文章目錄

• 一、torchvision 影象資料讀取 [0, 1]
• 二、torchvision 的 Transform
• 三、讀取影象資料類
• 3.1 class torchvision.datasets.ImageFolder 預設讀取影象資料方法：
• 3.2 自定義資料讀取方法

conda install pytorch -c pytorch
pip install config
pip install tqdm             #包裝迭代器，顯示進度條
pip install torchvision
pip install scikit-image
 

一、torchvision 影象資料讀取 [0, 1]

import torchvision.transforms as transforms
transforms 模組提供了一般的影象轉換操作類。
class torchvision.transforms.ToTensor
功能：
把shape=(H x W x C) 的畫素值為 [0, 255] 的 PIL.Image 和 numpy.ndarray
轉換成shape=(C x H x W)的畫素值範圍為[0.0, 1.0]的 torch.FloatTensor。

class torchvision.transforms.Normalize(mean, std)

import torchvision 
import torchvision.transforms as transforms 
import cv2 
import numpy as np 
from PIL import Image 

img_path = "./data/timg.jpg" 

# 引入transforms.ToTensor()功能： range [0, 255] -> [0.0,1.0] 
 

transform1 = transforms.Compose([transforms.ToTensor()])

# 直接讀取：numpy.ndarray 
img = cv2.imread(img_path)
print("img = ", img[0])      #只輸出其中一個通道
print("img.shape = ", img.shape)

# 歸一化，轉化為numpy.ndarray並顯示
img1 = transform1(img) 
img2 = img1.numpy()*255 
img2 = img2.astype('uint8') 
img2 = np.transpose(img2 , (1,2,0)) 
 
print("img1 = ", img1)
cv2.imshow('img2 ', img2 ) 
cv2.waitKey() 


# PIL 讀取影象
img = Image.open(img_path).convert('RGB') # 讀取影象 
img2 = transform1(img) # 歸一化到 [0.0,1.0] 
print("img2 = ",img2) #轉化為PILImage並顯示 
img_2 = transforms.ToPILImage()(img2).convert('RGB') 
print("img_2 = ",img_2) 
img_2.show()


從上到下依次輸出：---------------------------------------------
img =   [[197 203 202]
	 [195 203 202]
	 ...
	 [200 208 207]
	 [200 208 207]]
img.shape =  (362, 434, 3)

img1 =  tensor([[[0.7725, 0.7647, 0.7686,  ..., 0.7804, 0.7843, 0.7843],
         [0.7765, 0.7725, 0.7686,  ..., 0.7686, 0.7608, 0.7569],
         [0.7843, 0.7725, 0.7686,  ..., 0.7725, 0.7686, 0.7569],
         ...,

img_transform =  tensor([[[0.7922, 0.7922, 0.7961,  ..., 0.8078, 0.8118, 0.8118],
         [0.7961, 0.8000, 0.7961,  ..., 0.7922, 0.7882, 0.7843],
         [0.8039, 0.8000, 0.7961,  ..., 0.8118, 0.8039, 0.7922],
         ...,

transforms.Compose 歸一化到 [-1.0, 1.0 ]

transform2 = transforms.Compose([transforms.ToTensor()])
transforms.Normalize(mean = (0.5, 0.5, 0.5), std = (0.5, 0.5, 0.5))]) 

二、torchvision 的 Transform

在深度學習時關於影象的資料讀取：由於Tensorflow不支援與numpy的無縫切換，導致難以使用現成的pandas等格式化資料讀取工具，造成了很多不必要的麻煩，而pytorch解決了這個問題。

pytorch自定義讀取資料和進行Transform的部分請見文件：
http://pytorch.org/tutorials/beginner/data_loading_tutorial.html

但是按照文件中所描述所完成的自定義Dataset只能夠使用自定義的Transform步驟，而torchvision包中已經給我們提供了很多影象transform步驟的實現，為了使用這些已經實現的Transform步驟，我們可以使用如下方法定義Dataset：

from __future__ import print_function, division 
import os 
import torch 
import pandas as pd 
from PIL import Image 
import numpy as np 
from torch.utils.data import Dataset, DataLoader 
from torchvision import transforms 

class FaceLandmarkDataset(Dataset): 
    def __len__(self) -> int: 
        return len(self.landmarks_frame)
		
    def __init__(self, csv_file: str, root_dir: str, transform=None) -> None: 
        super().__init__() 
        self.landmarks_frame = pd.read_csv(csv_file) 
        self.root_dir = root_dir 
        self.transform = transform 

    def __getitem__(self, index:int): 
        img_name = self.landmarks_frame.ix[index, 0] 
        img_path = os.path.join('./faces', img_name) 
        with Image.open(img_path) as img: 
            image = img.convert('RGB') 
        landmarks = self.landmarks_frame.as_matrix()[index, 1:].astype('float') 
        landmarks = np.reshape(landmarks,newshape=(-1,2)) 
        if self.transform is not None: 
            image = self.transform(image) 
        return image, landmarks 

########################以上為資料讀取類（返回：image,landmarks）###############################
trans = transforms.Compose(transforms = [transforms.RandomSizedCrop(size=128), 
                                         transforms.ToTensor()]) 

face_dataset = FaceLandmarkDataset(csv_file='faces/face_landmarks.csv', 
				   root_dir='faces', transform= trans) 
loader = DataLoader(dataset = face_dataset, 
                    batch_size=4,
		    shuffle=True,
		    num_workers=4)

