宜配屋

环境

系统：win10

cpu：i7-6700HQ

gpu：gtx965m

python : 3.6

pytorch ：0.3

数据下载

来源自Sasank Chilamkurthy 的教程；数据：下载链接。

下载后解压放到项目根目录：

数据集为用来分类蚂蚁和蜜蜂。有大约120个训练图像，每个类有75个验证图像。

数据导入

可以使用 torchvision.datasets.ImageFolder(root,transforms) 模块可以将图片转换为 tensor。

先定义transform：

ata_transforms = {
  'train': transforms.Compose([
    # 随机切成224x224 大小图片 统一图片格式
    transforms.RandomResizedCrop(224),
    # 图像翻转
    transforms.RandomHorizontalFlip(),
    # totensor 归一化(0,255) >> (0,1)  normalize  channel=（channel-mean）/std
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
  ]),
  "val" : transforms.Compose([
    # 图片大小缩放 统一图片格式
    transforms.Resize(256),
    # 以中心裁剪
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
  ])
}

导入，加载数据：

data_dir = './hymenoptera_data'
# trans data
image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x]) for x in ['train', 'val']}
# load data
data_loaders = {x: DataLoader(image_datasets[x], batch_size=BATCH_SIZE, shuffle=True) for x in ['train', 'val']}

data_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']}
class_names = image_datasets['train'].classes
print(data_sizes, class_names)

{'train': 244, 'val': 153} ['ants', 'bees']

训练集 244图片，测试集153图片。

可视化部分图片看看，由于visdom支持tensor输入，不用换成numpy，直接用tensor计算即可：

inputs, classes = next(iter(data_loaders['val']))

out = torchvision.utils.make_grid(inputs)
inp = torch.transpose(out, 0, 2)
mean = torch.FloatTensor([0.485, 0.456, 0.406])
std = torch.FloatTensor([0.229, 0.224, 0.225])
inp = std * inp + mean
inp = torch.transpose(inp, 0, 2)
viz.images(inp)

创建CNN

net 根据上一篇的处理cifar10的改了一下规格：

class CNN(nn.Module):
  def __init__(self, in_dim, n_class):
    super(CNN, self).__init__()
    self.cnn = nn.Sequential(
      nn.BatchNorm2d(in_dim),
      nn.ReLU(True),
      nn.Conv2d(in_dim, 16, 7), # 224 >> 218
      nn.BatchNorm2d(16),
      nn.ReLU(inplace=True),
      nn.MaxPool2d(2, 2), # 218 >> 109
      nn.ReLU(True),
      nn.Conv2d(16, 32, 5), # 105
      nn.BatchNorm2d(32),
      nn.ReLU(True),
      nn.Conv2d(32, 64, 5), # 101
      nn.BatchNorm2d(64),
      nn.ReLU(True),
      nn.Conv2d(64, 64, 3, 1, 1),
      nn.BatchNorm2d(64),
      nn.ReLU(True),
      nn.MaxPool2d(2, 2), # 101 >> 50
      nn.Conv2d(64, 128, 3, 1, 1), #
      nn.BatchNorm2d(128),
      nn.ReLU(True),
      nn.MaxPool2d(3), # 50 >> 16
    )
    self.fc = nn.Sequential(
      nn.Linear(128*16*16, 120),
      nn.BatchNorm1d(120),
      nn.ReLU(True),
      nn.Linear(120, n_class))
  def forward(self, x):
    out = self.cnn(x)
    out = self.fc(out.view(-1, 128*16*16))
    return out

# 输入3层rgb ，输出 分类 2    
model = CNN(3, 2)

loss，优化函数：

line = viz.line(Y=np.arange(10))
loss_f = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=LR, momentum=0.9)
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=7, gamma=0.1)

参数：

BATCH_SIZE = 4
LR = 0.001
EPOCHS = 10

运行 10个 epoch 看看：

[9/10] train_loss:0.650|train_acc:0.639|test_loss:0.621|test_acc0.706
[10/10] train_loss:0.645|train_acc:0.627|test_loss:0.654|test_acc0.686
Training complete in 1m 16s
Best val Acc: 0.712418

运行 20个看看：

[19/20] train_loss:0.592|train_acc:0.701|test_loss:0.563|test_acc0.712
[20/20] train_loss:0.564|train_acc:0.721|test_loss:0.571|test_acc0.706
Training complete in 2m 30s
Best val Acc: 0.745098

准确率比较低：只有74.5%

我们使用models 里的 resnet18 运行 10个epoch：

model = torchvision.models.resnet18(True)
num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, 2)

[9/10] train_loss:0.621|train_acc:0.652|test_loss:0.588|test_acc0.667
[10/10] train_loss:0.610|train_acc:0.680|test_loss:0.561|test_acc0.667
Training complete in 1m 24s
Best val Acc: 0.686275

效果也很一般，想要短时间内就训练出效果很好的models，我们可以下载训练好的state，在此基础上训练：

model = torchvision.models.resnet18(pretrained=True)
num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, 2)

[9/10] train_loss:0.308|train_acc:0.877|test_loss:0.160|test_acc0.941
[10/10] train_loss:0.267|train_acc:0.885|test_loss:0.148|test_acc0.954
Training complete in 1m 25s
Best val Acc: 0.954248

10个epoch直接的到95%的准确率。