pytorch模型微调finetuning训练image_dog(kaggle)

对kaggle上的比赛 狗的品种识别进行训练。

 

​train文件夹有10222张狗的图片，test文件夹有10357张狗的图片，labels.csv文件放着train文件夹里的图片名称对应的label值。

模型：选择预训练的resnet50，冻住前面层的参数，将最后的全连接层改变使得输出类别变成120并对该层进行学习

代码：

import torch
import torchvision
from torch import nn
from torch.utils.data import Dataset
from PIL import Image
import os
from torchvision.transforms import transforms
import pandas as pd
from matplotlib import pyplot as plt
import numpy as np
import torch.optim as optim
from tqdm import tqdm #显示进度条的库


class dataset(Dataset):  # 获取文件夹里的图片，并返回图片和label

    def __init__(self, root_dir, imgs_labels, transform=None, valid_ratio=0.2, mode='train'):

        self.imgName = os.listdir(root_dir)
        self.data_len = len(self.imgName) - 1
        self.imgs_labels = imgs_labels
        self.transform = transform
        self.root_dir = root_dir
        self.mode = mode
        self.train_len = int(self.data_len * (1 - valid_ratio))

        if mode == 'train':
            self.real_len = int(len(self.imgName) * (1 - valid_ratio))
        elif mode == 'valid':
            self.real_len = self.data_len - self.train_len
        elif mode == 'test':
            self.real_len = len(self.imgName)

        print('Finished reading the {} set of Dataset ({} samples found)'
              .format(mode, self.real_len))

    def __getitem__(self, idx):
        if self.mode == 'train':
            self.imgtrainName = self.imgName[0:self.train_len]
            self.real_len = len(self.imgtrainName)
            img_item_path = os.path.join(self.root_dir, self.imgtrainName[idx])

            img = Image.open(img_item_path).convert('RGB')  # 转换成RGB形式的三通道图像
            if self.transform is not None:
                img = self.transform(img)
            label = self.imgs_labels[self.imgtrainName[idx].split('.')[0]]
            num_label = class_to_num[label]
            return img, num_label
        elif self.mode == 'valid':
            self.imgvalidName = self.imgName[self.train_len:]
            img_item_path = os.path.join(self.root_dir, self.imgvalidName[idx])
            img = Image.open(img_item_path).convert('RGB')  # 转换成RGB形式的三通道图像
            if self.transform is not None:
                img = self.transform(img)
            label = self.imgs_labels[self.imgvalidName[idx].split('.')[0]]
            num_label = class_to_num[label]
            return img, num_label
        elif self.mode == 'test':
            img_item_path = os.path.join(self.root_dir, self.imgPath[idx])
            img = Image.open(img_item_path).convert('RGB')  # 转换成RGB形式的三通道图像
            if self.transform is not None:
                img = self.transform(img)
            return img

    def __len__(self):
        return self.real_len


def get_net(device):
    finetune_net = nn.Sequential()
    finetune_net.features = torchvision.models.resnet50(pretrained=True)
    finetune_net.output_new = nn.Sequential(nn.Linear(1000, 256),
                                            nn.ReLU(),
                                            nn.Linear(256, 120))
    # 将模型参数分配给用于计算的CPU或GPU
    finetune_net = finetune_net.to(device)
    # 冻结参数
    for param in finetune_net.features.parameters():
        param.requires_grad = False
    return finetune_net


def train(net, loss, optimizer, train_dataloader, device, batch_size):
    # 训练开始
    net.train()
    train_acc = 0
    # for imgs, targets in train_dataloader:
    for batch in tqdm(train_dataloader, desc='训练'):
        imgs, targets = batch
        imgs = imgs.to(device)
        targets = targets.to(device)
        output = net(imgs)
        Loss = loss(output, targets)
        # 优化器优化模型
        optimizer.zero_grad()
        Loss.backward()
        optimizer.step()
        _, pred = output.max(1)
        num_correct = (pred == targets).sum().item()
        acc = num_correct / (batch_size)
        train_acc += acc

    print("epoch: {}, Loss: {}, Acc: {}".format(epoch, Loss.item(), train_acc / len(train_dataloader)))
    train_acces.append(train_acc / len(train_dataloader))
    train_losses.append(Loss.item())
    return train_losses, train_acces


def valid(net, loss, valid_dataloader, device, batch_size):
    net.eval()
    eval_loss = 0
    eval_acc = 0
    with torch.no_grad():
        # for imgs, targets in test_dataloader:
        for batch in tqdm(valid_dataloader, desc='验证'):
            imgs, targets = batch
            imgs = imgs.to(device)
            targets = targets.to(device)
            output = net(imgs)
            Loss = loss(output, targets)

