模型训练的准确率一直才百分之五十多 怎么调参数

Question

初学者，拿网上的代码跑自己的数据集，使用Resnet50预训练，调过学习率、batch_size、img_size的参数、换过优化器不过训练效果都差不多，还有什么办法提高准确率呢？

以下是训练代码：

import time
import torch
from torch import nn
from torch.utils.data import DataLoader
from PIL import Image
import torchvision.transforms as transforms
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
from torchvision.models import resnet50  # ResNet系列
from torchvision import models
from torch.utils.data import Dataset

transform_BZ= transforms.Normalize(
    mean=[0.44660836, 0.30800995, 0.21277349],# 取决于数据集
    std=[0.18726136, 0.13087553, 0.09021662]
)


class LoadData(Dataset):
    def __init__(self, txt_path, train_flag=True):
        self.imgs_info = self.get_images(txt_path)
        self.train_flag = train_flag
        self.img_size = 512

        self.train_tf = transforms.Compose([
                transforms.Resize(self.img_size),
                transforms.RandomHorizontalFlip(),#对图片进行随机的水平翻转
                transforms.RandomVerticalFlip(),#随机的垂直翻转
                transforms.ToTensor(),#把图片改为Tensor格式
                transform_BZ#图片标准化的步骤
            ])
        self.val_tf = transforms.Compose([##简单把图片压缩了变成Tensor模式
                transforms.Resize(self.img_size),
                transforms.ToTensor(),
                transform_BZ#标准化操作
            ])

    def get_images(self, txt_path):
        with open(txt_path, 'r', encoding='utf-8') as f:
            imgs_info = f.readlines()
            imgs_info = list(map(lambda x:x.strip().split('\t'), imgs_info))
        return imgs_info#返回图片信息

    def padding_black(self, img):   # 如果尺寸太小可以扩充
        w, h  = img.size
        scale = self.img_size / max(w, h)
        img_fg = img.resize([int(x) for x in [w * scale, h * scale]])
        size_fg = img_fg.size
        size_bg = self.img_size
        img_bg = Image.new("RGB", (size_bg, size_bg))
        img_bg.paste(img_fg, ((size_bg - size_fg[0]) // 2,
                              (size_bg - size_fg[1]) // 2))
        img = img_bg
        return img

    def __getitem__(self, index):#返回真正想返回的东西
        img_path, label = self.imgs_info[index]
        img = Image.open(img_path)#打开图片
        img = img.convert('RGB')#转换为RGB 格式
        img = self.padding_black(img)
        if self.train_flag:
            img = self.train_tf(img)
        else:
            img = self.val_tf(img)
        label = int(label)

        return img, label

    def __len__(self):
        return len(self.imgs_info)

def WriteData(fname, *args):
    with open(fname, 'a+') as f:
        for data in args:
            f.write(str(data)+"\t")
        f.write("\n")


def train(dataloader, model, loss_fn, optimizer,device):
    size = len(dataloader.dataset)
    avg_loss = 0
    # 从数据加载器中读取batch（一次读取多少张，即批次数），X(图片数据)，y（图片真实标签）。
    for batch, (X, y) in enumerate(dataloader):#固定格式：batch：第几批数据，不是批次大小，（X，y）：数值用括号

        # print(size)
        # 将数据存到显卡
        X, y = X.to(device), y.to(device)
        # 得到预测的结果pred
        pred = model(X)
        loss = loss_fn(pred, y)
        avg_loss += loss
        # 反向传播，更新模型参数
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        # 每训练10次，输出一次当前信息
        if batch % 10 == 0:
            loss, current = loss.item(), batch * len(X)
            print(f"loss: {loss:>7f}  [{current:>5d}/{size:>5d}]")

