on`"Densely Connected Convolutional Networks" <https://arxiv.org/pdf/1608.06993.pdf>`Args:growth_rate (int) - how many filters to add each layer (`k` in paper)block_config (list of 3 or 4 ints) - how many layers in each pooling blocknum_init_features (int) - the number of filters to learn in the first convolution layerbn_size (int) - multiplicative factor for number of bottle neck layers(i.e. bn_size * k features in the bottleneck layer)drop_rate (float) - dropout rate after each dense layernum_classes (int) - number of classification classes"""def __init__(self, growth_rate=32, block_config=(6, 12, 24, 16),num_init_features=24, bn_size=4, compression=0.5, drop_rate=0,num_classes=1000):super(DenseNet, self).__init__()# First Conv2dself.features = nn.Sequential(OrderedDict([('conv0', nn.Conv2d(3, num_init_features, kernel_size=7, stride=2, padding=3, bias=False)),('norm0', nn.BatchNorm2d(num_init_features)),('relu0', nn.ReLU(inplace=True)),('pool0', nn.MaxPool2d(kernel_size=3, stride=2, padding=1))]))# Each denseblocknum_features = num_init_featuresfor i, num_layers in enumerate(block_config):block = _DenseBlock(num_layers, num_features, bn_size, growth_rate, drop_rate)self.features.add_module('denseblock%d' % (i + 1), block)num_features += num_layers * growth_rateif i != len(block_config) - 1:transition = _Transition(num_input_features=num_features,num_output_features=int(num_features * compression))self.features.add_module('transition%d' % (i + 1), transition)num_features = int(num_features * compression)# Final bn+reluself.features.add_module('norm5', nn.BatchNorm2d(num_features))self.features.add_module('relu5', nn.ReLU(inplace=True))# classification layerself.classifier = nn.Linear(num_features, num_classes)# params initializationfor m in self.modules():if isinstance(m, nn.Conv2d):nn.init.kaiming_normal_(m.weight)elif isinstance(m, nn.BatchNorm2d):nn.init.constant_(m.bias, 0)nn.init.constant_(m.weight, 1)elif isinstance(m, nn.Linear):nn.init.constant_(m.bias, 0)def forward(self, x):features = self.features(x)out = F.avg_pool2d(features, 7, stride=1).view(features.size(0), -1)out = self.classifier(out)return outmodel_urls = {'densenet121': 'https://download.pytorch.org/models/densenet121-a639ec97.pth','densenet169': 'https://download.pytorch.org/models/densenet169-b2777c0a.pth','densenet201': 'https://download.pytorch.org/models/densenet201-c1103571.pth','densenet161': 'https://download.pytorch.org/models/densenet161-8d451a50.pth'}def densenet121(pretrained=False, **kwargs):"""DenseNet121"""model = DenseNet(num_init_features=64, growth_rate=32, block_config=(6, 12, 24, 16), **kwargs)if pretrained:# '.'s are no longer allowed in module names, but pervious _DenseLayer# has keys 'norm.1', 'relu.1', 'conv.1', 'norm.2', 'relu.2', 'conv.2'.