たった1行でデータ拡張！torchvision のコードまとめ！

# モジュールのインポート
import matplotlib.pyplot as plt
from torchvision import transforms as transforms
import torch
import torchvision
import numpy as np

# 描画用の関数(チャンネル数の関係で、グレースケール画像とカラー画像で表示を分けています！)
def show_image(x):
  fig = plt.figure(figsize=(10,10))

  for s in range(len(x)):
      img = x[s]
      npimg = img.numpy()
      if npimg.shape[0]==1 :  # グレースケール用
        npimg = npimg.reshape(32, 32)
        ax1 = fig.add_subplot(1, len(x), s+1)
        plt.axis('off')
        plt.imshow(npimg, cmap='gray')
      else :  # カラー用
        npimg = np.transpose(npimg, (1, 2, 0))
        ax1 = fig.add_subplot(1, len(x), s+1)
        plt.axis('off')
        plt.imshow(npimg)

# CIFAR10 をダウンロード
cifar10set = torchvision.datasets.CIFAR10(root='./data', train=True, download=True, transform=transforms.ToTensor())
cifar10loader = torch.utils.data.DataLoader(cifar10set, batch_size=3, shuffle=False)

# 変換前の画像の描画
iter_ = iter(cifar10loader)
x, t = next(iter_)
x, t = next(iter_)  # <- 2回繰り返してるのはカエルの画像が嫌だからです！

show_image(x)

transform = transforms.Resize(size=(100, 100))
transed_x = transform(x)

show_image(transed_x)

transform = transforms.CenterCrop(size=(20, 20))
transed_x = transform(x)

show_image(transed_x)

transform = transforms.FiveCrop(size=(20, 20))
transed_x = transform(x)

for i in range(5): # 4隅+中央で5セット出てくるので, 5回ループ
  show_image(transed_x[i])

transform = transforms.TenCrop(size=(20, 20))
transed_x = transform(x)

for i in range(10):# 4隅+中央で5セット * 画像2枚 = 10セット出てくるので, 10回ループ
  show_image(transed_x[i])

transform = transforms.RandomCrop(size=(20, 20))
transed_x = transform(x)

show_image(transed_x)

transform = transforms.RandomResizedCrop(size=(40, 40), scale=(0.08, 1.0), ratio=(3 / 4, 4 / 3))
transed_x = transform(x)

show_image(transed_x)

transform = transforms.Pad(padding=3, padding_mode="constant", fill=0)
transed_x = transform(x)

show_image(transed_x)

transform = transforms.Pad(padding=3, padding_mode="edge")
transed_x = transform(x)

show_image(transed_x)

transform = transforms.Pad(padding=3, padding_mode="reflect")
transed_x = transform(x)

show_image(transed_x)

transform = transforms.Pad(padding=3, padding_mode="symmetric")
transed_x = transform(x)

show_image(transed_x)

transform = transforms.RandomHorizontalFlip(p=1.0)
transed_x = transform(x)

show_image(transed_x)

transform = transforms.RandomVerticalFlip(p=1.0)
transed_x = transform(x)

show_image(transed_x)

transform = transforms.RandomAffine(
    degrees=(-10, 10), translate=(0.1, 0.3), scale=(0.5, 0.8)
    ) 
transed_x = transform(x)

show_image(transed_x)

transform = transforms.transforms.RandomPerspective(distortion_scale=0.3, p=1.0)
transed_x = transform(x)

show_image(transed_x)

transform = transforms.RandomRotation(degrees=(-30, 30))
transed_x = transform(x)

show_image(transed_x)

transform = transforms.Grayscale(num_output_channels=1)
transed_x = transform(x)

show_image(transed_x)

transform = transforms.RandomGrayscale(p=1)
transed_x = transform(x)

show_image(transed_x)

transform = transforms.GaussianBlur(kernel_size=3, sigma=(0.1, 2.0))
transed_x = transform(x)

show_image(transed_x)

transform = transforms.ColorJitter(brightness=0.5, contrast=0.5, saturation=0.5, hue=0.5)
transed_x = transform(x)

show_image(transed_x)

transform = transforms.RandomInvert(p=1.0)
transed_x = transform(x)

show_image(transed_x)

transform = transforms.RandomPosterize(bits=3, p=1.0)
x_uint8 = (255*x).to(torch.uint8) # 0-1 -> 0-255 への逆正規化, float32->uint8 への変更

transed_x = transform(x_uint8)

show_image(transed_x)

transform = transforms.RandomSolarize(threshold=0.5)
transed_x = transform(x)

show_image(transed_x)

transform = transforms.RandomAdjustSharpness(sharpness_factor=5, p=0.5)
transed_x = transform(x)

show_image(transed_x)

transform = transforms.RandomAutocontrast(p=1.0)
transed_x = transform(x)

show_image(transed_x)

transform = transforms.RandomEqualize(p=1.0)
x_uint8 = (255*x).to(torch.uint8) # 0-1 -> 0-255 への逆正規化, float32->uint8への変更

transed_x = transform(x_uint8)

show_image(transed_x)