Pytorch 中的数据增强(自学用)

介绍Pytorch中torchvision.transforms几个数据增强函数的使用方法:

from torchvision import transforms
from PIL import Image
from torchvision.transforms import functional as TF
import torch
path = 'D:\\片儿\\1910pt课程\\pract\\001.png'
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 在转换成 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
  present a list which contains the mean and std of operated images, by which will normalize all the images, the type of data is Tensor

mean = [0.5, 0.5, 0.5,0.5]
std = [0.5, 0.5, 0.5,0.5]

transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(mean, std),
    transforms.ToPILImage()# visualization
    
])
new_img = transform(img)
new_img

operations towards PIL data

• ToTensor
  turn PIL Image or numpy.ndarray into tensor

transform = transforms.Compose([
    transforms.ToTensor(),   
])
new_img = transform(img)

• Resize
  resize the input PIL Image to the given size.

parameters

• size: (x) width=height; (h,w) height, width
• interpolation: default: Bilinear

size = (100, 100)
transform = transforms.Compose([
    transforms.Resize(size),
])

new_img = transform(img)
new_img

• Scale
  abandoned, replaced by Resize

• CenterCrop
  Crops the given PIL Image at the center.

parameter:

• 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 = (400, 300)
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.

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 = (100, 200, 300, 400)
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(300, fill=(0, 255, 255)), transforms.CenterCrop(400), transforms.RandomRotation(45)]
transform = transforms.Compose([
    transforms.RandomApply(transform, p=0.8)
])

new_img = transform(img)
new_img

• RandomChoice

Apply single transformation randomly picked from a list

transform = [transforms.Pad(300, fill=(0, 255, 255)), transforms.CenterCrop(400), transforms.RandomRotation(45)]
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(300, fill=(0, 255, 255)), transforms.CenterCrop(400), transforms.RandomRotation(45)]
transform = transforms.Compose([
    transforms.RandomOrder(transform)
])

new_img = transform(img)
new_img

• RandomCrop

Crop the given PIL Image at a random location.

• 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.

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.

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)

import numpy as np
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: (758, 1256, 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

(758, 1256, 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