# They are also in the checkpoints in model_urls. This pattern is used# to find such keys.pattern = re.compile(r'^(.*denselayer\d+\.(?:norm|relu|conv))\.((?:[12])\.(?:weight|bias|running_mean|running_var))$')state_dict = model_zoo.load_url(model_urls['densenet121'])for key in list(state_dict.keys()):res = pattern.match(key)if res:new_key = res.group(1) + res.group(2)state_dict[new_key] = state_dict[key]del state_dict[key]model.load_state_dict(state_dict)return model"""搭建densenet121模型"""
# model = densenet121().to(device)
model = densenet121(True).to(device) # 使用预训练模型
print(model)
print(summary.summary(model, (3, 224, 224))) # 查看模型的参数量以及相关指标输出
----------------------------------------------------------------Layer (type) Output Shape Param #
================================================================Conv2d-1 [-1, 64, 112, 112] 9,408BatchNorm2d-2 [-1, 64, 112, 112] 128ReLU-3 [-1, 64, 112, 112] 0MaxPool2d-4 [-1, 64, 56, 56] 0BatchNorm2d-5 [-1, 64, 56, 56] 128ReLU-6 [-1, 64, 56, 56] 0Conv2d-7 [-1, 128, 56, 56] 8,192BatchNorm2d-8 [-1, 128, 56, 56] 256ReLU-9 [-1, 128, 56, 56] 0Conv2d-10 [-1, 32, 56, 56] 36,864BatchNorm2d-11 [-1, 96, 56, 56] 192ReLU-12 [-1, 96, 56, 56] 0Conv2d-13 [-1, 128, 56, 56] 12,288BatchNorm2d-14 [-1, 128, 56, 56] 256ReLU-15 [-1, 128, 56, 56] 0Conv2d-16 [-1, 32, 56, 56] 36,864BatchNorm2d-17 [-1, 128, 56, 56] 256ReLU-18 [-1, 128, 56, 56] 0Conv2d-19 [-1, 128, 56, 56] 16,384BatchNorm2d-20 [-1, 128, 56, 56] 256ReLU-21 [-1, 128, 56, 56] 0Conv2d-22 [-1, 32, 56, 56] 36,864BatchNorm2d-23 [-1, 160, 56, 56] 320ReLU-24 [-1, 160, 56, 56] 0Conv2d-25 [-1, 128, 56, 56] 20,480BatchNorm2d-26 [-1, 128, 56, 56] 256ReLU-27 [-1, 128, 56, 56] 0Conv2d-28 [-1, 32, 56, 56] 36,864BatchNorm2d-29 [-1, 192, 56, 56] 384ReLU-30 [-1, 192, 56, 56] 0Conv2d-31 [-1, 128, 56, 56] 24,576BatchNorm2d-32 [-1, 128, 56, 56] 256ReLU-33 [-1, 128, 56, 56] 0Conv2d-34 [-1, 32, 56, 56] 36,864BatchNorm2d-35 [-1, 224, 56, 56] 448ReLU-36 [-1, 224, 56, 56] 0Conv2d-37 [-1, 128, 56, 56] 28,672BatchNorm2d-38 [-1, 128, 56, 56] 256ReLU-39 [-1, 128, 56, 56] 0Conv2d-40 [-1, 32, 56, 56] 36,864BatchNorm2d-41 [-1, 256, 56, 56] 512ReLU-42 [-1, 256, 56, 56] 0Conv2d-43 [-1, 128, 56, 56] 32,768AvgPool2d-44 [-1, 128, 28, 28] 0BatchNorm2d-45 [-1, 128, 28, 28] 256ReLU-46 [-1, 128, 28, 28] 0Conv2d-47 [-1, 128, 28, 28] 16,384BatchNorm2d-48 [-1, 128, 28, 28] 256ReLU-49 [-1, 128, 28, 28] 0Conv2d-50 [-1, 32, 28, 28] 36,864BatchNorm2d-51 [-1, 160, 28, 28] 320ReLU-52 [-1, 160, 28, 28] 0Conv2d-53 [-1, 128, 28, 28] 20,480BatchNorm2d-54 [-1, 128, 28, 28] 256ReLU-55 [-1, 128, 28, 28] 0Conv2d-56 [-1, 32, 28, 28] 36,864BatchNorm2d-57 [-1, 192, 28, 28] 384ReLU-58 [-1, 192, 28, 28] 0Conv2d-59 [-1, 128, 28, 28] 24,576BatchNorm2d-60 [-1, 128, 28, 28] 256ReLU-61 [-1, 128, 28, 28] 