pytorch Dataset,DataLoader产生自定义的训练数据案例

1. torch.utils.data.Dataset

datasets这是一个pytorch定义的dataset的源码集合。下面是一个自定义Datasets的基本框架，初始化放在__init__()中，其中__getitem__()和__len__()两个方法是必须重写的。

__getitem__()返回训练数据，如图片和label，而__len__()返回数据长度。

class CustomDataset(data.Dataset):#需要继承data.Dataset
 def __init__(self):
  # TODO
  # 1. Initialize file path or list of file names.
  pass
 def __getitem__(self, index):
  # TODO
  # 1. Read one data from file (e.g. using numpy.fromfile, PIL.Image.open).
  # 2. Preprocess the data (e.g. torchvision.Transform).
  # 3. Return a data pair (e.g. image and label).
  #这里需要注意的是，第一步：read one data，是一个data
  pass
 def __len__(self):
  # You should change 0 to the total size of your dataset.
  return 0

2. torch.utils.data.DataLoader

DataLoader(object)可用参数：

dataset(Dataset) 传入的数据集

batch_size(int, optional)每个batch有多少个样本

shuffle(bool, optional)在每个epoch开始的时候，对数据进行重新排序

sampler(Sampler, optional) 自定义从数据集中取样本的策略，如果指定这个参数，那么shuffle必须为False

batch_sampler(Sampler, optional) 与sampler类似，但是一次只返回一个batch的indices（索引），需要注意的是，一旦指定了这个参数，那么batch_size,shuffle,sampler,drop_last就不能再制定了（互斥——Mutually exclusive）

num_workers (int, optional) 这个参数决定了有几个进程来处理data loading。0意味着所有的数据都会被load进主进程。（默认为0）

collate_fn (callable, optional) 将一个list的sample组成一个mini-batch的函数

pin_memory (bool, optional) 如果设置为True，那么data loader将会在返回它们之前，将tensors拷贝到CUDA中的固定内存（CUDA pinned memory）中.

drop_last (bool, optional) 如果设置为True：这个是对最后的未完成的batch来说的，比如你的batch_size设置为64，而一个epoch只有100个样本，那么训练的时候后面的36个就被扔掉了。 如果为False（默认），那么会继续正常执行，只是最后的batch_size会小一点。

timeout(numeric, optional) 如果是正数，表明等待从worker进程中收集一个batch等待的时间，若超出设定的时间还没有收集到，那就不收集这个内容了。这个numeric应总是大于等于0。默认为0

worker_init_fn (callable, optional) 每个worker初始化函数 If not None, this will be called on eachworker subprocess with the worker id (an int in [0, num_workers - 1]) as input, after seeding and before data loading. (default: None)

3. 使用Dataset, DataLoader产生自定义训练数据

假设TXT文件保存了数据的图片和label,格式如下：第一列是图片的名字，第二列是label

0.jpg 0
1.jpg 1
2.jpg 2
3.jpg 3
4.jpg 4
5.jpg 5
6.jpg 6
7.jpg 7
8.jpg 8
9.jpg 9

也可以是多标签的数据，如：

0.jpg 0 10
1.jpg 1 11
2.jpg 2 12
3.jpg 3 13
4.jpg 4 14
5.jpg 5 15
6.jpg 6 16
7.jpg 7 17
8.jpg 8 18
9.jpg 9 19

图库十张原始图片放在./dataset/images目录下，然后我们就可以自定义一个Dataset解析这些数据并读取图片，再使用DataLoader类产生batch的训练数据

3.1 自定义Dataset

首先先自定义一个TorchDataset类，用于读取图片数据，产生标签：

注意初始化函数：

import torch
from torch.autograd import Variable
from torchvision import transforms
from torch.utils.data import Dataset, DataLoader
import numpy as np
from utils import image_processing
import os

class TorchDataset(Dataset):
 def __init__(self, filename, image_dir, resize_height=256, resize_width=256, repeat=1):
  '''
  :param filename: 数据文件TXT：格式：imge_name.jpg label1_id labe2_id
  :param image_dir: 图片路径：image_dir+imge_name.jpg构成图片的完整路径
  :param resize_height 为None时，不进行缩放
  :param resize_width 为None时，不进行缩放，
        PS：当参数resize_height或resize_width其中一个为None时，可实现等比例缩放
  :param repeat: 所有样本数据重复次数，默认循环一次，当repeat为None时，表示无限循环<sys.maxsize
  '''
  self.image_label_list = self.read_file(filename)
  self.image_dir = image_dir
  self.len = len(self.image_label_list)
  self.repeat = repeat
  self.resize_height = resize_height
  self.resize_width = resize_width

