结合代码和说明,带你理解 Pytorch 中的数据增强操作
from torchvision import transforms
from PIL import Image
from torchvision.transforms import functional as TF
import torch
所有的变换
[“Compose”, “ToTensor”, “ToPILImage”, “Normalize”, “Resize”, “Scale”, “CenterCrop”, “Pad”,
“Lambda”, “RandomApply”, “RandomChoice”, “RandomOrder”, “RandomCrop”, “RandomHorizontalFlip”,
“RandomVerticalFlip”, “RandomResizedCrop”, “RandomSizedCrop”, “FiveCrop”, “TenCrop”, “LinearTransformation”,
“ColorJitter”, “RandomRotation”, “RandomAffine”, “Grayscale”, “RandomGrayscale”,
“RandomPerspective”]
# 读取一张测试图片
path = "F:/jupyter/OpenCV-Python-Tutorial/Tutorial/sample_img/lena.jpg"
img = Image.open(path)
img
对 Torch 数据操作的变换
ToPILImage
Convert a tensor or an ndarray to PIL Image.
# 将 ``PIL Image`` or ``numpy.ndarray`` 转换成 tensor 再转成 PIL Image.
transform = transforms.Compose([
transforms.ToTensor(), # Convert a ``PIL Image`` or ``numpy.ndarray`` to tensor.
transforms.ToPILImage() # Convert a tensor or an ndarray to PIL Image.
])
new_img = transform(img)
new_img
Normalize
提供一个所有通道的均值(mean) 和方差(std),会将原始数据进行归一化,操作的数据格式是 Tensor
mean = [0.5, 0.5, 0.5]
std = [0.5, 0.5, 0.5]
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean, std),
transforms.ToPILImage() # 这里是为了可视化,故将其再转为 PIL,以下同理
])
new_img = transform(img)
new_img
对 PIL 数据操作的变换
ToTensor
将 PIL Image 或者 numpy.ndarray 格式的数据转换成 tensor
transform = transforms.Compose([
transforms.ToTensor(), # Convert a ``PIL Image`` or ``numpy.ndarray`` to tensor.
])
new_img = transform(img)
Resize
Resize the input PIL Image to the given size.
参数
- size: 一个值的话,高和宽共享,否则对应是 (h, w)
- interpolation: 插值方式 默认 PIL.Image.BILINEAR
size = (100, 100)
transform = transforms.Compose([
transforms.Resize(size),
])
new_img = transform(img)
new_img
Scale
这个已经废弃,由 Resize 替代
CenterCrop
Crops the given PIL Image at the center.
裁剪一定 size 的图片,以图片的中心往外
参数
- size: 一个值的话,高和宽共享,否则对应是 (h, w),若是该值超过原始图片尺寸,则外围用 0 填充
size = (200, 500)
transform = transforms.Compose([
transforms.CenterCrop(size),
])
new_img = transform(img)
new_img
Pad
Pad the given PIL Image on all sides with the given “pad” value. 填充图片的外部轮廓 PIL 数据格式
参数
- padding:填充的宽度,可以是一个 值、或者元组,分别对应 4 个边
- fill:填充的值,可以是一个值(所有通道都用该值填充),或者一个 3 元组(RGB 三通道) 当 padding_mode=constant 起作用
- padding_mode:填充的模式:constant, edge(填充值为边缘), reflect(从边缘往内一个像素开始做镜像) or symmetric(从边缘做镜像).
padding = (10, 20, 30, 40)
transform = transforms.Compose([
transforms.Pad(padding, padding_mode="symmetric"),
])
new_img = transform(img)
new_img
Lambda
根据用户自定义的方式进行变换
lambd = lambda x: TF.rotate(x, 100)
transform = transforms.Compose([
transforms.Lambda(lambd)
])
new_img = transform(img)
new_img
RandomApply
给定一定概率从一组 transformations 应用
transform = [transforms.Pad(100, fill=(0, 255, 255)), transforms.CenterCrop(100), transforms.RandomRotation(20)]
transform = transforms.Compose([
transforms.RandomApply(transform, p=0.5)
])
new_img = transform(img)
new_img
RandomChoice
Apply single transformation randomly picked from a list
随机从一组变换中选择一个
transform = [transforms.Pad(100, fill=(0, 255, 255)), transforms.CenterCrop((100, 300))]
transform = transforms.Compose([
transforms.RandomChoice(transform)
])
new_img = transform(img)
new_img
RandomOrder
Apply a list of transformations in a random order
transform = [transforms.Pad(100, fill=(0, 255, 255)), transforms.CenterCrop((50, 50))]
transform = transforms.Compose([
transforms.RandomOrder(transform)
])
new_img = transform(img)
new_img
RandomCrop
Crop the given PIL Image at a random location. 随机进行裁剪
参数:
- size
- padding=None
- pad_if_needed=False
- fill=0
- padding_mode=‘constant’
transform = transforms.Compose([
transforms.RandomCrop((100, 300))
])
new_img = transform(img)
new_img
RandomHorizontalFlip & RandomVerticalFlip
Horizontally/Vertically flip the given PIL Image randomly with a given probability. 按一定概率进行水平/竖直翻转
transform = transforms.Compose([