0Conv2d-62 [-1, 32, 28, 28] 36,864BatchNorm2d-63 [-1, 224, 28, 28] 448ReLU-64 [-1, 224, 28, 28] 0Conv2d-65 [-1, 128, 28, 28] 28,672BatchNorm2d-66 [-1, 128, 28, 28] 256ReLU-67 [-1, 128, 28, 28] 0Conv2d-68 [-1, 32, 28, 28] 36,864BatchNorm2d-69 [-1, 256, 28, 28] 512ReLU-70 [-1, 256, 28, 28] 0Conv2d-71 [-1, 128, 28, 28] 32,768BatchNorm2d-72 [-1, 128, 28, 28] 256ReLU-73 [-1, 128, 28, 28] 0Conv2d-74 [-1, 32, 28, 28] 36,864BatchNorm2d-75 [-1, 288, 28, 28] 576ReLU-76 [-1, 288, 28, 28] 0Conv2d-77 [-1, 128, 28, 28] 36,864BatchNorm2d-78 [-1, 128, 28, 28] 256ReLU-79 [-1, 128, 28, 28] 0Conv2d-80 [-1, 32, 28, 28] 36,864BatchNorm2d-81 [-1, 320, 28, 28] 640ReLU-82 [-1, 320, 28, 28] 0Conv2d-83 [-1, 128, 28, 28] 40,960BatchNorm2d-84 [-1, 128, 28, 28] 256ReLU-85 [-1, 128, 28, 28] 0Conv2d-86 [-1, 32, 28, 28] 36,864BatchNorm2d-87 [-1, 352, 28, 28] 704ReLU-88 [-1, 352, 28, 28] 0Conv2d-89 [-1, 128, 28, 28] 45,056BatchNorm2d-90 [-1, 128, 28, 28] 256ReLU-91 [-1, 128, 28, 28] 0Conv2d-92 [-1, 32, 28, 28] 36,864BatchNorm2d-93 [-1, 384, 28, 28] 768ReLU-94 [-1, 384, 28, 28] 0Conv2d-95 [-1, 128, 28, 28] 49,152BatchNorm2d-96 [-1, 128, 28, 28] 256ReLU-97 [-1, 128, 28, 28] 0Conv2d-98 [-1, 32, 28, 28] 36,864BatchNorm2d-99 [-1, 416, 28, 28] 832ReLU-100 [-1, 416, 28, 28] 0Conv2d-101 [-1, 128, 28, 28] 53,248BatchNorm2d-102 [-1, 128, 28, 28] 256ReLU-103 [-1, 128, 28, 28] 0Conv2d-104 [-1, 32, 28, 28] 36,864BatchNorm2d-105 [-1, 448, 28, 28] 896ReLU-106 [-1, 448, 28, 28] 0Conv2d-107 [-1, 128, 28, 28] 57,344BatchNorm2d-108 [-1, 128, 28, 28] 256ReLU-109 [-1, 128, 28, 28] 0Conv2d-110 [-1, 32, 28, 28] 36,864BatchNorm2d-111 [-1, 480, 28, 28] 960ReLU-112 [-1, 480, 28, 28] 0Conv2d-113 [-1, 128, 28, 28] 61,440BatchNorm2d-114 [-1, 128, 28, 28] 256ReLU-115 [-1, 128, 28, 28] 0Conv2d-116 [-1, 32, 28, 28] 36,864BatchNorm2d-117 [-1, 512, 28, 28] 1,024ReLU-118 [-1, 512, 28, 28] 0Conv2d-119 [-1, 256, 28, 28] 131,072AvgPool2d-120 [-1, 256, 14, 14] 0BatchNorm2d-121 [-1, 256, 14, 14] 512ReLU-122 [-1, 256, 14, 14] 0Conv2d-123 [-1, 128, 14, 14] 32,768BatchNorm2d-124 [-1, 128, 14, 14] 256ReLU-125 [-1, 128, 14, 14] 0Conv2d-126 [-1, 32, 14, 14] 36,864BatchNorm2d-127 [-1, 288, 14, 14] 576ReLU-128 [-1, 288, 14, 14] 0Conv2d-129 [-1, 128, 14, 14] 36,864BatchNorm2d-130 [-1, 128, 14, 14] 256ReLU-131 [-1, 128, 14, 14] 0Conv2d-132 [-1, 32, 14, 14] 36,864BatchNorm2d-133 [-1, 320, 14, 14] 640ReLU-134 [-1, 320, 14, 14] 0Conv2d-135 [-1, 128, 14, 14] 40,960BatchNorm2d-136 [-1, 128, 14, 14] 256ReLU-137 [-1, 128, 14, 14] 0Conv2d-138 [-1, 32, 14, 14] 36,864BatchNorm2d-139 [-1, 352, 14, 14] 704ReLU-140 [-1, 352, 14, 14] 0Conv2d-141 [-1, 128, 14, 14] 45,056BatchNorm2d-142 [-1, 128, 14, 14] 256ReLU-143 [-1, 128, 14, 14] 0Conv2d-144 [-1, 32, 14, 14] 36,864BatchNorm2d-145 [-1, 384, 14, 14] 768ReLU-146 [-1, 