  # 相关预处理的初始化
  '''class torchvision.transforms.ToTensor'''
  # 把shape=(H,W,C)的像素值范围为[0, 255]的PIL.Image或者numpy.ndarray数据
  # 转换成shape=(C,H,W)的像素数据，并且被归一化到[0.0, 1.0]的torch.FloatTensor类型。
  self.toTensor = transforms.ToTensor()

  '''class torchvision.transforms.Normalize(mean, std)
  此转换类作用于torch. * Tensor,给定均值(R, G, B) 和标准差(R, G, B)，
  用公式channel = (channel - mean) / std进行规范化。
  '''
  # self.normalize=transforms.Normalize()

 def __getitem__(self, i):
  index = i % self.len
  # print("i={},index={}".format(i, index))
  image_name, label = self.image_label_list[index]
  image_path = os.path.join(self.image_dir, image_name)
  img = self.load_data(image_path, self.resize_height, self.resize_width, normalization=False)
  img = self.data_preproccess(img)
  label=np.array(label)
  return img, label

 def __len__(self):
  if self.repeat == None:
   data_len = 10000000
  else:
   data_len = len(self.image_label_list) * self.repeat
  return data_len

 def read_file(self, filename):
  image_label_list = []
  with open(filename, 'r') as f:
   lines = f.readlines()
   for line in lines:
    # rstrip：用来去除结尾字符、空白符(包括\n、\r、\t、' '，即：换行、回车、制表符、空格)
    content = line.rstrip().split(' ')
    name = content[0]
    labels = []
    for value in content[1:]:
     labels.append(int(value))
    image_label_list.append((name, labels))
  return image_label_list

 def load_data(self, path, resize_height, resize_width, normalization):
  '''
  加载数据
  :param path:
  :param resize_height:
  :param resize_width:
  :param normalization: 是否归一化
  :return:
  '''
  image = image_processing.read_image(path, resize_height, resize_width, normalization)
  return image

 def data_preproccess(self, data):
  '''
  数据预处理
  :param data:
  :return:
  '''
  data = self.toTensor(data)
  return data

3.2 DataLoader产生批训练数据

if __name__=='__main__':
 train_filename="../dataset/train.txt"
 # test_filename="../dataset/test.txt"
 image_dir='../dataset/images'

 epoch_num=2 #总样本循环次数
 batch_size=7 #训练时的一组数据的大小
 train_data_nums=10
 max_iterate=int((train_data_nums+batch_size-1)/batch_size*epoch_num) #总迭代次数

 train_data = TorchDataset(filename=train_filename, image_dir=image_dir,repeat=1)
 # test_data = TorchDataset(filename=test_filename, image_dir=image_dir,repeat=1)
 train_loader = DataLoader(dataset=train_data, batch_size=batch_size, shuffle=False)
 # test_loader = DataLoader(dataset=test_data, batch_size=batch_size,shuffle=False)

 # [1]使用epoch方法迭代，TorchDataset的参数repeat=1
 for epoch in range(epoch_num):
  for batch_image, batch_label in train_loader:
   image=batch_image[0,:]
   image=image.numpy()#image=np.array(image)
   image = image.transpose(1, 2, 0) # 通道由[c,h,w]->[h,w,c]
   image_processing.cv_show_image("image",image)
   print("batch_image.shape:{},batch_label:{}".format(batch_image.shape,batch_label))
   # batch_x, batch_y = Variable(batch_x), Variable(batch_y)

