- 导入相关库
import collectionsimport mathimport osimport shutilimport pandas as pdimport torchimport torchvisionfrom torch import nnfrom d2l import torch as d2l- 下载数据集
d2l.DATA_HUB['cifar10_tiny'] = (d2l.DATA_URL + 'kaggle_cifar10_tiny.zip','2068874e4b9a9f0fb07ebe0ad2b29754449ccacd')# 如果使用完整的Kaggle竞赛的数据集，设置demo为Falsedemo = Trueif demo:data_dir = d2l.download_extract('cifar10_tiny')else:data_dir = '../data/kaggle/cifar-10/'- 整理数据集
# 查看数据集def read_csv_labels(fname):"""读取‘fname’来给标签字典返回一个文件名"""with open(fname, 'r') as f:lines = f.readlines()[1:]# readlines(): 每次读文档的一行，以后还需要逐步循环tokens = [l.rstrip().split(',') for l in lines]# rstrip(): 删除字符串后面（右面）的空格或特殊字符， 还有lstrip（左面）、strip(两面)return dict((name, label) for name, label in tokens)labels = read_csv_labels(os.path.join(data_dir, 'trainLabels.csv'))print('训练样本：', len(labels))print('类别：', len(set(labels.values())))# set(): 集合，里面不能包含重复的元素，接受一个list作为参数
将验证集从原始的训练集钟拆分出来
# 拆分数据集：训练集、验证集def copyfile(filename, target_dir):"""将文件复制到目标目录"""os.makedirs(target_dir, exist_ok=True)# 创建多层目录，exist_ok为True：在目标目录已存在的情况下不会触发FileExistsError异常。shutil.copy(filename, target_dir)#拷贝文件，filename：要拷贝的文件；target_dir：目标文件夹def reorg_train_valid(data_dir, labels, valid_ratio):"""将验证集从原始训练集钟拆分出来"""# 训练数据集中样本数量最少的类别中的样本数# Counter: 计数器，返回一个字典，键为元素，值为元素个数;# .most_common(): 返回一个列表， 列表元素为（元素，出现次数），默认按出现频率排序# [-1]: 样本数量最少的类别（类别， 样本数），[-1][1]: 样本数数量最少的类别中的样本数n = collections.Counter(labels.values()).most_common()[-1][1]# 验证集中每个类别的样本数n_valid_per_label= max(1, math.floor((n * valid_ratio)))# math.floor(): 向下取整math.ceil(): 向上取整label_count = {}# 遍历原始训练集中的每个样本for train_file in os.listdir(os.path.join(data_dir, 'train')):label = labels[train_file.split('.')[0]]# 从文件名中提取标签fname = os.path.join(data_dir, 'train', train_file)copyfile(fname, os.path.join(data_dir, 'train_valid_test', 'train_valid', label))# 如果该类别的样本数还未达到在验证集中的设定数量，则将样本复制到验证集中if label not in label_count or label_count[label] < n_valid_per_label:copyfile(fname, os.path.join(data_dir, 'train_valid_test', 'valid', label))label_count[label] = label_count.get(label, 0) + 1else:copyfile(fname, os.path.join(data_dir, 'train_valid_test', 'train', label))return n_valid_per_label# reorg_test函数用来在预测期间整理测试集，以方便读取def reorg_test(data_dir):"""在预测期间整理测试集，以方便读取"""# 遍历测试集中的每个样本for test_file in os.listdir(os.path.join(data_dir, 'test')):# 将测试集中的样本复制到新的目录结构中的 'test' 子目录下，标签为 'unknown'copyfile(os.path.join(data_dir, 'test', test_file), os.path.join(data_dir, 'train_valid_test', 'test', 'unknown'))# 整个处理数据集函数def reorg_cifar10_data(data_dir, valid_ratio):labels = read_csv_labels(os.path.join(data_dir, 'trainLabels.csv'))reorg_train_valid(data_dir, labels, valid_ratio)reorg_test(data_dir)- 这个小规模数据集的批量大小是32，在实际的cifar-10数据集中，可以设为128
- 将10%的训练样本作为调整超参数的验证集
batch_size = 32 if demo else 128valid_ratio = 0.1reorg_cifar10_data(data_dir, valid_ratio)结果会生成一个train_valid_test的文件夹，里面有：- test文件夹---unknow文件夹：5张没有标签的测试照片- train_valid文件夹---10个类被的文件夹：每个文件夹包含所属类别的全部照片- train文件夹--10个类别的文件夹：每个文件夹下包含90%的照片用于训练- valid文件夹--10个类别的文件夹：每个文件夹下包含10%的照片用于验证- 图像增广
transform_train = torchvision.transforms.Compose([# 原本图像是32*32，先放大成40*40， 在随机裁剪为32*32，实现训练数据的增强torchvision.transforms.Resize(40),torchvision.transforms.RandomResizedCrop(32, scale=(0.64, 1.0), ratio=(1.0, 1.0)),torchvision.transforms.RandomHorizontalFlip(),torchvision.transforms.ToTensor(),torchvision.transforms.Normalize([0.4914, 0.4822, 0.4465],[0.2023, 0.1994, 0.2010])])transform_test = torchvision.transforms.Compose([torchvision.transforms.ToTensor(),# 标准化图像的每个通道 ： 消除评估结果中的随机性torchvision.transforms.Normalize([0.4914, 0.4822, 0.4465],[0.2023, 0.1994, 0.2010])])- 加载数据集