384, 14, 14] 0Conv2d-147 [-1, 128, 14, 14] 49,152BatchNorm2d-148 [-1, 128, 14, 14] 256ReLU-149 [-1, 128, 14, 14] 0Conv2d-150 [-1, 32, 14, 14] 36,864BatchNorm2d-151 [-1, 416, 14, 14] 832ReLU-152 [-1, 416, 14, 14] 0Conv2d-153 [-1, 128, 14, 14] 53,248BatchNorm2d-154 [-1, 128, 14, 14] 256ReLU-155 [-1, 128, 14, 14] 0Conv2d-156 [-1, 32, 14, 14] 36,864BatchNorm2d-157 [-1, 448, 14, 14] 896ReLU-158 [-1, 448, 14, 14] 0Conv2d-159 [-1, 128, 14, 14] 57,344BatchNorm2d-160 [-1, 128, 14, 14] 256ReLU-161 [-1, 128, 14, 14] 0Conv2d-162 [-1, 32, 14, 14] 36,864BatchNorm2d-163 [-1, 480, 14, 14] 960ReLU-164 [-1, 480, 14, 14] 0Conv2d-165 [-1, 128, 14, 14] 61,440BatchNorm2d-166 [-1, 128, 14, 14] 256ReLU-167 [-1, 128, 14, 14] 0Conv2d-168 [-1, 32, 14, 14] 36,864BatchNorm2d-169 [-1, 512, 14, 14] 1,024ReLU-170 [-1, 512, 14, 14] 0Conv2d-171 [-1, 128, 14, 14] 65,536BatchNorm2d-172 [-1, 128, 14, 14] 256ReLU-173 [-1, 128, 14, 14] 0Conv2d-174 [-1, 32, 14, 14] 36,864BatchNorm2d-175 [-1, 544, 14, 14] 1,088ReLU-176 [-1, 544, 14, 14] 0Conv2d-177 [-1, 128, 14, 14] 69,632BatchNorm2d-178 [-1, 128, 14, 14] 256ReLU-179 [-1, 128, 14, 14] 0Conv2d-180 [-1, 32, 14, 14] 36,864BatchNorm2d-181 [-1, 576, 14, 14] 1,152ReLU-182 [-1, 576, 14, 14] 0Conv2d-183 [-1, 128, 14, 14] 73,728BatchNorm2d-184 [-1, 128, 14, 14] 256ReLU-185 [-1, 128, 14, 14] 0Conv2d-186 [-1, 32, 14, 14] 36,864BatchNorm2d-187 [-1, 608, 14, 14] 1,216ReLU-188 [-1, 608, 14, 14] 0Conv2d-189 [-1, 128, 14, 14] 77,824BatchNorm2d-190 [-1, 128, 14, 14] 256ReLU-191 [-1, 128, 14, 14] 0Conv2d-192 [-1, 32, 14, 14] 36,864BatchNorm2d-193 [-1, 640, 14, 14] 1,280ReLU-194 [-1, 640, 14, 14] 0Conv2d-195 [-1, 128, 14, 14] 81,920BatchNorm2d-196 [-1, 128, 14, 14] 256ReLU-197 [-1, 128, 14, 14] 0Conv2d-198 [-1, 32, 14, 14] 36,864BatchNorm2d-199 [-1, 672, 14, 14] 1,344ReLU-200 [-1, 672, 14, 14] 0Conv2d-201 [-1, 128, 14, 14] 86,016BatchNorm2d-202 [-1, 128, 14, 14] 256ReLU-203 [-1, 128, 14, 14] 0Conv2d-204 [-1, 32, 14, 14] 36,864BatchNorm2d-205 [-1, 704, 14, 14] 1,408ReLU-206 [-1, 704, 14, 14] 0Conv2d-207 [-1, 128, 14, 14] 90,112BatchNorm2d-208 [-1, 128, 14, 14] 256ReLU-209 [-1, 128, 14, 14] 0Conv2d-210 [-1, 32, 14, 14] 36,864BatchNorm2d-211 [-1, 736, 14, 14] 1,472ReLU-212 [-1, 736, 14, 14] 0Conv2d-213 [-1, 128, 14, 14] 94,208BatchNorm2d-214 [-1, 128, 14, 14] 256ReLU-215 [-1, 128, 14, 14] 0Conv2d-216 [-1, 32, 14, 14] 36,864BatchNorm2d-217 [-1, 768, 14, 14] 1,536ReLU-218 [-1, 768, 14, 14] 0Conv2d-219 [-1, 128, 14, 14] 98,304BatchNorm2d-220 [-1, 128, 14, 14] 256ReLU-221 [-1, 128, 14, 14] 0Conv2d-222 [-1, 32, 14, 14] 36,864BatchNorm2d-223 [-1, 800, 14, 14] 1,600ReLU-224 [-1, 800, 14, 14] 0Conv2d-225 [-1, 128, 14, 14] 102,400BatchNorm2d-226 [-1, 128, 14, 14] 256ReLU-227 [-1, 128, 14, 14] 0Conv2d-228 [-1, 32, 14, 14] 36,864BatchNorm2d-229 [-1, 