不吹不黑一份代码即可进 Kaggle 排行榜！
发布于: 7 小时前
大家好，我是 cv 君，今天分享一份源码，有了这份代码，你就是一个 Kaggle 的一个图像分类赛排行榜因为这是很久以前的比赛，不知道还开不开榜，比如选择你还可以去刷 Kaggle 的 Mnist 排行榜 99.93+分以上就能进入排行榜（没记错的话）这个 Mnist，每个 AI 的选手都做过，大家可以看我专栏其他文章，是进 Mnist 排行榜的代码；
我们今天不做 Mnist 了，我们做几年前，特别著名的 cat-dog 分类！这份代码，你可以得到 99.3 以上的准确率，你想知道当时这个比赛进排行榜多少分吗？98.9！我们做了什么呢？
迁移学习 模型集成 数据增强 调优调参 优化提分 冻结网络欢迎来到动手撸深度学习~cv 君力挺的，最值得看的专栏系列：动手撸深度学习学习完本系列，你能从调包侠到一个能独立打比赛的朋友周边的神!文章付费，不想让太多人白嫖！！！因为全部原创，很珍贵!!看完你会感谢我的~感兴趣的话，就来看看吧~首先，大家通过上两节，我们已经通过 Keras 搭建了自己的神经网络，并且学习了搭建方法，数据增强方法，Dropout 等的使用方法和原理，并且学习了迁移学习的几种方式，并且实现了得到了 98.5 以上的准确率，我们距离排行榜还有一些些距离。这一节我们开始：迁移学习 模型集成 数据增强 调优调参 优化提分 冻结网络
现在来看代码：老规矩，导入库
model_resnet18 = torch.hub.load('pytorch/vision', 'resnet18', pretrained=True)model_resnet34 = torch.hub.load('pytorch/vision', 'resnet34', pretrained=True)复制代码
for name, param in model_resnet18.named_parameters():if ("bn" not in name):param.requires_grad = False
for name, param in model_resnet34.named_parameters():if ("bn" not in name):param.requires_grad = False
训练分类器权重复制代码
num_classes = 2
model_resnet18.fc = nn.Sequential(nn.Linear(model_resnet18.fc.in_features, 512),nn.ReLU(),nn.Dropout(),nn.Linear(512, num_classes))
model_resnet34.fc = nn.Sequential(nn.Linear(model_resnet34.fc.in_features, 512),nn.ReLU(),nn.Dropout(),nn.Linear(512, num_classes))
123456789101112定义训练
def train(model, optimizer, loss_fn, train_loader, val_loader, epochs=2, device="cpu"):for epoch in range(epochs):training_loss = 0.0valid_loss = 0.0model.train()for batch in train_loader:optimizer.zero_grad()inputs, targets = batchinputs = inputs.to(device)targets = targets.to(device)output = model(inputs)loss = loss_fn(output, targets)loss.backward()optimizer.step()training_loss += loss.data.item() * inputs.size(0)print(loss)training_loss /= len(train_loader.dataset)
复制代码
model.eval() num_correct = 0 num_examples = 0 for batch in val_loader: inputs, targets = batch inputs = inputs.to(device) output = model(inputs) targets = targets.to(device) optimizer.zero_grad()
loss = loss_fn(output, targets) valid_loss += loss.data.item() * inputs.size(0)
correct = torch.eq(torch.max(F.softmax(output, dim=1), dim=1)[1], targets).view(-1) num_correct += torch.sum(correct).item() num_examples += correct.shape[0] valid_loss /= len(val_loader.dataset)
print( 'Epoch: {}, Training Loss: {:.4f}, Validation Loss: {:.4f}, accuracy = {:.4f}'.format(epoch, training_loss,
复制代码
数据读取
batch_size=40img_dimensions = 224
Normalize to the ImageNet mean and standard deviation
Could calculate it for the cats/dogs data set, but the ImageNet
values give acceptable results here.
