GoogleNet_V3，构建程序

导入必要模块

import os
import numpy as np
import torch.nn as nn
import torch
from torch.utils.data import DataLoader
import torchvision.transforms as transforms
import torch.optim as optim
from matplotlib import pyplot as plt

导入数据处理模块和改造后的标签平滑损失函数

from GoogLeNet.C_GoogLeNet.tools.my_dataset import NCFMDataSet
from GoogLeNet_V3.E_GoogLeNet_v3.tools.common_tools import get_googlenet_v3, LabelSmoothingCrossEntropy

标签平滑：主要是给系统增加正则化，就是让标签不再是[0100]，变成[0.1 0.7 0.1 0.1]

class LabelSmoothingCrossEntropy(nn.Module):
    """
    Label Smoothing Cross Entropy Loss
    """
    def __init__(self, eps=0.001):
        super(LabelSmoothingCrossEntropy, self).__init__()
        self.eps = eps

    def forward(self, x, target):
        # CE(q, p) = - sigma(q_i * log(p_i))
        log_probs = torch.nn.functional.log_softmax(x, dim=-1)  # 实现  log(p_i)

        # H(q, p)
        H_pq = -log_probs.gather(dim=-1, index=target.unsqueeze(1))  # 只需要q_i == 1的地方， 此时已经得到CE
        H_pq = H_pq.squeeze(1)

        # H(u, p)
        H_uq = -log_probs.mean()  # 由于u是均匀分布，等价于求均值

        loss = (1 - self.eps) * H_pq + self.eps * H_uq
        return loss.mean()

通过os模块获取加载数据集的地址，和模型预训练参数
加载预训练参数，可以让后面训练时候加速，不用太多给epoch才收敛，也可以不加载，就是训练会久一点点
可以参考这篇：https://www.sohu.com/a/277322487_129720

BASE_DIR = os.path.dirname(os.path.abspath(__file__))
data_dir = os.path.join(BASE_DIR, "..", "..", "Data", "NCFM", "train")
path_state_dict = os.path.join(BASE_DIR, "..", "data", "inception_v3_google-1a9a5a14.pth")

进行数据预处理，对训练数据进行多尺度随机裁剪，反转，totensor，normalize
对验证集也是做了处理（老师说可以改善准确率，我一直不明白）
用了multcrop和dense

train_transform = transforms.Compose([
        transforms.Resize((342)),  # 299 / (224/256) = 342
        transforms.CenterCrop(299),
        transforms.RandomCrop(299),
        transforms.RandomHorizontalFlip(p=0.5),
        transforms.ToTensor(),
        transforms.Normalize(norm_mean, norm_std),
    ])

    normalizes = transforms.Normalize(norm_mean, norm_std)
    valid_transform = transforms.Compose([
        transforms.Resize((342, 342)),
        transforms.TenCrop(299, vertical_flip=False),
        transforms.Lambda(lambda crops: torch.stack([normalizes(transforms.ToTensor()(crop)) for crop in crops])),#这里用stack进行堆叠，前面裁剪出来的10张进行堆叠在一个,BCHW,等于B=10，b是0维度
    ])

    # 构建MyDataset实例
    train_data = NCFMDataSet(data_dir=data_dir, mode="train", transform=train_transform)
    valid_data = NCFMDataSet(data_dir=data_dir, mode="valid", transform=valid_transform)

    # 构建DataLoder
    train_loader = DataLoader(dataset=train_data, batch_size=BATCH_SIZE, shuffle=True)
    valid_loader = DataLoader(dataset=valid_data, batch_size=4)

加载数据的方式

class NCFMDataSet(Dataset):
    def __init__(self, data_dir, mode="train", split_n=0.9, rng_seed=620, transform=None):
        """
        鱼类分类任务的Dataset
        :param data_dir: str, 数据集所在路径
        :param transform: torch.transform，数据预处理
        """
        self.mode = mode
        self.data_dir = data_dir
        self.rng_seed = rng_seed
        self.split_n = split_n
        self.data_info = self._get_img_info()  # data_info存储所有图片路径和标签，在DataLoader中通过index读取样本
        self.transform = transform

    def __getitem__(self, index):#配合下面读取图片信息，返回img和标签
        path_img, label = self.data_info[index]
        img = Image.open(path_img).convert('RGB')     # 0~255