832, 14, 14] 1,664ReLU-230 [-1, 832, 14, 14] 0Conv2d-231 [-1, 128, 14, 14] 106,496BatchNorm2d-232 [-1, 128, 14, 14] 256ReLU-233 [-1, 128, 14, 14] 0Conv2d-234 [-1, 32, 14, 14] 36,864BatchNorm2d-235 [-1, 864, 14, 14] 1,728ReLU-236 [-1, 864, 14, 14] 0Conv2d-237 [-1, 128, 14, 14] 110,592BatchNorm2d-238 [-1, 128, 14, 14] 256ReLU-239 [-1, 128, 14, 14] 0Conv2d-240 [-1, 32, 14, 14] 36,864BatchNorm2d-241 [-1, 896, 14, 14] 1,792ReLU-242 [-1, 896, 14, 14] 0Conv2d-243 [-1, 128, 14, 14] 114,688BatchNorm2d-244 [-1, 128, 14, 14] 256ReLU-245 [-1, 128, 14, 14] 0Conv2d-246 [-1, 32, 14, 14] 36,864BatchNorm2d-247 [-1, 928, 14, 14] 1,856ReLU-248 [-1, 928, 14, 14] 0Conv2d-249 [-1, 128, 14, 14] 118,784BatchNorm2d-250 [-1, 128, 14, 14] 256ReLU-251 [-1, 128, 14, 14] 0Conv2d-252 [-1, 32, 14, 14] 36,864BatchNorm2d-253 [-1, 960, 14, 14] 1,920ReLU-254 [-1, 960, 14, 14] 0Conv2d-255 [-1, 128, 14, 14] 122,880BatchNorm2d-256 [-1, 128, 14, 14] 256ReLU-257 [-1, 128, 14, 14] 0Conv2d-258 [-1, 32, 14, 14] 36,864BatchNorm2d-259 [-1, 992, 14, 14] 1,984ReLU-260 [-1, 992, 14, 14] 0Conv2d-261 [-1, 128, 14, 14] 126,976BatchNorm2d-262 [-1, 128, 14, 14] 256ReLU-263 [-1, 128, 14, 14] 0Conv2d-264 [-1, 32, 14, 14] 36,864BatchNorm2d-265 [-1, 1024, 14, 14] 2,048ReLU-266 [-1, 1024, 14, 14] 0Conv2d-267 [-1, 512, 14, 14] 524,288AvgPool2d-268 [-1, 512, 7, 7] 0BatchNorm2d-269 [-1, 512, 7, 7] 1,024ReLU-270 [-1, 512, 7, 7] 0Conv2d-271 [-1, 128, 7, 7] 65,536BatchNorm2d-272 [-1, 128, 7, 7] 256ReLU-273 [-1, 128, 7, 7] 0Conv2d-274 [-1, 32, 7, 7] 36,864BatchNorm2d-275 [-1, 544, 7, 7] 1,088ReLU-276 [-1, 544, 7, 7] 0Conv2d-277 [-1, 128, 7, 7] 69,632BatchNorm2d-278 [-1, 128, 7, 7] 256ReLU-279 [-1, 128, 7, 7] 0Conv2d-280 [-1, 32, 7, 7] 36,864BatchNorm2d-281 [-1, 576, 7, 7] 1,152ReLU-282 [-1, 576, 7, 7] 0Conv2d-283 [-1, 128, 7, 7] 73,728BatchNorm2d-284 [-1, 128, 7, 7] 256ReLU-285 [-1, 128, 7, 7] 0Conv2d-286 [-1, 32, 7, 7] 36,864BatchNorm2d-287 [-1, 608, 7, 7] 1,216ReLU-288 [-1, 608, 7, 7] 0Conv2d-289 [-1, 128, 7, 7] 77,824BatchNorm2d-290 [-1, 128, 7, 7] 256ReLU-291 [-1, 128, 7, 7] 0Conv2d-292 [-1, 32, 7, 7] 36,864BatchNorm2d-293 [-1, 640, 7, 7] 1,280ReLU-294 [-1, 640, 7, 7] 0Conv2d-295 [-1, 128, 7, 7] 81,920BatchNorm2d-296 [-1, 128, 7, 7] 256ReLU-297 [-1, 128, 7, 7] 0Conv2d-298 [-1, 32, 7, 7] 36,864BatchNorm2d-299 [-1, 672, 7, 7] 1,344ReLU-300 [-1, 672, 7, 7] 0Conv2d-301 [-1, 128, 7, 7] 86,016BatchNorm2d-302 [-1, 128, 7, 7] 256ReLU-303 [-1, 128, 7, 7] 0Conv2d-304 [-1, 32, 7, 7] 36,864BatchNorm2d-305 [-1, 704, 7, 7] 1,408ReLU-306 [-1, 704, 7, 7] 0Conv2d-307 [-1, 128, 7, 7] 90,112BatchNorm2d-308 [-1, 128, 7, 7] 256ReLU-309 [-1, 128, 7, 7] 0Conv2d-310 [-1, 32, 7, 7] 36,864BatchNorm2d-311 [-1, 736, 7, 7] 1,472ReLU-312 [-1, 736, 7, 7] 0Conv2d-313 [-1, 128, 7, 7] 94,208BatchNorm2d-314 [-1, 128, 