pytorch学习笔记——基础知识

推荐一个入门学习的地址https://github.com/bharathgs/Awesome-pytorch-list

pyTorch doc:https://pytorch.org/docs/stable/index.html

 

anaconda 安装pytorch  (support cuda)

pip install torch==1.1.0 -f https://download.pytorch.org/whl/cu90/stable

 

疑问：

1.优化器的选择；loss函数的选择；参数的设置

目录

1.读取图像数据

method1：测试单张图片

method2：自定义类，读取真个文件夹

2.cuda 使用

3.tensor和Variable

4.transforms

5.DataLoader 和 Dataset

5.1 直接调用

5.2 自定义

5.2.1 each class one folder

5.2.2 all class in one folderr

6.torchversion

 *.不熟悉的code记录

*.1 torch.backends.cudnn.benchmark = true

---

1.读取图像数据

method1：测试单张图片

from PIL import Image

import torchvision.transforms as transforms

#transform=transform=transforms.ToTensor()
transform = transforms.Compose([transforms.ToTensor(),
                                transforms.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225))])
img_path="5_283.bmp"
img = Image.open(img_path)
data = Variable(transform(img)) // transform可以根据自己的需要来设置
#print(data.size())
data = data.unsqueeze(0) //这一步是因为我在测试时维度部队，net需要一个四维的输入，而data只是一个三维的数据，所以在外面在加一层维度

 

method2：自定义类，读取真个文件夹

Pytorch自定义读取数据的方式，主要用到两个类：
torch.utils.data.Dataset和torch.utils.data.DataLoade

import cv2
import os
import numpy as np
import torch 
import torchvision.transforms as transforms
from torch.utils.data import Dataset, DataLoader


name = 'C:\\Users\\Administrator\\Desktop\\history\\HED-BSDS\\test1.lst'
base_path = 'C:\\Users\\Administrator\\Desktop\\history\\HED-BSDS'

#首先定义一个Dataset的子类->myDataset
class myDataset(Dataset):
    def __init__(self, name, base_path):
        f = open(name)
        self.filenames = f.readlines()
        f.close()
 #override这两个方法
    def __getitem__(self, index):
        path = self.filenames[index]
        print(os.path.join(base_path, path))
        img = cv2.imread(os.path.join(base_path, path).strip())
        img = torch.Tensor(img)
        return img

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

dataset = myDataset

train_loader = DataLoader(dataset(name=name, base_path=base_path), 
    batch_size=4, shuffle=True)
for img in train_loader:
    print(img.size())
    cv2.imshow('we', np.uint8(img.numpy()[0]))
    cv2.waitKey()

 

2.cuda 使用

1.确认设备是否支持cuda

improt torch
torch.cuuda.is_available()//如果返回false，则不支持

2.数据转换到cuda上，相关的有网络模型net，训练数据data，损失函数criterion（可选）

转换的方式有两种：1.to(device)  2.  net.cuda()

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = Net().to(device)

for batch_idx, (data, target) in enumerate(train_loader):
        data, target = Variable(data.to(device)), Variable(target.to(device))

 

3.tensor和Variable

PyTorch 在autograd模块中实现了计算图的相关功能，autograd的核心数据结构是Variable。Variable封装了Tensor，并记录对Tensor的操作记录来构建计算图。
forward的输入输出都是Variable，只有Variable才具有自动求导功能，Tensor是没有的，所以在输入时，需要把Tensor封装成Varible。

4.transforms

https://pytorch-cn.readthedocs.io/zh/latest/torchvision/torchvision-transform/

1）scale(size): 将w，h 缩放成size，最小边为size

2）CenterCrop(size):size可是一个int，也可以是（w,h）中心切割

3）RandomCrop(size)；随机切割

4）RandomHorizontalFlip：随机水平翻转

5）RandomSizedCrop(size, interpolation=2)：随机切，然后缩放成size

6）Pad：填充

normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
transforms.Compose([
                                 transforms.RandomSizedCrop(224),
                                 transforms.RandomHorizontalFlip(),
                                 transforms.ToTensor(),
                                 normalize,
                  ])

5.DataLoader 和 Dataset

5.1 直接调用

大多数其他的PyTorch教程和示例都希望你先按照训练集和验证集来组织文件夹，然后在训练集和验证集中再按照类别进行组织。

normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
	train_loader = data.DataLoader(
        datasets.ImageFolder(filepath,
                             transforms.Compose([
                                 transforms.RandomSizedCrop(224),
                                 transforms.RandomHorizontalFlip(),
                                 transforms.ToTensor(),
                                 normalize,
                             ])),
        batch_size=16,
        shuffle=True)