上面的迭代代码是通过两个for实现，其中参数epoch_num表示总样本循环次数，比如epoch_num=2，那就是所有样本循环迭代2次。

但这会出现一个问题，当样本总数train_data_nums与batch_size不能整取时，最后一个batch会少于规定batch_size的大小，比如这里样本总数train_data_nums=10，batch_size=7，第一次迭代会产生7个样本，第二次迭代会因为样本不足，只能产生3个样本。

我们希望，每次迭代都会产生相同大小的batch数据，因此可以如下迭代：注意本人在构造TorchDataset类时，就已经考虑循环迭代的方法，因此，你现在只需修改repeat为None时，就表示无限循环了，调用方法如下：

 '''
 下面两种方式，TorchDataset设置repeat=None可以实现无限循环，退出循环由max_iterate设定
 '''
 train_data = TorchDataset(filename=train_filename, image_dir=image_dir,repeat=None)
 train_loader = DataLoader(dataset=train_data, batch_size=batch_size, shuffle=False)
 # [2]第2种迭代方法
 for step, (batch_image, batch_label) in enumerate(train_loader):
  image=batch_image[0,:]
  image=image.numpy()#image=np.array(image)
  image = image.transpose(1, 2, 0) # 通道由[c,h,w]->[h,w,c]
  image_processing.cv_show_image("image",image)
  print("step:{},batch_image.shape:{},batch_label:{}".format(step,batch_image.shape,batch_label))
  # batch_x, batch_y = Variable(batch_x), Variable(batch_y)
  if step>=max_iterate:
   break
 # [3]第3种迭代方法
 # for step in range(max_iterate):
 #  batch_image, batch_label=train_loader.__iter__().__next__()
 #  image=batch_image[0,:]
 #  image=image.numpy()#image=np.array(image)
 #  image = image.transpose(1, 2, 0) # 通道由[c,h,w]->[h,w,c]
 #  image_processing.cv_show_image("image",image)
 #  print("batch_image.shape:{},batch_label:{}".format(batch_image.shape,batch_label))
 #  # batch_x, batch_y = Variable(batch_x), Variable(batch_y)

3.3 附件：image_processing.py

上面代码，用到image_processing，这是本人封装好的图像处理包，包含读取图片，画图等基本方法：

# -*-coding: utf-8 -*-
"""
 @Project: IntelligentManufacture
 @File : image_processing.py
 @Author : panjq
 @E-mail : pan_jinquan@163.com
 @Date : 2019-02-14 15:34:50
"""

import os
import glob
import cv2
import numpy as np
import matplotlib.pyplot as plt

def show_image(title, image):
 '''
 调用matplotlib显示RGB图片
 :param title: 图像标题
 :param image: 图像的数据
 :return:
 '''
 # plt.figure("show_image")
 # print(image.dtype)
 plt.imshow(image)
 plt.axis('on') # 关掉坐标轴为 off
 plt.title(title) # 图像题目
 plt.show()

def cv_show_image(title, image):
 '''
 调用OpenCV显示RGB图片
 :param title: 图像标题
 :param image: 输入RGB图像
 :return:
 '''
 channels=image.shape[-1]
 if channels==3:
  image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR) # 将BGR转为RGB
 cv2.imshow(title,image)
 cv2.waitKey(0)

def read_image(filename, resize_height=None, resize_width=None, normalization=False):
 '''
 读取图片数据,默认返回的是uint8,[0,255]
 :param filename:
 :param resize_height:
 :param resize_width:
 :param normalization:是否归一化到[0.,1.0]
 :return: 返回的RGB图片数据
 '''

 bgr_image = cv2.imread(filename)
 # bgr_image = cv2.imread(filename,cv2.IMREAD_IGNORE_ORIENTATION|cv2.IMREAD_COLOR)
 if bgr_image is None:
  print("Warning:不存在:{}", filename)
  return None
 if len(bgr_image.shape) == 2: # 若是灰度图则转为三通道
  print("Warning:gray image", filename)
  bgr_image = cv2.cvtColor(bgr_image, cv2.COLOR_GRAY2BGR)