7, 7] 256ReLU-315 [-1, 128, 7, 7] 0Conv2d-316 [-1, 32, 7, 7] 36,864BatchNorm2d-317 [-1, 768, 7, 7] 1,536ReLU-318 [-1, 768, 7, 7] 0Conv2d-319 [-1, 128, 7, 7] 98,304BatchNorm2d-320 [-1, 128, 7, 7] 256ReLU-321 [-1, 128, 7, 7] 0Conv2d-322 [-1, 32, 7, 7] 36,864BatchNorm2d-323 [-1, 800, 7, 7] 1,600ReLU-324 [-1, 800, 7, 7] 0Conv2d-325 [-1, 128, 7, 7] 102,400BatchNorm2d-326 [-1, 128, 7, 7] 256ReLU-327 [-1, 128, 7, 7] 0Conv2d-328 [-1, 32, 7, 7] 36,864BatchNorm2d-329 [-1, 832, 7, 7] 1,664ReLU-330 [-1, 832, 7, 7] 0Conv2d-331 [-1, 128, 7, 7] 106,496BatchNorm2d-332 [-1, 128, 7, 7] 256ReLU-333 [-1, 128, 7, 7] 0Conv2d-334 [-1, 32, 7, 7] 36,864BatchNorm2d-335 [-1, 864, 7, 7] 1,728ReLU-336 [-1, 864, 7, 7] 0Conv2d-337 [-1, 128, 7, 7] 110,592BatchNorm2d-338 [-1, 128, 7, 7] 256ReLU-339 [-1, 128, 7, 7] 0Conv2d-340 [-1, 32, 7, 7] 36,864BatchNorm2d-341 [-1, 896, 7, 7] 1,792ReLU-342 [-1, 896, 7, 7] 0Conv2d-343 [-1, 128, 7, 7] 114,688BatchNorm2d-344 [-1, 128, 7, 7] 256ReLU-345 [-1, 128, 7, 7] 0Conv2d-346 [-1, 32, 7, 7] 36,864BatchNorm2d-347 [-1, 928, 7, 7] 1,856ReLU-348 [-1, 928, 7, 7] 0Conv2d-349 [-1, 128, 7, 7] 118,784BatchNorm2d-350 [-1, 128, 7, 7] 256ReLU-351 [-1, 128, 7, 7] 0Conv2d-352 [-1, 32, 7, 7] 36,864BatchNorm2d-353 [-1, 960, 7, 7] 1,920ReLU-354 [-1, 960, 7, 7] 0Conv2d-355 [-1, 128, 7, 7] 122,880BatchNorm2d-356 [-1, 128, 7, 7] 256ReLU-357 [-1, 128, 7, 7] 0Conv2d-358 [-1, 32, 7, 7] 36,864BatchNorm2d-359 [-1, 992, 7, 7] 1,984ReLU-360 [-1, 992, 7, 7] 0Conv2d-361 [-1, 128, 7, 7] 126,976BatchNorm2d-362 [-1, 128, 7, 7] 256ReLU-363 [-1, 128, 7, 7] 0Conv2d-364 [-1, 32, 7, 7] 36,864BatchNorm2d-365 [-1, 1024, 7, 7] 2,048ReLU-366 [-1, 1024, 7, 7] 0Linear-367 [-1, 1000] 1,025,000
================================================================
Total params: 7,978,856
Trainable params: 7,978,856
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 294.58
Params size (MB): 30.44
Estimated Total Size (MB): 325.59
----------------------------------------------------------------
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losstrain_acc += (pred.argmax(1) == y).type(torch.float).sum().item()train_loss += loss.item()train_acc /= sizetrain_loss /= num_batchesreturn 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()#统计预测正确的个数test_acc /= sizetest_loss /= num_batchesreturn test_acc, test_loss2.3.4 正式训练
"""训练模型--正式训练"""
epochs = 10
train_loss = []
train_acc = []
test_loss = []
test_acc = []
best_test_acc=0for 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')
输出最高精度为Test_acc:100%
Epoch: 1, duration:74420ms, Train_acc:83.7%, Train_loss:0.902, Test_acc:85.8%，Test_loss:0.345, Lr:1.00E-04,Update the best model
Epoch: 2, duration:72587ms, Train_acc:86.4%, Train_loss:0.329, Test_acc:85.5%，Test_loss:0.343, Lr:1.00E-04
