数据增强2

Transform的剩余方法

2020.7.6修改，补全了对应的测试代码

1. Pad
   对图片进行填充
   transform.Pad(padding, fill=0, padding_mode='contant')

• padding设置填充大小
  参数三种方式(a),(a,b),(a,b,c,d)分别为四面，上下和左右，以及上下左右
• padding)mode：填充模式，有四种，constant,edge,reflect和symmetric
• fill：Contant时，设置填充的像素值(R,G,B)or(Gray)

import torch
import torchvision
from torchvision.transforms import Compose, ToTensor, Pad, ColorJitter, RandomAffine, RandomErasing
from PIL import Image
import matplotlib.pyplot as plt
import numpy as np

unloader = torchvision.transforms.ToPILImage()

transforms1 = Compose([Pad(100, fill=(0, 0, 0)),
                      ToTensor()])

transforms2 = Compose([Pad((100,50), fill=(0, 0, 0)),
                      ToTensor()])

transforms3 = Compose([Pad((100,200,300,50), fill=(0, 255, 0)),
                      ToTensor()])

image = Image.open("lena.png").convert("RGB")

plt.subplot(2,2,1)
plt.imshow(image)
plt.subplot(2,2,2)
img=transforms1(image)
img=unloader(img)
plt.grid()
plt.imshow(img)
plt.subplot(2,2,3)
img=transforms2(image)
img=unloader(img)
plt.grid()
plt.imshow(img)
plt.subplot(2,2,4)
img=transforms3(image)
img=np.transpose(img.numpy(),(1,2,0))
plt.imshow(img)
plt.grid()
plt.savefig("out.png",dpi=300)
plt.show()

2.ColorJitter
调整亮度、对比度、饱和度和色相

transforms.ColorJitter(brightness=0,contrast=0,saturation=0,hue=0)

• brightness:亮度调整因子
  当为a时，从[max(0,1-a),a+1]中随机选择，当为(a,b)时，从[a,b]中选择

• contrast :对比度参数，同brightness

• saturation：饱和度参数，同brightness

• hue:色相参数，当为a时，从[-a,a]中选择参数，注：0<=a<=0.5,当为(a,b)时从[a,b]中选择，注：-0.5<=a<=b<=0.5

import torch
from torchvision.utils import make_grid
import torchvision
from torchvision.transforms import Compose, ToTensor, Pad, ColorJitter, RandomAffine, RandomErasing
from PIL import Image
import matplotlib.pyplot as plt
import numpy as np

unloader = torchvision.transforms.ToPILImage()
tensor = ToTensor()
transforms = Compose([ColorJitter(brightness=0.8),
                      ToTensor()])
transforms1=Compose([ColorJitter(contrast=0.8),
                     ToTensor()])
transforms2=Compose([ColorJitter(saturation=0.8),
                     ToTensor()])
transforms3=Compose([ColorJitter(hue=0.4),
                     ToTensor()])
transforms4=Compose([ColorJitter(brightness=0.8,saturation=0.8,contrast=0.8,hue=0.4),
                     ToTensor()])
image = Image.open("lena.png").convert("RGB")
img = tensor(image)
imgs=list()
imgs1=list()
imgs2=list()
imgs3=[]
imgs4=[]
for i in range(9):
    imgT = transforms(image)
    imgT1 = transforms1(image)
    imgT2 = transforms2(image)
    imgT3 = transforms3(image)
    imgT4 = transforms4(image)
    imgs.append(imgT)
    imgs1.append(imgT1)
    imgs2.append(imgT2)
    imgs3.append(imgT3)
    imgs4.append(imgT4)
img=torch.stack(imgs,dim=0)
img = make_grid(img, nrow=3,padding=5,normalize=True,scale_each=True)
img = unloader(img)
plt.imshow(img)
plt.show()

img=torch.stack(imgs1,dim=0)
img = make_grid(img, nrow=3,padding=5,normalize=True,scale_each=True)
img = unloader(img)
plt.imshow(img)
plt.show()

img=torch.stack(imgs2,dim=0)
img = make_grid(img, nrow=3,padding=5,normalize=True,scale_each=True)
img = unloader(img)
plt.imshow(img)
plt.show()

img=torch.stack(imgs3,dim=0)
img = make_grid(img, nrow=3,padding=5,normalize=True,scale_each=True)
img = unloader(img)
plt.imshow(img)
plt.show()

img=torch.stack(imgs4,dim=0)
img = make_grid(img, nrow=3,padding=5,normalize=True,scale_each=True)
img = unloader(img)
plt.imshow(img)
plt.show()

3.Grayscalea和RandomGrayscale
依据概率把图片转变成灰度图

• num_output_channels 输出的通道数
• p:转换的概率

import torch
from torchvision.utils import make_grid
import torchvision
from torchvision.transforms import Compose, ToTensor, Pad, ColorJitter,RandomGrayscale, Grayscale,RandomAffine, RandomErasing
from PIL import Image
import matplotlib.pyplot as plt
import numpy as np

unloader = torchvision.transforms.ToPILImage()
tensor = ToTensor()
transforms = Compose([Grayscale(num_output_channels=3),
                      ToTensor()])

transforms1=Compose([RandomGrayscale(p=0.4),
                     ToTensor()])
image=Image.open('lena.png').convert('RGB')
imgs=[]
for i in range(4):
    if i==0:
        img=transforms(image)
        imgs.append(img)

    else:
        imgs.append(transforms1(image))
out=torch.stack(imgs)
out_img=make_grid(out,nrow=2,normalize=True,scale_each=True)
out_img=np.transpose(out_img.numpy(),(1,2,0))
plt.imshow(out_img)
plt.show()

4.RandomAffine
对图像进行仿射变换，仿射变换是二维的线性变换，由五种基本原子变换构成，分别是旋转、平移、缩放、错切和翻转

• degrees:旋转角度设置
• translate:平移区间设置，如(a,b),a设置宽(width),b设置高(height)所图像在宽纬度平移的区间为-img_widtha<dx<img_widtha
• scale:缩放比例(以面积为单位)
• fill_color:填充颜色设置
• shear:错切角度设置，有水平错切和垂直错切，若为a,则仅在x轴错切，错切角度在(-a,a)之间，若为(a,b)则a设置x轴角度，b设置为y轴角度，若为(a,b,c,d)则a,b,设置为x角度，cd设置为y轴角度
• resample:重采样方法，有NEAREST,BILINEAR,BICUBIC三种模式

5.RandomErasing

对图片随机遮挡

• p:概率值，执行该操作的概率
• scale:遮挡区域的面积
• ratio遮挡区域的长宽比
• 设置遮挡区域的像素值，(RGB)or(GRAY)
  
  此操作需要在张量前提下进行操作

transforms.ToTensor()
transforms.RandomErasing(p=1,scale=(0.02,0.33),ratio=(0.3,3.3),value=(254,255,0,0))

6.lambda
用户自定义lambda方法

transforms.TenCrop(200,vertical_flip=True)
transforms.Lambda(lambda crops:torch.stack([transforms.ToTensor()(crop) for crop in crops]))

对Transforms的选择化操作

自定义transforms方法

定义的要素：

• 仅接受一个参数，返回一个参数
• 注意上下游的输入与输出