    # 当一个epoch完了后返回平均 loss
    avg_loss /= size
    avg_loss = avg_loss.detach().cpu().numpy()
    return avg_loss


def validate(dataloader, model, loss_fn, device):
    size = len(dataloader.dataset)
    # 将模型转为验证模式
    model.eval()
    # 初始化test_loss 和 correct， 用来统计每次的误差
    test_loss, correct = 0, 0
    # 测试时模型参数不用更新，所以no_gard()
    # 非训练， 推理期用到
    with torch.no_grad():
        # 加载数据加载器，得到里面的X（图片数据）和y(真实标签）

        for X, y in dataloader:
            # 将数据转到GPU
            X, y = X.to(device), y.to(device)
            # 将图片传入到模型当中就，得到预测的值pred
            pred = model(X)
            # 计算预测值pred和真实值y的差距
            test_loss += loss_fn(pred, y).item()
            # 统计预测正确的个数(针对分类)
            correct += (pred.argmax(1) == y).type(torch.float).sum().item()
    test_loss /= size
    correct /= size
    print(f"correct = {correct}, Test Error: \n Accuracy: {(100 * correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")
    return correct, test_loss


if __name__=='__main__':
    batch_size = 16

    # # 给训练集和测试集分别创建一个数据集加载器
    train_data = LoadData("/content/drive/MyDrive/data/train2500.txt", True)
    valid_data = LoadData("/content/drive/MyDrive/data/test2500.txt", False)


    train_dataloader = DataLoader(dataset=train_data, num_workers=4, pin_memory=True, batch_size=batch_size, shuffle=True)
    valid_dataloader = DataLoader(dataset=valid_data, num_workers=4, pin_memory=True, batch_size=batch_size)

    # 如果显卡可用，则用显卡进行训练
    device = "cuda" if torch.cuda.is_available() else "cpu"
    print(f"Using {device} device")

    finetune_net = resnet50(num_classes=5).to(device)
    state_dict=torch.load("/content/drive/MyDrive/data/resnet50_best.pth",map_location='cpu')
    finetune_net.load_state_dict(state_dict)

    nn.init.xavier_normal_(finetune_net.fc.weight)

    parms_1x = [value for name, value in finetune_net.named_parameters()
                if name not in ["fc.weight", "fc.bias"]]
    # 最后一层10倍学习率
    parms_10x = [value for name, value in finetune_net.named_parameters()
                 if name in ["fc.weight", "fc.bias"]]

    finetune_net = finetune_net.to(device)
    # 定义损失函数，计算相差多少，交叉熵，
    loss_fn = nn.CrossEntropyLoss()

    # 定义优化器，用来训练时候优化模型参数，随机梯度下降法
    learning_rate = 1e-3
    optimizer = torch.optim.Adam([
        {
            'params': parms_1x
        },
        {
            'params': parms_10x,
            'lr': learning_rate * 10
        }], lr=learning_rate)

    epochs = 3
    loss_ = 10
    save_root = "/content/drive/MyDrive/data/"


    for t in range(epochs):
        print(f"Epoch {t + 1}\n-------------------------------")
        time_start = time.time()
        avg_loss = train(train_dataloader, finetune_net, loss_fn, optimizer, device)
        time_end = time.time()
        print(f"train time: {(time_end - time_start)}")

        val_accuracy, val_loss = validate(valid_dataloader, finetune_net,loss_fn, device)
        # 写入数据
        WriteData(save_root + "resnet50_3.txt",
                  "epoch", t,
                  "train_loss", avg_loss,
                  "val_loss", val_loss,
                  "val_accuracy", val_accuracy)
        if t % 5 == 0:
            torch.save(finetune_net.state_dict(), save_root + "resnet50_3_epoch" + str(t) + "_loss_" + str(avg_loss) + ".pth")
        torch.save(finetune_net.state_dict(), save_root + "resnet50_3_last.pth")
        if avg_loss < loss_:
            loss_ = avg_loss
            torch.save(finetune_net.state_dict(), save_root + "resnet50_3_best.pth")

Answer 1

尝试增加数据集数量或者做数据增强，如果背景太复杂多样化会影响准确率的。