Epoch: 3, duration:72941ms, Train_acc:87.9%, Train_loss:0.292, Test_acc:89.2%，Test_loss:0.262, Lr:1.00E-04,Update the best model
Epoch: 4, duration:74155ms, Train_acc:88.8%, Train_loss:0.279, Test_acc:89.7%，Test_loss:0.248, Lr:1.00E-04,Update the best model
Epoch: 5, duration:75123ms, Train_acc:89.1%, Train_loss:0.265, Test_acc:89.0%，Test_loss:0.277, Lr:1.00E-04
Epoch: 6, duration:74381ms, Train_acc:89.6%, Train_loss:0.255, Test_acc:90.5%，Test_loss:0.249, Lr:1.00E-04,Update the best model
Epoch: 7, duration:73710ms, Train_acc:90.2%, Train_loss:0.243, Test_acc:84.1%，Test_loss:0.369, Lr:1.00E-04
Epoch: 8, duration:73995ms, Train_acc:90.7%, Train_loss:0.230, Test_acc:89.5%，Test_loss:0.250, Lr:1.00E-04
Epoch: 9, duration:73017ms, Train_acc:90.7%, Train_loss:0.223, Test_acc:89.3%，Test_loss:0.263, Lr:1.00E-04
Epoch:10, duration:73960ms, Train_acc:91.2%, Train_loss:0.224, Test_acc:91.6%，Test_loss:0.209, Lr:1.00E-04,Update the best model
Epoch:11, duration:74113ms, Train_acc:91.2%, Train_loss:0.219, Test_acc:90.5%，Test_loss:0.225, Lr:1.00E-04
Epoch:12, duration:73573ms, Train_acc:91.5%, Train_loss:0.213, Test_acc:88.5%，Test_loss:0.273, Lr:1.00E-04
Epoch:13, duration:73206ms, Train_acc:92.2%, Train_loss:0.202, Test_acc:85.1%，Test_loss:0.377, Lr:1.00E-04
Epoch:14, duration:73540ms, Train_acc:92.1%, Train_loss:0.195, Test_acc:91.2%，Test_loss:0.225, Lr:1.00E-04
Epoch:15, duration:73378ms, Train_acc:92.3%, Train_loss:0.192, Test_acc:87.6%，Test_loss:0.796, Lr:1.00E-04
Epoch:16, duration:73195ms, Train_acc:92.5%, Train_loss:0.187, Test_acc:92.5%，Test_loss:0.197, Lr:1.00E-04,Update the best model
Epoch:17, duration:73737ms, Train_acc:93.1%, Train_loss:0.174, Test_acc:92.7%，Test_loss:0.186, Lr:1.00E-04,Update the best model
Epoch:18, duration:73884ms, Train_acc:93.4%, Train_loss:0.171, Test_acc:80.6%，Test_loss:0.463, Lr:1.00E-04
Epoch:19, duration:73239ms, Train_acc:93.2%, Train_loss:0.168, Test_acc:91.2%，Test_loss:0.221, Lr:1.00E-04
Epoch:20, duration:73386ms, Train_acc:93.7%, Train_loss:0.159, Test_acc:92.5%，Test_loss:0.196, Lr:1.00E-04
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) / "Cockatoo/001.jpg"),model=model,transform=train_transforms,classes=classes)
输出
预测结果是：0
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
预测结果是：0
0.9268929503916449 0.1855085204311073 知识点详解
3.1 nn.Sequential和nn.Module区别与选择
3.1.1 nn.Sequential
torch.nn.Sequential是一个Sequential容器，模块将按照构造函数中传递的顺序添加到模块中。另外，也可以传入一个有序模块。 为了更容易理解，官方给出了一些案例：
# Sequential使用实例model = nn.Sequential(nn.Conv2d(1,20,5),nn.ReLU(),nn.Conv2d(20,64,5),nn.ReLU())# Sequential with OrderedDict使用实例
model = nn.Sequential(OrderedDict([('conv1', nn.Conv2d(1,20,5)),('relu1', nn.ReLU()),('conv2', nn.Conv2d(20,64,5)),('relu2', nn.ReLU())]))
3.1.2 nn.Module
下面我们再用 Module 定义这个模型，下面是使用 Module 的模板