            _, pred = output.max(1)
            num_correct = (pred == targets).sum().item()
            eval_loss += Loss
            acc = num_correct / (batch_size)
            eval_acc += acc

        eval_losses = eval_loss / (len(valid_dataloader))
        eval_acc = eval_acc / (len(valid_dataloader))
        eval_acces.append(eval_acc)
        print("整体验证集上的Loss: {}".format(eval_losses))
        print("整体验证集上的正确率: {}".format(eval_acc))
    return eval_acces

def show_acces(train_losses, train_acces, valid_acces, num_epoch):#对准确率和loss画图显得直观
    plt.plot(1 + np.arange(len(train_losses)), train_losses, linewidth=1.5, linestyle='dashed', label='train_losses')
    plt.plot(1 + np.arange(len(train_acces)), train_acces, linewidth=1.5, linestyle='dashed', label='train_acces')
    plt.plot(1 + np.arange(len(valid_acces)), valid_acces, linewidth=1.5, linestyle='dashed', label='valid_acces')
    plt.grid()
    plt.xlabel('epoch')
    # plt.xlim((1,num_epoch))
    plt.xticks(range(1, 1 + num_epoch, 1))
    plt.legend()
    plt.show()


if __name__ == '__main__':
    train_path = r'/kaggle/input/dogbreedidentification/train/'
    test_path = r'/kaggle/input/dogbreedidentification/test/'
    imgs_train = os.listdir(train_path)
    imgs_test = os.listdir(test_path)
    print('训练集图片数：', len(imgs_train))
    print('验证集图片数：', len(imgs_test))
    data_info = pd.read_csv(r'/kaggle/input/dogbreedidentification/labels.csv')
    # print(data_info.sample(10))
    imgs = data_info.iloc[0:, 0]
    labels = data_info.iloc[0:, 1]
    imgs_labels = dict(zip(imgs, labels))


    train_root_dir = train_path
    test_root_dir = test_path
    labels = imgs_labels

    # 独热编码
    labels_sorted = sorted(list(set(data_info['breed'])))
    n_classes = len(labels_sorted)
    class_to_num = dict(zip(labels_sorted, range(n_classes)))
    print('len:', len(labels_sorted))
    num_to_class = {v: k for k, v in class_to_num.items()}

    # config
    batchsize = 64
    num_epoch = 5
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    net = get_net(device)
    net.to(device)
    loss = nn.CrossEntropyLoss()
    optimizer = optim.Adam([{'params': net.output_new.parameters()}], lr=0.001)

    train_losses = []
    train_acces = []
    eval_acces = []

    transform = {"train": transforms.Compose([
        # 随机裁剪图像，所得图像为原始面积的0.08到1之间，高宽比在3/4和4/3之间。
        # 然后，缩放图像以创建224x224的新图像
        transforms.RandomResizedCrop(224, scale=(0.08, 1.0),
                                     ratio=(3.0 / 4.0, 4.0 / 3.0)),
        transforms.RandomHorizontalFlip(),
        # 随机更改亮度，对比度和饱和度
        transforms.ColorJitter(brightness=0.4,
                               contrast=0.4,
                               saturation=0.4),
        # 添加随机噪声
        transforms.ToTensor(),
        # 标准化图像的每个通道
        transforms.Normalize([0.485, 0.456, 0.406],
                             [0.229, 0.224, 0.225])]),
        "valid": transforms.Compose([
            torchvision.transforms.Resize(256),
            # 从图像中心裁切224x224大小的图片
            torchvision.transforms.CenterCrop(224),
            torchvision.transforms.ToTensor(),
            torchvision.transforms.Normalize([0.485, 0.456, 0.406],
                                             [0.229, 0.224, 0.225])])}
    train_dataset = dataset(train_root_dir, imgs_labels, transform=transform["train"], valid_ratio=0.1, mode='train')
    valid_dataset = dataset(train_root_dir, imgs_labels, transform=transform["valid"], valid_ratio=0.1, mode='valid')
    test_dataset = dataset(test_root_dir, imgs_labels, transform=transform_train, mode='test')
    train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=batchsize, shuffle=True)
    valid_dataloader = torch.utils.data.DataLoader(valid_dataset, batch_size=int(batchsize / 2), shuffle=True)
    test_dataloader = torch.utils.data.DataLoader(test_dataset, batch_size=batch_size, shuffle=True)

    #放几张样本图片看看
    examples = enumerate(train_dataloader)  # 组合成一个索引序列
    batch_idx, (example_data, example_targets) = next(examples)
    fig = plt.figure()
    for i in range(6):
        plt.subplot(2, 3, i + 1)
        # plt.tight_layout()
        img = example_data[i]
        img = np.asarray(img)
        plt.imshow(img.transpose((1, 2, 0)))
        plt.title(num_to_class[int(example_targets[i])])
        plt.xticks([])
        plt.yticks([])
    plt.show()

    print("参数数：{}".format(sum(x.numel() for x in net.parameters())))
    for epoch in range(num_epoch):
        print("——————第 {} 轮训练开始——————".format(epoch + 1))
        train_losses, train_acces = train(net, loss, optimizer, train_dataloader, device, batchsize)
        valid_acces = valid(net, loss, valid_dataloader, device, batchsize / 2)
    show_acces(train_losses, train_acces, valid_acces, num_epoch)

结果：

​

编辑

添加图片注释，不超过 140 字（可选）

​

编辑

添加图片注释，不超过 140 字（可选）

5个epoch使验证集上准确率0.83，不过2到3个epoch就已经收敛了（可能是这就是微调的吧，冻住了其他参数，就学习了最后的Linear层所以参数不多）。