?

import torch
import torchvision


class ToTensor:
    # 定義一個(gè)將原數(shù)據(jù)轉(zhuǎn)為 Torch 的操作
    def __call__(self, X):
        return torch.from_numpy(X)


class MulTransform:
    # 定義一個(gè)將所有數(shù)據(jù)翻倍的預(yù)處理操作
    def __call__(self, X):
        X *= 2
        return X


# 定義預(yù)處理集合器
composed = torchvision.transforms.Compose([ToTensor(), MulTransform()])
composed

"""
Compose(
    <__main__.ToTensor object at 0x7fa24c2c6450>
    <__main__.MulTransform object at 0x7fa24bbfdf90>
)
"""

import numpy as np
data = np.array([1, 2, 3])
composed(data)
# tensor([2, 4, 6])

from torch.utils.data import Dataset
import pandas as pd


class WineDataset(Dataset):
    # 建立一個(gè)數(shù)據(jù)集合繼承  Dataset 即可
    def __init__(self, transform):
        # I初始化數(shù)據(jù)
        # 以pandas的形式讀入數(shù)據(jù)
        xy = pd.read_csv(
            "https://labfile.oss.aliyuncs.com/courses/2316/wine.csv", header=None)
        self.n_samples = xy.shape[0]

        # 這里我們就不做Tensor的轉(zhuǎn)換了，將其全部放入 transform 中
        self.x_data = xy.values[:, 1:]
        self.y_data = xy.values[:, 0].reshape(-1,1) # 轉(zhuǎn)為二維向量，和x匹配
        # 數(shù)據(jù)預(yù)處理集合
        self.transform = transform

    # 返回 dataset[index]
    def __getitem__(self, index):
        sample = self.x_data[index], self.y_data[index]

        if self.transform:
            sample = self.transform(sample)

        return sample
    # 返回?cái)?shù)據(jù)長(zhǎng)度

    def __len__(self):
        return self.n_samples

class Normalization:
    def __call__(self, sample):
        inputs, targets = sample
        amin, amax = inputs.min(), inputs.max()  # 求最大最小值
        inputs = (inputs-amin)/(amax-amin)  # (矩陣元素-最小值)/(最大值-最小值)
        return inputs, targets


# 測(cè)試代碼
a = 10*np.random.random((5, 5))
# 測(cè)試數(shù)據(jù) 前4列表示特征，最后一列表示標(biāo)簽
data = [a[:, 0:4], a[:, 4]]
Normalization()(data)

class ToTensor:
    def __call__(self, sample):
        inputs, targets = sample
        return torch.from_numpy(inputs), torch.from_numpy(targets)


# 測(cè)試代碼
a = 10*np.random.random((5, 5))
# 測(cè)試數(shù)據(jù) 前4列表示特征，最后一列表示標(biāo)簽
data = [a[:, 0:4], a[:, 4]]
ToTensor()(data)

# 定義 composed
composed = torchvision.transforms.Compose([Normalization(), ToTensor()])
# 傳入該參數(shù)，即可獲得一系列預(yù)處理之后的數(shù)據(jù)
dataset = WineDataset(transform=composed)
first_data = dataset[0]
features, labels = first_data
# 輸出類(lèi)型觀察數(shù)據(jù)是否發(fā)生改變
print(type(features), type(labels))
# 輸出內(nèi)容觀察數(shù)據(jù)是否進(jìn)行了歸一
print(features, labels)

!wget -nc  "https://labfile.oss.aliyuncs.com/courses/2534/cat.jpg"

import torchvision.transforms as transforms
from PIL import Image
import matplotlib.pyplot as plt
from matplotlib.pyplot import imshow
%matplotlib inline
img = Image.open("cat.jpg")
imshow(img)

# torchvision.transforms.CenterCrop（size）：從中心開(kāi)始，裁剪給定大小的 PIL 圖像
transform = transforms.CenterCrop((64, 200))
new_img = transform(img)
imshow(new_img)

# transforms.ColorJitter(brightness=0, contrast=0, saturation=0, hue=0)：
# 改變圖片的亮度、對(duì)比度和飽和度
plt.subplot(221)
imshow(img)
# r隨機(jī)改變亮度
my_img1 = transforms.ColorJitter((0.5, 0.6))(img)
plt.subplot(222)
imshow(my_img1)
# 隨機(jī)改變對(duì)比度
my_img2 = transforms.ColorJitter(0, (0.5, 0.6))(img)
plt.subplot(223)
imshow(my_img2)
# 隨機(jī)改變飽和度
my_img3 = transforms.ColorJitter(0, 0, (0.5, 0.6))(img)
plt.subplot(224)
imshow(my_img3)
plt.show()

plt.subplot(1, 3, 1)
imshow(img)
my_img1 = transforms.Grayscale(1)(img)
plt.subplot(132)
imshow(my_img1, 'gray')
my_img2 = transforms.Grayscale(3)(img)
plt.subplot(133)
imshow(my_img2)

# transforms.Pad（padding，fill = 0，padding_mode ='constant' ）：
# 使用給定的 pad 值將給定的 PIL 圖像四處填充
plt.subplot(121)
imshow(img)
# 四周加邊界
my_img = transforms.Pad(padding=20, fill=(0, 255, 255),
                        padding_mode='constant')(img)
plt.subplot(122).set_title("Pad")
imshow(my_img)

# transforms.RandomAffine(degrees, translate=None, scale=None,
#                       shear=None, resample=False, fillcolor=0)：
# 保持圖像中心不變的中心仿射變換(可以理解為不同程度的旋轉(zhuǎn)，再在空余位置補(bǔ) 0)
my_img1 = transforms.RandomAffine(60)(img)
plt.subplot(221).set_title("rotate_only")
imshow(my_img1)
my_img2 = transforms.RandomAffine(60, translate=(0.3, 0.3))(img)
plt.subplot(222).set_title("rotate_translate")
imshow(my_img2)
my_img3 = transforms.RandomAffine(60, scale=(2.0, 2.1))(img)
plt.subplot(223).set_title("rotate_scale")
imshow(my_img3)
my_img4 = transforms.RandomAffine(0, shear=60)(img)
plt.subplot(224).set_title("shear_only")
imshow(my_img4)
plt.tight_layout()

#torchvision.transforms.RandomResizedCrop(size, scale,ratio,interpolation)
new_img = transforms.RandomResizedCrop((128, 126), scale=(0.08, 1.0),
                                       ratio=(0.75, 1.333333333), interpolation=2)(img)
imshow(new_img)