class 网络名字(nn.Module):def __init__(self, 一些定义的参数):super(网络名字, self).__init__()self.layer1 = nn.Linear(num_input, num_hidden)self.layer2 = nn.Sequential(...)...定义需要用的网络层def forward(self, x): # 定义前向传播x1 = self.layer1(x)x2 = self.layer2(x)x = x1 + x2...return x
注意的是，Module 里面也可以使用 Sequential，同时 Module 非常灵活，具体体现在 forward 中，如何复杂的操作都能直观的在 forward 里面执行
3.1.3 对比
为了方便比较，我们先用普通方法搭建一个神经网络。
class Net(torch.nn.Module):def __init__(self, n_feature, n_hidden, n_output):super(Net, self).__init__()self.hidden = torch.nn.Linear(n_feature, n_hidden)self.predict = torch.nn.Linear(n_hidden, n_output)def forward(self, x):x = F.relu(self.hidden(x))x = self.predict(x)return x
net1 = Net(1, 10, 1)net2 = torch.nn.Sequential(torch.nn.Linear(1, 10),torch.nn.ReLU(),torch.nn.Linear(10, 1)
)
打印这两个net
print(net1)
"""
Net ((hidden): Linear (1 -> 10)(predict): Linear (10 -> 1)
)
"""
print(net2)
"""
Sequential ((0): Linear (1 -> 10)(1): ReLU ()(2): Linear (10 -> 1)
)
"""
我们可以发现，打印torch.nn.Sequential会自动加入激励函数,
在 net1 中, 激励函数实际上是在 forward() 功能中被调用的,没有在init中定义,所以在打印网络结构时不会有激励函数的信息.
解析源码,在torch.nn.Sequential中:
def forward(self, input):for module in self:input = module(input)return input
可以看到,torch.nn.Sequential的forward只是简单的顺序传播,操作性有限.
3.1.4 总结
使用torch.nn.Module，我们可以根据自己的需求改变传播过程，如RNN等
如果你需要快速构建或者不需要过多的过程，直接使用torch.nn.Sequential即可。
参考链接：nn.Sequential和nn.Module区别与选择
3.2 python中OrderedDict的使用
很多人认为python中的字典是无序的，因为它是按照hash来存储的，但是python中有个模块collections(英文，收集、集合)，里面自带了一个子类
OrderedDict，实现了对字典对象中元素的排序。请看下面的实例：
import collections
print "Regular dictionary"
d={}
d['a']='A'
d['b']='B'
d['c']='C'
for k,v in d.items():print k,vprint "\nOrder dictionary"
d1 = collections.OrderedDict()
d1['a'] = 'A'
d1['b'] = 'B'
d1['c'] = 'C'
d1['1'] = '1'
d1['2'] = '2'
for k,v in d1.items():print k,v
输出：
Regular dictionary
a A
c C
b BOrder dictionary
a A
b B
c C
1 1
2 2
可以看到，同样是保存了ABC等几个元素，但是使用OrderedDict会根据放入元素的先后顺序进行排序。所以输出的值是排好序的。
OrderedDict对象的字典对象，如果其顺序不同那么Python也会把他们当做是两个不同的对象，请看事例：
print 'Regular dictionary:'
d2={}
d2['a']='A'
d2['b']='B'
d2['c']='C'd3={}
d3['c']='C'
d3['a']='A'
d3['b']='B'print d2 == d3print '\nOrderedDict:'
d4=collections.OrderedDict()
d4['a']='A'
d4['b']='B'
d4['c']='C'd5=collections.OrderedDict()
d5['c']='C'
d5['a']='A'
d5['b']='B'print d1==d2输出：
Regular dictionary:
TrueOrderedDict:
False
再看几个例子：
dd = {'banana': 3, 'apple':4, 'pear': 1, 'orange': 2}
#按key排序
kd = collections.OrderedDict(sorted(dd.items(), key=lambda t: t[0]))
print kd
#按照value排序
vd = collections.OrderedDict(sorted(dd.items(),key=lambda t:t[1]))
print vd#输出
OrderedDict([('apple', 4), ('banana', 3), ('orange', 2), ('pear', 1)])
OrderedDict([('pear', 1), ('orange', 2), ('banana', 3), ('apple', 4)])
总结
数据量越大，训练时间越长，在DataLoader中增加num_workers，即增加线程数量，可能会导致内存不足出现，Couldn‘t open shared file mapping或者Out of memery的错误，可尝试减小num_corkers。