三、讀取影象資料類

3.1 class torchvision.datasets.ImageFolder 預設讀取影象資料方法：

• __init__（ 初始化）
  • classes, class_to_idx = find_classes(root) :得到分類的類別名（classes）和類別名與數字類別的對映關係字典（class_to_idx）
    其中 classes (list): List of the class names.
    其中 class_to_idx (dict): Dict with items (class_name, class_index).
  • imgs = make_dataset(root, class_to_idx)得到imags列表。
    其中 imgs (list): List of (image path, class_index) tuples
    每個值是一個tuple，每個tuple包含兩個元素：影象路徑和標籤
• __getitem__（影象獲取）
  • path, target = self.imgs[index] 獲取影象（路徑，標籤）
  • img = self.loader(path)資料讀取。
  • img = self.transform(img)資料、標籤 轉換成 tensor
  • target = self.target_transform(target)
• __len__（ 資料集數量）
  • return len(self.imgs)

class ImageFolder(data.Dataset):
    """預設影象資料目錄結構
    root
    .
    ├──dog
    |   ├──001.png
    |   ├──002.png
    |   └──...
    └──cat  
    |   ├──001.png
    |   ├──002.png
    |   └──...
    └──...
    """
    def __init__(self, root, transform=None, target_transform=None,
                 loader=default_loader):
        classes, class_to_idx = find_classes(root)
        imgs = make_dataset(root, class_to_idx)
        if len(imgs) == 0:
            raise(RuntimeError("Found 0 images in subfolders of: " + root + "\n"
                               "Supported image extensions are: " + ",".join(IMG_EXTENSIONS)))

        self.root = root
        self.imgs = imgs
        self.classes = classes
        self.class_to_idx = class_to_idx
        self.transform = transform
        self.target_transform = target_transform
        self.loader = loader

    def __getitem__(self, index):
        """
        index (int): Index
	Returns:tuple: (image, target) where target is class_index of the target class.
        """
        path, target = self.imgs[index]
        img = self.loader(path)
        if self.transform is not None:
            img = self.transform(img)
        if self.target_transform is not None:
            target = self.target_transform(target)
        return img, target

    def __len__(self):
        return len(self.imgs)

影象獲取 __getitem__ 中，self.loader(path) 採用的是default_loader，如下

def pil_loader(path):    # 一般採用pil_loader函式。
    with open(path, 'rb') as f:
        with Image.open(f) as img:
            return img.convert('RGB')

def accimage_loader(path):
    import accimage
    try:
        return accimage.Image(path)
    except IOError:
        # Potentially a decoding problem, fall back to PIL.Image
        return pil_loader(path)

def default_loader(path):
    from torchvision import get_image_backend
    if get_image_backend() == 'accimage':
        return accimage_loader(path)
    else:
        return pil_loader(path)

3.2 自定義資料讀取方法

PyTorch中和資料讀取相關的類都要繼承一個基類：torch.utils.data.Dataset。
故需要改寫其中的 __init__、__len__、__getitem__ 等三個方法即可。

• __init__（）初始化傳入引數：
  • img_path 裡面為所有影象資料（包括訓練和測試）
    txt_path 裡面有 train.txt和val.txt兩個檔案：txt檔案中每行都是影象路徑，tab鍵，標籤。
  • 其中 self.img_name 和 self.img_label 的讀取方式就跟你資料的存放方式有關（需要調整的地方)
• __getitem__（）依然採用default_loader方法來讀取影象。
• Transform中將每張影象都封裝成 Tensor

class customData(Dataset):
    def __init__(self, img_path, txt_path, dataset = '',data_transforms=None, loader = default_loader):
        with open(txt_path) as input_file:
            """
	    關於json檔案解析：
	    https://blog.csdn.net/wsp_1138886114/article/details/83302339
	    txt檔案解析如下，具體文字解析具體分析，沒有定數
            """
            lines = input_file.readlines()
            self.img_name = [os.path.join(img_path, line.strip().split('\t')[0]) for line in lines]
            self.img_label = [int(line.strip().split('\t')[-1]) for line in lines]
        self.data_transforms = data_transforms
        self.dataset = dataset
        self.loader = loader

    def __len__(self):
        return len(self.img_name)

    def __getitem__(self, item):
        img_name = self.img_name[item]
        label = self.img_label[item]
        img = self.loader(img_name)

        if self.data_transforms is not None:
            try:
                img = self.data_transforms[self.dataset](img)
            except:
                print("Cannot transform image: {}".format(img_name))
        return img, label
#####################以上為影象資料讀取，返回（img, label）#########################

# 保證image_datasets與torchvision.datasets.ImageFolder類返回的資料型別一樣
image_datasets = {x: customData(img_path='/ImagePath',
                                txt_path=('/TxtFile/' + x + '.txt'),
                                data_transforms=data_transforms,
                                dataset=x) for x in ['train', 'val']}

#用torch.utils.data.DataLoader類，將這個batch的影象資料和標籤都分別封裝成Tensor。
dataloders = {x: torch.utils.data.DataLoader(image_datasets[x],
                                             batch_size=batch_size,
                                             shuffle=True) for x in ['train', 'val']}

# 模型儲存
torch.save(model, 'output/resnet_epoch{}.pkl'.format(epoch))

https://pytorch-cn.readthedocs.io/zh/latest/package_references/data/#torchutilsdata
鳴謝
https://www.cnblogs.com/denny402/p/5096001.html
https://blog.csdn.net/VictoriaW/article/details/72822005
https://blog.csdn.net/hao5335156/article/details/80593349
https://blog.csdn.net/u014380165/article/details/78634829