train_ds, train_valid_ds = [torchvision.datasets.ImageFolder(os.path.join(data_dir, 'train_valid_test', folder),transform=transform_train) for folder in ['train', 'train_valid']]valid_ds, test_ds = [torchvision.datasets.ImageFolder(os.path.join(data_dir, 'train_valid_test', folder), transform=transform_test) for folder in ['valid', 'test']]- 定义迭代器，方便快速迭代数据
train_iter, train_valid_iter = [torch.utils.data.DataLoader(dataset, batch_size, shuffle=True, drop_last=True) for dataset in (train_ds, train_valid_ds)]valid_iter = torch.utils.data.DataLoader(valid_ds, batch_size, shuffle=False, drop_last=True)test_iter = torch.utils.data.DataLoader(test_ds, batch_size, shuffle=False, drop_last=False)- 定义模型与损失函数
# 对resnet18做微调，输入通道数为3， 输出类别数为10def get_net():num_classes = 10net = d2l.resnet18(num_classes, in_channels=3)return net# 查看网络模型get_net()
# 使用交叉熵损失函数作为损失函数: 直接返回n分样本的lossloss = nn.CrossEntropyLoss(reduction='none')- 定义训练函数
# 定义训练函数def train(net, train_iter, valid_iter, num_epochs, lr, wd, devices, lr_period, lr_decay):trainer = torch.optim.SGD(net.parameters(), lr=lr, momentum=0.9, weight_decay=wd)scheduler = torch.optim.lr_scheduler.StepLR(trainer, lr_period, lr_decay)num_batches, timer = len(train_iter), d2l.Timer()legend = ['train loss', 'train acc']if valid_iter is not None:legend.append('valid acc')animator = d2l.Animator(xlabel='epoch', xlim=[1, num_epochs], legend=legend)net = nn.DataParallel(net, device_ids=devices).to(devices[0])for epoch in range(num_epochs):net.train()metric = d2l.Accumulator(3)for i, (features, labels) in enumerate(train_iter):timer.start()l, acc = d2l.train_batch_ch13(net, features, labels, loss, trainer, devices)metric.add(l, acc, labels.shape[0])timer.stop()if (i + 1) % (num_batches // 5) == 0 or i == num_batches - 1:animator.add(epoch + (i + 1) / num_batches, (metric[0]/ metric[2], metric[1] / metric[2], None))if valid_iter is not None:valid_acc = d2l.evaluate_accuracy_gpu(net, valid_iter)animator.add(epoch+1, (None, None, valid_acc))scheduler.step()measures = (f'train loss {metric[0] / metric[2]:.3f},'f'train acc{metric[1] / metric[2]:.3f}')if valid_iter is not None:measures += f', valid acc {valid_acc:.3f}'print(measures + f'\n{metric[2] * num_epochs /timer.sum():.1f}' f'example/sec on {str(devices)}')- 训练模型
- （数据集太小，导致精度不高）
import time# 在开头设置开始时间start = time.perf_counter()# start = time.clock() python3.8之前可以# 训练和验证模型devices, num_epochs, lr, wd = d2l.try_all_gpus(), 20, 2e-4, 5e-4lr_period, lr_decay, net = 4, 0.9, get_net()train(net, train_iter, valid_iter, num_epochs, lr, wd, devices, lr_period, lr_decay)# 在程序运行结束的位置添加结束时间end = time.perf_counter()# end = time.clock()python3.8之前可以# 再将其进行打印，即可显示出程序完成的运行耗时print(f'运行耗时{(end-start):.4f}')
10. 对测试集进行分类并提交结果
net, preds = get_net(), []train(net ,train_valid_iter, None, num_epochs, lr, wd, devices, lr_period, lr_decay)for X, _ in test_iter:y_hat = net(X.to(devices[0]))preds.extend(y_hat.argmax(dim=1).type(torch.int32).cpu().numpy())sorted_ids = list(range(1, len(test_ds) + 1))sorted_ids.sort(key=lambda x: str(x))df = pd.DataFrame({'id' : sorted_ids, 'label': preds})df['label'] = df['label'].apply(lambda x: train_valid_ds.classes[x])df.to_csv('submission.csv', index=False)