MindSpore报错ValueError: when loss_fn is not None, train_dataset should return two elements, but got 3

经验分享 Tech Blogs 数据加载及处理-Data Loading&Processing
1

1.系统环境

硬件环境(Ascend/GPU/CPU): Ascend/GPU/CPU
MindSpore版本: 1.1
执行模式（动态图）:不限
Python版本: Python=3.7.5
操作系统平台: 不限

2 报错信息

2.1 问题描述

自定义数据集在第一轮可正常遍历，但第二轮时出现list index out of range

2.2 报错信息

ValueError: when loss_fn is not None, train_dataset should return two elements, but got 3

2.3 脚本代码

import mindspore.dataset as ds  
import matplotlib.pyplot as plt  
import mindspore.dataset.vision.py_transforms as py_trans  
from mindspore.dataset.transforms.py_transforms import Compose  
import mindspore.ops as ops  
# from mindspore import Tensor  
import numpy as np  
from PIL import Image  
import os  
import random  
from option import opt  
    
class DatasetGenerator:  
    def __init__(self, path, train, format='.png'):  
        # self.size = size  
        self.format = format  
        self.train = train  
        if train:  
            self.haze_imgs_dir = os.listdir(os.path.join(path, 'train', 'hazy'))  
            self.haze_imgs = [os.path.join(path, 'train', 'hazy', img) for img in self.haze_imgs_dir]  
            self.clear_dir = os.path.join(path, 'train', 'gt')  
        else:  
            self.haze_imgs_dir = os.listdir(os.path.join(path, 'test', 'hazy'))  
            self.haze_imgs = [os.path.join(path, 'test', 'hazy', img) for img in self.haze_imgs_dir]  
            self.clear_dir = os.path.join(path, 'test', 'gt')  
        # print(self.haze_imgs_dir, self.clear_dir)  
        np.random.seed(58)  
        self.__random_seed = []  
        for _ in range(len(self.haze_imgs)):  
            self.__random_seed.append(random.randint(0, 1000000))  
        self.__index = -1  
    
    def __getitem__(self, index):  
        self.__index += 1  
        haze = Image.open(self.haze_imgs[index])  
        # if isinstance(self.size,int):  
        #     while haze.size[0]<self.size or haze.size[1]<self.size :  
        #         index=random.randint(0,20000)  
        #         haze=Image.open(self.haze_imgs[index])  
        img = self.haze_imgs[index].split('\\')[-1]  
        # img=self.haze_imgs[index].split('/')[-1]  
        img_name = img.split('_')  
        # img_name=img.split('\\')[-1].split('_')  
        # print(img_name)  
        clear_name=f"{img_name[0]}_gt_{img_name[2]}"  
        # print(self.clear_dir, clear_name, os.path.join(self.clear_dir,clear_name))  
        clear=Image.open(os.path.join(self.clear_dir,clear_name))  
        w, h = clear.size  
        nw, nh = haze.size  
        left = (w - nw)/2  
        top = (h - nh)/2  
        right = (w + nw)/2  
        bottom = (h + nh)/2  
        clear = clear.crop((left, top, right, bottom))  
        return (haze, clear, index)  
    
    def __len__(self):  
        print(len(self.haze_imgs))  
        return len(self.haze_imgs)  
    
    def get_seed(self):  
        seed = self.__random_seed[self.__index]  
        return seed  
    
def decode(img):  
    return Image.fromarray(img)  
    
def set_random_seed(img_name, seed):  
    random.seed(seed)  
    return img_name  
    
ds.config.set_seed(8)  
# DATA_DIR = opt.data_url  
DATA_DIR = 'C:\Users\kle\Desktop\NTIRE2021'  
    
train_dataset_generator = DatasetGenerator(DATA_DIR, train=True)  
train_dataset = ds.GeneratorDataset(train_dataset_generator, ["hazy", "gt", "img_name"], shuffle=True)  
test_dataset_generator = DatasetGenerator(DATA_DIR, train=False)  
test_dataset = ds.GeneratorDataset(test_dataset_generator, ["hazy", "gt", "img_name"], shuffle=False)  
transforms_list = [  
    decode,  
    (lambda img_name: set_random_seed(img_name, train_dataset_generator.get_seed())),  
    py_trans.RandomCrop(opt.crop_size),  
    py_trans.ToTensor(),  
]  
compose_trans = Compose(transforms_list)  
train_dataset = train_dataset.map(operations=compose_trans, input_columns=["hazy"])  
train_dataset = train_dataset.map(operations=compose_trans, input_columns=["gt"])  
train_dataset = train_dataset.batch(opt.bs, drop_remainder=True)  
    
if __name__ == '__main__':  
    for i in range(2):  
        print(i)  
        for batch in train_dataset.create_dict_iterator():  
            # print(batch)  
            # hazy = Tensor(batch["hazy"], dtype=mindspore.float32)  
            # clear = Tensor(batch["gt"], dtype=mindspore.float32)  
            print(batch["hazy"].shape, batch["gt"].shape)

