torchvision指南

本筆記引用自PyTorch中文文檔

1. torchvision介紹

torchvision包含了目前流行的數(shù)據(jù)集，模型結(jié)構(gòu)和常用的圖片轉(zhuǎn)換工具愕贡。

1.1 torchvision.datasets

• torchvision.datasets中包含以下數(shù)據(jù)集：
  • MNIST
  • COCO
  • LSUN Classification
  • ImageFolder
  • Imagenet-12
  • CIFAR10 and CIFAR100
  • STL10
• dataset擁有以下API：__getitem__, __len__
• 這些datasets都是torch.utils.data.Dataset的子類

1.1.1 MNIST

• dest.MNIST(root, train=True, transform=None, target_transform=None, download=False)
  -download: 是否聯(lián)網(wǎng)下載

1.1.2 COCO

• 圖像標(biāo)注
• dest.CocoCaptions(root="dir where images are", annFile="json annotation file", [transform, target_transform, transform])

# sample
import torchvision.datasets as dset
import torchvision.transforms as transforms
cap = dset.CocoCaptions(root = 'dir where images are',
          annFile = 'json annotation file',
          transform=transforms.ToTensor())
print(Number of sample: ', len(cap))
img, target = cap[3]
print('Image size:', img.size())
print(target)

'''
output
Number of samples: 82783
Image Size: (3L, 427L, 640L)
[u'A plane emitting smoke stream flying over a mountain.',
u'A plane darts across a bright blue sky behind a mountain covered in snow',
u'A plane leaves a contrail above the snowy mountain top.',
u'A mountain that has a plane flying overheard in the distance.',
u'A mountain view with a plume of smoke in the background']
'''

• 檢測
• `dset.CocoDetection(root='dir where images are', annFile='json annotation file', [transform, target_transform])

1.1.3 ImageFolder

• 一個通用的數(shù)據(jù)加載器孵运，數(shù)據(jù)集中數(shù)據(jù)以以下方式組織：

root/dog/xxx.png
root/dog/xxy.png
root/dog/xxz.png

root/cat/123.png
root/cat/nsdf3.png
root/cat/asd932_.png

• dset.ImageFolder(root='root folder path', [transform, target_transform])
  • self.classes: 用一個list保存類名
  • self.class_to_idx: 類名對應(yīng)的索引
  • self.imgs: 保存(img_path, class)tuple的list

1.1.4 CIFAR

• dset.CIFAR10(root, train=True, transform=None, target_transform=None, download=False)
• dset.CIFAR100(root, train=True, transform=None, target_transform=None, download=False)

1.2 torchvision.models

• torchvision.models模塊的子模塊中包含以下模型結(jié)構(gòu)：
  • AlexNet
  • VGG
  • ResNet
  • SqueezeNet
  • DenseNet
• 可使用預(yù)訓(xùn)練模型螟够，如：

torchvision.models.alexnet(pretrained=False, **kwargs)
torchvision.models.resnet18(pretrained=False, **kwargs)
torchvision.models.resnet34(pretrained=False, ** kwargs)
torchvision.models.resnet50(pretrained=False, ** kwargs)
torchvision.models.resnet101(pretrained=False, ** kwargs)
torchvision.models.resnet152(pretrained=False, ** kwargs)
torchvision.models.vgg11(pretrained=False, ** kwargs)
torchvision.models.vgg11_bn(** kwargs)
torchvision.models.vgg13(pretrained=False, ** kwargs)
torchvision.models.vgg13_bn(** kwargs)
torchvision.models.vgg16(pretrained=False, ** kwargs)
torchvision.models.vgg16_bn(** kwargs)
torchvision.models.vgg19(pretrained=False, ** kwargs)
torchvision.models.vgg19_bn(** kwargs)

1.3 torchvision.transforms

• 對PIL.Image進(jìn)行變換
• 使用torchvision.transforms.Compose(transforms)將多個transform組合起來使用
• transforms.CenterCrop(size)：將給定PIL.Image進(jìn)行中心切割彤断，得到給定的size怕品，size可以是tuple或Integer
• transforms.RandomCrop(size, padding=0)：切割中心點(diǎn)位置隨機(jī)選取妇垢，size可以是tuple或Integer
• transforms.RandomHorizontalFlip(p=0.5)：隨機(jī)水平翻轉(zhuǎn)
• transforms.RandomSizedCrop(size, interpolation=2)：先隨機(jī)切，再resize成給定size大小肉康。
• transforms.Pad(padding, fill=0)：給所有邊用給定的值填充闯估。padding：要填充多少像素
• transforms.ToTensor()：將一個取值范圍是[0, 255]的PIL.Image或shape為(H, W, C)的numpy.ndarray，轉(zhuǎn)換成形狀為[C, H, W]吼和，取值范圍是[0, 1.0]的torch.FloatTensor
• transforms.Normalize(mean, std)：給定均值與方差睬愤，正則化，即Normalized_image=(image-mean)/std
• 通用變換：使用lambda作為轉(zhuǎn)換器纹安，transforms.Lambda(lambda)

1.4 torchvision.utils

• utils.make_grid(tensor, nrow=8, padding=2, normalize=False, range=None, scale_each=False)
  給定4D-mini-batch Tensor尤辱，形狀為(B*C*H*W)，或者一個a list of image厢岂，做成一個size為(B / nrow, nrow)的子圖集
  • normalize=True, 對圖像像素歸一化
  • range=(min, max)光督，min和max是數(shù)字，則min, max用來規(guī)范化image
  • scale_each=True, 每個圖片獨(dú)立規(guī)范化塔粒。
• utils.save_image(tensor, filename, nrow=8, padding=2, normalize=False, range=None, scale_each=False)
  將給定的Tensor保存成image文件结借，如果是mini-batch tensor，就用make-grid做成子圖集再保存卒茬。