img_transforms = transforms.Compose([transforms.Resize((img_dimensions, img_dimensions)),transforms.ToTensor(),transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])])
img_test_transforms = transforms.Compose([transforms.Resize((img_dimensions,img_dimensions)),transforms.ToTensor(),transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])])
train_data_path = "train_total/"train_data = torchvision.datasets.ImageFolder(root=train_data_path,transform=img_transforms, is_valid_file=check_image)
validation_data_path = "val_total/"validation_data = torchvision.datasets.ImageFolder(root=validation_data_path,transform=img_test_transforms, is_valid_file=check_image)
test_data_path = "val_total/"test_data = torchvision.datasets.ImageFolder(root=test_data_path, transform=img_test_transforms, is_valid_file=check_image)
num_workers = 0train_data_loader = torch.utils.data.DataLoader(train_data, batch_size=batch_size, shuffle=True, num_workers=num_workers)validation_data_loader = torch.utils.data.DataLoader(validation_data, batch_size=batch_size, shuffle=False, num_workers=num_workers)test_data_loader = torch.utils.data.DataLoader(test_data, batch_size=batch_size, shuffle=False, num_workers=num_workers)
复制代码
测试
def test_model(model):correct = 0total = 0with torch.no_grad():for data in test_data_loader:images, labels = data[0].to(device), data[1].to(device)outputs = model(images)_, predicted = torch.max(outputs.data, 1)total += labels.size(0)correct += (predicted == labels).sum().item()print('correct: {:d} total: {:d}'.format(correct, total))print('accuracy = {:f}'.format(correct / total))复制代码
训练部分
model_resnet18.to(device)optimizer = optim.Adam(model_resnet18.parameters(), lr=0.001)train(model_resnet18, optimizer, torch.nn.CrossEntropyLoss(), train_data_loader, validation_data_loader, epochs=2, device=device)test_model(model_resnet18)
model_resnet34.to(device)optimizer = optim.Adam(model_resnet34.parameters(), lr=0.001)train(model_resnet34, optimizer, torch.nn.CrossEntropyLoss(), train_data_loader, validation_data_loader, epochs=2, device=device)test_model(model_resnet34)
torch.save(model_resnet18.state_dict(), "./models/model_resnet18.pth")torch.save(model_resnet34.state_dict(), "./models/model_resnet34.pth")
复制代码
训练完毕 可以实际验证结果啦~
resnet18 = torch.hub.load('pytorch/vision', 'resnet18')resnet18.fc = nn.Sequential(nn.Linear(resnet18.fc.in_features, 512), nn.ReLU(), nn.Dropout(), nn.Linear(512, num_classes))resnet18.load_state_dict(torch.load('./models/model_resnet18.pth'))resnet18.eval()
resnet34 = torch.hub.load('pytorch/vision', 'resnet34')resnet34.fc = nn.Sequential(nn.Linear(resnet34.fc.in_features, 512), nn.ReLU(), nn.Dropout(), nn.Linear(512, num_classes))resnet34.load_state_dict(torch.load('./models/model_resnet34.pth'))resnet34.to(device)resnet34.eval()def predict():import osdef find_classes(dir):classes = os.listdir(dir)classes.sort()class_to_idx = {classes[i]: i for i in range(len(classes))}return classes, class_to_idx
复制代码
def make_prediction(model, filename): labels, _ = find_classes('test1/') img = Image.open(filename) img = img_test_transforms(img) img = img.unsqueeze(0) prediction = model(img.to(device)) prediction = prediction.argmax() print(prediction)
make_prediction(model_resnet34.to(device), 'test1/14.jpg')make_prediction(model_resnet34.to(device), 'test1/10.jpg')def val():models_ensemble = [resnet18.to(device), resnet34.to(device)] correct = 0total = 0with torch.no_grad():for data in test_data_loader:images, labels = data[0].to(device), data[1].to(device)predictions = [i(images).data for i in models_ensemble]avg_predictions = torch.mean(torch.stack(predictions), dim=0)_, predicted = torch.max(avg_predictions, 1)复制代码
total += labels.size(0) correct += (predicted == labels).sum().item()
print('accuracy = {:f}'.format(correct / total))print('correct: {:d} total: {:d}'.format(correct, total))val()predict()复制代码
测试时集成模型
models_ensemble = [resnet18.to(device), resnet34.to(device)]correct = 0total = 0with torch.no_grad():for data in test_data_loader:images, labels = data[0].to(device), data[1].to(device)predictions = [i(images).data for i in models_ensemble]avg_predictions = torch.mean(torch.stack(predictions), dim=0)_, predicted = torch.max(avg_predictions, 1)复制代码
total += labels.size(0) correct += (predicted == labels).sum().item()
复制代码
大家可以看到模型集成地方式，当然这只是一种集成方法
何谓模型集成即通过训练多个模型，对带预测数据进行拟合（训练时集成）和验证时集成（测试时多模型投票或者平均或者比例方式）
以上通过 10 折交叉验证，可训练得到 10 个 CNN 模型，集成方法有：
平均法：将 10 个模型预测结果的概率取平均值，然后解码为具体字符投票法：对 10 个模型预测结果进行投票，得到最终字符 12 各个代码，函数的具体意思，请看我的视频~视频很详细地介绍了。视频地址：链接：https://pan.baidu.com/s/1teabC-TIfc-8MPVcV_Q76g提取码：deep 一定要看看嗷，这里讲的清晰一些。
划线
评论
复制
发布于: 7 小时前阅读数: 4
cv君
关注
还未添加个人签名 2021.03.22 加入
还未添加个人简介
评论