 rgb_image = cv2.cvtColor(bgr_image, cv2.COLOR_BGR2RGB) # 将BGR转为RGB
 # show_image(filename,rgb_image)
 # rgb_image=Image.open(filename)
 rgb_image = resize_image(rgb_image,resize_height,resize_width)
 rgb_image = np.asanyarray(rgb_image)
 if normalization:
  # 不能写成:rgb_image=rgb_image/255
  rgb_image = rgb_image / 255.0
 # show_image("src resize image",image)
 return rgb_image

def fast_read_image_roi(filename, orig_rect, ImreadModes=cv2.IMREAD_COLOR, normalization=False):
 '''
 快速读取图片的方法
 :param filename: 图片路径
 :param orig_rect:原始图片的感兴趣区域rect
 :param ImreadModes: IMREAD_UNCHANGED
      IMREAD_GRAYSCALE
      IMREAD_COLOR
      IMREAD_ANYDEPTH
      IMREAD_ANYCOLOR
      IMREAD_LOAD_GDAL
      IMREAD_REDUCED_GRAYSCALE_2
      IMREAD_REDUCED_COLOR_2
      IMREAD_REDUCED_GRAYSCALE_4
      IMREAD_REDUCED_COLOR_4
      IMREAD_REDUCED_GRAYSCALE_8
      IMREAD_REDUCED_COLOR_8
      IMREAD_IGNORE_ORIENTATION
 :param normalization: 是否归一化
 :return: 返回感兴趣区域ROI
 '''
 # 当采用IMREAD_REDUCED模式时，对应rect也需要缩放
 scale=1
 if ImreadModes == cv2.IMREAD_REDUCED_COLOR_2 or ImreadModes == cv2.IMREAD_REDUCED_COLOR_2:
  scale=1/2
 elif ImreadModes == cv2.IMREAD_REDUCED_GRAYSCALE_4 or ImreadModes == cv2.IMREAD_REDUCED_COLOR_4:
  scale=1/4
 elif ImreadModes == cv2.IMREAD_REDUCED_GRAYSCALE_8 or ImreadModes == cv2.IMREAD_REDUCED_COLOR_8:
  scale=1/8
 rect = np.array(orig_rect)*scale
 rect = rect.astype(int).tolist()
 bgr_image = cv2.imread(filename,flags=ImreadModes)

 if bgr_image is None:
  print("Warning:不存在:{}", filename)
  return None
 if len(bgr_image.shape) == 3: #
  rgb_image = cv2.cvtColor(bgr_image, cv2.COLOR_BGR2RGB) # 将BGR转为RGB
 else:
  rgb_image=bgr_image #若是灰度图
 rgb_image = np.asanyarray(rgb_image)
 if normalization:
  # 不能写成:rgb_image=rgb_image/255
  rgb_image = rgb_image / 255.0
 roi_image=get_rect_image(rgb_image , rect)
 # show_image_rect("src resize image",rgb_image,rect)
 # cv_show_image("reROI",roi_image)
 return roi_image

def resize_image(image,resize_height, resize_width):
 '''
 :param image:
 :param resize_height:
 :param resize_width:
 :return:
 '''
 image_shape=np.shape(image)
 height=image_shape[0]
 width=image_shape[1]
 if (resize_height is None) and (resize_width is None):#错误写法：resize_height and resize_width is None
  return image
 if resize_height is None:
  resize_height=int(height*resize_width/width)
 elif resize_width is None:
  resize_width=int(width*resize_height/height)
 image = cv2.resize(image, dsize=(resize_width, resize_height))
 return image
def scale_image(image,scale):
 '''
 :param image:
 :param scale: (scale_w,scale_h)
 :return:
 '''
 image = cv2.resize(image,dsize=None, fx=scale[0],fy=scale[1])
 return image

def get_rect_image(image,rect):
 '''
 :param image:
 :param rect: [x,y,w,h]
 :return:
 '''
 x, y, w, h=rect
 cut_img = image[y:(y+ h),x:(x+w)]
 return cut_img
def scale_rect(orig_rect,orig_shape,dest_shape):
 '''
 对图像进行缩放时，对应的rectangle也要进行缩放
 :param orig_rect: 原始图像的rect=[x,y,w,h]
 :param orig_shape: 原始图像的维度shape=[h,w]
 :param dest_shape: 缩放后图像的维度shape=[h,w]
 :return: 经过缩放后的rectangle
 '''
 new_x=int(orig_rect[0]*dest_shape[1]/orig_shape[1])
 new_y=int(orig_rect[1]*dest_shape[0]/orig_shape[0])
 new_w=int(orig_rect[2]*dest_shape[1]/orig_shape[1])
 new_h=int(orig_rect[3]*dest_shape[0]/orig_shape[0])
 dest_rect=[new_x,new_y,new_w,new_h]
 return dest_rect