transforms.RandomHorizontalFlip(p=0.5),
transforms.RandomVerticalFlip(p=0.5)
])
new_img = transform(img)
new_img
RandomResizedCrop
Crop the given PIL Image to random size and aspect ratio. 裁剪给定的 PIL 图像到随机的尺寸和长宽比。
参数:
- size: expected output size of each edge
- scale: range of size of the origin size cropped
- ratio: range of aspect ratio of the origin aspect ratio cropped
- interpolation: Default: PIL.Image.BILINEAR
transform = transforms.Compose([
transforms.RandomResizedCrop((200, 300))
])
new_img = transform(img)
new_img
RandomSizedCrop
已经废弃,由 RandomResizedCrop. 取代了
FiveCrop
将给定的 PIL 图像裁剪成四个角和中间的裁剪
UNIT_SIZE = 200 # 每张图片的宽度是固定的
size = (100, UNIT_SIZE)
transform = transforms.Compose([
transforms.FiveCrop(size)
])
new_img = transform(img)
delta = 20 # 偏移量,几个图片间隔看起来比较明显
new_img_2 = Image.new("RGB", (UNIT_SIZE*5+delta, 100))
top_right = 0
for im in new_img:
new_img_2.paste(im, (top_right, 0)) # 将image复制到target的指定位置中
top_right += UNIT_SIZE + int(delta/5) # 左上角的坐标,因为是横向的图片,所以只需要 x 轴的值变化就行
new_img_2
TenCrop
裁剪一张图片的 4 个角以及中间得到指定大小的图片,并且进行水平翻转/竖直翻转 共 10 张
参数:
- size
- vertical_flip=False (默认是水平翻转)
UNIT_SIZE = 200 # 每张图片的宽度是固定的
size = (100, UNIT_SIZE)
transform = transforms.Compose([
transforms.TenCrop(size, vertical_flip=True)
])
new_img = transform(img)
delta = 50 # 偏移量,几个图片间隔看起来比较明显
new_img_2 = Image.new("RGB", (UNIT_SIZE*10+delta, 100))
top_right = 0
for im in new_img:
new_img_2.paste(im, (top_right, 0)) # 将image复制到target的指定位置中
top_right += UNIT_SIZE + int(delta/10) # 左上角的坐标,因为是横向的图片,所以只需要 x 轴的值变化就行
new_img_2
LinearTransformation
白化变换,笔者不是很理解,但是好像消耗的内存应该比较大
ColorJitter
Randomly change the brightness, contrast and saturation of an image. 随机改变图像的亮度、对比度和饱和度
参数:
- brightness:亮度
- contrast:对比度
- saturation:饱和度
- hue:色调 0<= hue <= 0.5 or -0.5 <= min <= max <= 0.5.
transform = transforms.Compose([
transforms.ColorJitter(brightness=(0, 36), contrast=(
0, 10), saturation=(0, 25), hue=(-0.5, 0.5))
])
new_img = transform(img)
new_img
RandomRotation
一定角度旋转图像
参数:
- degrees:旋转的角度
- resample=False:重采样过滤器 可选{PIL.Image.NEAREST, PIL.Image.BILINEAR, PIL.Image.BICUBIC}
- expand=False:如果为 True ,则展开输出,使其足够大以容纳整个旋转后的图像。 如果为 Fales 或省略,使输出图像的大小与输入图像相同。
- center=None 旋转中心
transform = transforms.Compose([
transforms.RandomRotation(30, resample=Image.BICUBIC, expand=False, center=(100, 300))
])
new_img = transform(img)
new_img
RandomAffine
保持图像中心不变的随机仿射变换,可以进行随心所欲的变化
参数:
- degrees:旋转角度
- translate:水平偏移
- scale:
- shear: 裁剪
- resample ({PIL.Image.NEAREST, PIL.Image.BILINEAR, PIL.Image.BICUBIC}, optional)
- fillcolor:图像外部填充颜色 int
transform = transforms.Compose([
transforms.RandomAffine(degrees=30, translate=(0, 0.2), scale=(0.9, 1), shear=(6, 9), fillcolor=66)
])
new_img = transform(img)
new_img
Grayscale
转换图像灰度。
参数:
- num_output_channels:1 或者 3 输出图像所需的通道数(若是 3 的话,则代表三个通道的值是一样的)
transform = transforms.Compose([
transforms.Grayscale(num_output_channels=3)
])
new_img = transform(img)
new_img_array = np.array(new_img)
r, g, b = new_img_array[:, :, 0], new_img_array[:, :, 1], new_img_array[:, :, 2]
print(r == b)
print("shape:", new_img_array.shape)
new_img
[[ True True True ... True True True]
[ True True True ... True True True]
[ True True True ... True True True]
...
[ True True True ... True True True]
[ True True True ... True True True]
[ True True True ... True True True]]
shape: (400, 400, 3)
RandomGrayscale
Randomly convert image to grayscale with a probability of p (default 0.1). 以一定的概率对图像进行灰度化,转换后的图片还是 3 通道的
transform = transforms.Compose([
transforms.RandomGrayscale(p=0.6)
])
new_img = transform(img)
print(np.array(new_img).shape)
new_img
(400, 400, 3)
RandomPerspective
对给定的PIL图像以给定的概率随机进行透视变换。
transform = transforms.Compose([
transforms.RandomPerspective(distortion_scale=1, p=1, interpolation=3)
])
new_img = transform(img)
new_img
其他
transforms.Compose
函数是将几个变化整合在一起的,变换是有顺序的,需要注意是变换函数是对 PIL 数据格式进行还是 Torch 数据格式进行变换