5.2 自定义

5.2.1 each class one folder

class CatDogDataset(Dataset):
    def __init__(self, root, resize, mode, namefile):
        self.root = root
        self.size = resize
        self.namefile = namefile
        self.name2label = {}
        for name in sorted(os.listdir(os.path.join(root))):
            #过滤掉非目录文件
            if not os.path.isdir(os.path.join(root, name)):
                continue
            #构建字典，名字：0~4数字
            self.name2label[name] = len(self.name2label.keys())
        print(self.name2label)
        self.images, self.labels = self.load_csv()

        if mode == "train": # 80%
            self.images = self.images[:int(0.8*len(self.images))]
            self.labels = self.labels[:int(0.8*len(self.labels))]
        elif mode == "val": # 10% = 80%~95%
            self.images = self.images[int(0.8*len(self.images)):int(0.95 * len(self.images))]
            self.labels = self.labels[int(0.8*len(self.labels)):int(0.95 * len(self.labels))]
        else: # 5% = 95%~100%
            self.images = self.images[int(0.95 * len(self.images)):]
            self.labels = self.labels[int(0.95 * len(self.labels)):]

    def load_csv(self):
            if not os.path.isdir(os.path.join(self.root, self.namefile)):
                images = []
                for name in self.name2label.keys():
                    images += glob.glob(os.path.join(self.root, name, '*jpg'))
                    images += glob.glob(os.path.join(self.root, name, '*jpeg'))
                    images += glob.glob(os.path.join(self.root, name, '*png'))

		# 1165 ['pokemon/pokeman/bulbasaur/00000159.jpg',
		#print(len(images), images)
                random.shuffle(images)
                with open(os.path.join(self.root, self.namefile), mode='w',newline='') as f:
                    writer = csv.writer(f)
                    for img in images:
                        name = img.split(os.sep)[-2]
                        label = self.name2label[name]
                        writer.writerow([img, label])
                    print('writer into cvs file:', self.namefile)
            
            images,labels = [],[]
            #images, labels = [],[]
            
            with open(os.path.join(self.root, self.namefile)) as f:
                reader = csv.reader(f)
                for row in reader:
                    img, label = row
                    label = int(label)
                    images.append(img)
                    labels.append(label)
            assert len(images) == len(labels), 'num of imgs != labels'
            return images, labels

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

 
    def denormalize(self, x_hat):
 
        mean = [0.485, 0.456, 0.406]
        std = [0.229, 0.224, 0.225]
 
        # x_hat = (x-mean)/std
        # x = x_hat*std = mean
        # x: [c, h, w]
        # mean: [3] => [3, 1, 1]
        mean = torch.tensor(mean).unsqueeze(1).unsqueeze(1)
        std = torch.tensor(std).unsqueeze(1).unsqueeze(1)
        # print(mean.shape, std.shape)
        x = x_hat * std + mean
 
        return x
 
    def __getitem__(self, idx):
        # idx~[0~len(images)]
        # self.images, self.labels
        # img: 'pokemon/bulbasaur/00000000.png'
        # label: 0
        img, label = self.images[idx], self.labels[idx]
 
        tf = transforms.Compose([
            lambda x:Image.open(x).convert("RGB"), # string path => image data
            transforms.Scale(256),
            transforms.CenterCrop(224),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                 std=[0.229, 0.224, 0.225])
        ])
 
        img = tf(img)
        label = torch.tensor(label)
 
        return img, label

5.2.2 all class in one folderr

6.torchversion

专门处理图像的库。主要包含datasets  models transforms  utils

包含的数据集：

• MNISTCOCO
• Captions
• Detection
• LSUN
• ImageFolder
• Imagenet-12
• CIFAR
• STL10
• SVHN
• PhotoTour

 包含的模型：

• AlexNet
• VGG
• ResNet
• SqueezeNet
• DenseNet

加载模型example

import torchvision.models as models
resnet18 = models.resnet18()
alexnet = models.alexnet()
squeezenet = models.squeezenet1_0()
densenet = models.densenet_161()

也可以通过使用 pretrained=True 来加载一个别人预训练好的模型

 

 *.不熟悉的code记录

*.1 torch.backends.cudnn.benchmark = true

如果网络的输入数据维度或类型上变化不大，设置 可以增加运行效率；如果每次迭代都变化的话，会导致 cnDNN 每次都会去寻找一遍最优配置，这样反而会降低运行效率。