def show_image_rect(win_name,image,rect):
 '''
 :param win_name:
 :param image:
 :param rect:
 :return:
 '''
 x, y, w, h=rect
 point1=(x,y)
 point2=(x+w,y+h)
 cv2.rectangle(image, point1, point2, (0, 0, 255), thickness=2)
 cv_show_image(win_name, image)

def rgb_to_gray(image):
 image = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
 return image

def save_image(image_path, rgb_image,toUINT8=True):
 if toUINT8:
  rgb_image = np.asanyarray(rgb_image * 255, dtype=np.uint8)
 if len(rgb_image.shape) == 2: # 若是灰度图则转为三通道
  bgr_image = cv2.cvtColor(rgb_image, cv2.COLOR_GRAY2BGR)
 else:
  bgr_image = cv2.cvtColor(rgb_image, cv2.COLOR_RGB2BGR)
 cv2.imwrite(image_path, bgr_image)

def combime_save_image(orig_image, dest_image, out_dir,name,prefix):
 '''
 命名标准：out_dir/name_prefix.jpg
 :param orig_image:
 :param dest_image:
 :param image_path:
 :param out_dir:
 :param prefix:
 :return:
 '''
 dest_path = os.path.join(out_dir, name + "_"+prefix+".jpg")
 save_image(dest_path, dest_image)

 dest_image = np.hstack((orig_image, dest_image))
 save_image(os.path.join(out_dir, "{}_src_{}.jpg".format(name,prefix)), dest_image)

3.4 完整的代码

# -*-coding: utf-8 -*-
"""
 @Project: pytorch-learning-tutorials
 @File : dataset.py
 @Author : panjq
 @E-mail : pan_jinquan@163.com
 @Date : 2019-03-07 18:45:06
"""
import torch
from torch.autograd import Variable
from torchvision import transforms
from torch.utils.data import Dataset, DataLoader
import numpy as np
from utils import image_processing
import os

class TorchDataset(Dataset):
 def __init__(self, filename, image_dir, resize_height=256, resize_width=256, repeat=1):
  '''
  :param filename: 数据文件TXT：格式：imge_name.jpg label1_id labe2_id
  :param image_dir: 图片路径：image_dir+imge_name.jpg构成图片的完整路径
  :param resize_height 为None时，不进行缩放
  :param resize_width 为None时，不进行缩放，
        PS：当参数resize_height或resize_width其中一个为None时，可实现等比例缩放
  :param repeat: 所有样本数据重复次数，默认循环一次，当repeat为None时，表示无限循环<sys.maxsize
  '''
  self.image_label_list = self.read_file(filename)
  self.image_dir = image_dir
  self.len = len(self.image_label_list)
  self.repeat = repeat
  self.resize_height = resize_height
  self.resize_width = resize_width

  # 相关预处理的初始化
  '''class torchvision.transforms.ToTensor'''
  # 把shape=(H,W,C)的像素值范围为[0, 255]的PIL.Image或者numpy.ndarray数据
  # 转换成shape=(C,H,W)的像素数据，并且被归一化到[0.0, 1.0]的torch.FloatTensor类型。
  self.toTensor = transforms.ToTensor()

  '''class torchvision.transforms.Normalize(mean, std)
  此转换类作用于torch. * Tensor,给定均值(R, G, B) 和标准差(R, G, B)，
  用公式channel = (channel - mean) / std进行规范化。
  '''
  # self.normalize=transforms.Normalize()

 def __getitem__(self, i):
  index = i % self.len
  # print("i={},index={}".format(i, index))
  image_name, label = self.image_label_list[index]
  image_path = os.path.join(self.image_dir, image_name)
  img = self.load_data(image_path, self.resize_height, self.resize_width, normalization=False)
  img = self.data_preproccess(img)
  label=np.array(label)
  return img, label

 def __len__(self):
  if self.repeat == None:
   data_len = 10000000
  else:
   data_len = len(self.image_label_list) * self.repeat
  return data_len

 def read_file(self, filename):
  image_label_list = []
  with open(filename, 'r') as f:
   lines = f.readlines()
   for line in lines:
    # rstrip：用来去除结尾字符、空白符(包括\n、\r、\t、' '，即：换行、回车、制表符、空格)
    content = line.rstrip().split(' ')
    name = content[0]
    labels = []
    for value in content[1:]:
     labels.append(int(value))
    image_label_list.append((name, labels))
  return image_label_list

 def load_data(self, path, resize_height, resize_width, normalization):
  '''
  加载数据
  :param path:
  :param resize_height:
  :param resize_width:
  :param normalization: 是否归一化
  :return:
  '''
  image = image_processing.read_image(path, resize_height, resize_width, normalization)
  return image

 def data_preproccess(self, data):
  '''
  数据预处理
  :param data:
  :return:
  '''
  data = self.toTensor(data)
  return data

if __name__=='__main__':
 train_filename="../dataset/train.txt"
 # test_filename="../dataset/test.txt"
 image_dir='../dataset/images'

 epoch_num=2 #总样本循环次数
 batch_size=7 #训练时的一组数据的大小
 train_data_nums=10
 max_iterate=int((train_data_nums+batch_size-1)/batch_size*epoch_num) #总迭代次数

 train_data = TorchDataset(filename=train_filename, image_dir=image_dir,repeat=1)
 # test_data = TorchDataset(filename=test_filename, image_dir=image_dir,repeat=1)
 train_loader = DataLoader(dataset=train_data, batch_size=batch_size, shuffle=False)
 # test_loader = DataLoader(dataset=test_data, batch_size=batch_size,shuffle=False)

 # [1]使用epoch方法迭代，TorchDataset的参数repeat=1
 for epoch in range(epoch_num):
  for batch_image, batch_label in train_loader:
   image=batch_image[0,:]
   image=image.numpy()#image=np.array(image)
   image = image.transpose(1, 2, 0) # 通道由[c,h,w]->[h,w,c]
   image_processing.cv_show_image("image",image)
   print("batch_image.shape:{},batch_label:{}".format(batch_image.shape,batch_label))
   # batch_x, batch_y = Variable(batch_x), Variable(batch_y)

 '''
 下面两种方式，TorchDataset设置repeat=None可以实现无限循环，退出循环由max_iterate设定
 '''
 train_data = TorchDataset(filename=train_filename, image_dir=image_dir,repeat=None)
 train_loader = DataLoader(dataset=train_data, batch_size=batch_size, shuffle=False)
 # [2]第2种迭代方法
 for step, (batch_image, batch_label) in enumerate(train_loader):
  image=batch_image[0,:]
  image=image.numpy()#image=np.array(image)
  image = image.transpose(1, 2, 0) # 通道由[c,h,w]->[h,w,c]
  image_processing.cv_show_image("image",image)
  print("step:{},batch_image.shape:{},batch_label:{}".format(step,batch_image.shape,batch_label))
  # batch_x, batch_y = Variable(batch_x), Variable(batch_y)
  if step>=max_iterate:
   break
 # [3]第3种迭代方法
 # for step in range(max_iterate):
 #  batch_image, batch_label=train_loader.__iter__().__next__()
 #  image=batch_image[0,:]
 #  image=image.numpy()#image=np.array(image)
 #  image = image.transpose(1, 2, 0) # 通道由[c,h,w]->[h,w,c]
 #  image_processing.cv_show_image("image",image)
 #  print("batch_image.shape:{},batch_label:{}".format(batch_image.shape,batch_label))
 #  # batch_x, batch_y = Variable(batch_x), Variable(batch_y)

以上为个人经验，希望能给大家一个参考，也希望大家多多支持我们。如有错误或未考虑完全的地方，望不吝赐教。