3 根因分析

脚本中第37行利用魔术方法 def getitem(self, index)定义了随机访问的类，其中index参数作为样本元素索引，外部可以直接使用index来进行数据的访问，外部不需要对其进行修改（如：每个迭代开始前的复位操作）.
但是在代码的第35行户另外自定义了__index，所以需要在每个迭代前进行复位操作，来保证不出现数组越界访问。

4 解决方案

根据根因分析提供的信息，已经非常清楚，就是在另外每次迭代再重新使用__index时，每次迭代完后都进行了

self.__index = -1

造成了下次迭代没有复位的问题，那么可以很简单的让它等于初始值即可。

self.__index += 0

完整代码如下：

import mindspore.dataset as ds  
import matplotlib.pyplot as plt  
import mindspore.dataset.vision.py_transforms as py_trans  
from mindspore.dataset.transforms.py_transforms import Compose  
import mindspore.ops as ops  
# from mindspore import Tensor  
import numpy as np  
from PIL import Image  
import os  
import random  
from option import opt  
    
class DatasetGenerator:  
    def __init__(self, path, train, format='.png'):  
        # self.size = size  
        self.format = format  
        self.train = train  
        if train:  
            self.haze_imgs_dir = os.listdir(os.path.join(path, 'train', 'hazy'))  
            self.haze_imgs = [os.path.join(path, 'train', 'hazy', img) for img in self.haze_imgs_dir]  
            self.clear_dir = os.path.join(path, 'train', 'gt')  
        else:  
            self.haze_imgs_dir = os.listdir(os.path.join(path, 'test', 'hazy'))  
            self.haze_imgs = [os.path.join(path, 'test', 'hazy', img) for img in self.haze_imgs_dir]  
            self.clear_dir = os.path.join(path, 'test', 'gt')  
        # print(self.haze_imgs_dir, self.clear_dir)  
        np.random.seed(58)  
        self.__random_seed = []  
        for _ in range(len(self.haze_imgs)):  
            self.__random_seed.append(random.randint(0, 1000000))  
        self.__index += 0  
    
    def __getitem__(self, index):  
        self.__index += 1  
        haze = Image.open(self.haze_imgs[index])  
        # if isinstance(self.size,int):  
        #     while haze.size[0]<self.size or haze.size[1]<self.size :  
        #         index=random.randint(0,20000)  
        #         haze=Image.open(self.haze_imgs[index])  
        img = self.haze_imgs[index].split('\\')[-1]  
        # img=self.haze_imgs[index].split('/')[-1]  
        img_name = img.split('_')  
        # img_name=img.split('\\')[-1].split('_')  
        # print(img_name)  
        clear_name=f"{img_name[0]}_gt_{img_name[2]}"  
        # print(self.clear_dir, clear_name, os.path.join(self.clear_dir,clear_name))  
        clear=Image.open(os.path.join(self.clear_dir,clear_name))  
        w, h = clear.size  
        nw, nh = haze.size  
        left = (w - nw)/2  
        top = (h - nh)/2  
        right = (w + nw)/2  
        bottom = (h + nh)/2  
        clear = clear.crop((left, top, right, bottom))  
        return (haze, clear, index)  
    
    def __len__(self):  
        print(len(self.haze_imgs))  
        return len(self.haze_imgs)  
    
    def get_seed(self):  
        seed = self.__random_seed[self.__index]  
        return seed  
    
def decode(img):  
    return Image.fromarray(img)  
    
def set_random_seed(img_name, seed):  
    random.seed(seed)  
    return img_name  
    
ds.config.set_seed(8)  
# DATA_DIR = opt.data_url  
DATA_DIR = 'C:\Users\kle\Desktop\NTIRE2021'  
    
train_dataset_generator = DatasetGenerator(DATA_DIR, train=True)  
train_dataset = ds.GeneratorDataset(train_dataset_generator, ["hazy", "gt", "img_name"], shuffle=True)  
test_dataset_generator = DatasetGenerator(DATA_DIR, train=False)  
test_dataset = ds.GeneratorDataset(test_dataset_generator, ["hazy", "gt", "img_name"], shuffle=False)  
transforms_list = [  
    decode,  
    (lambda img_name: set_random_seed(img_name, train_dataset_generator.get_seed())),  
    py_trans.RandomCrop(opt.crop_size),  
    py_trans.ToTensor(),  
]  
compose_trans = Compose(transforms_list)  
train_dataset = train_dataset.map(operations=compose_trans, input_columns=["hazy"])  
train_dataset = train_dataset.map(operations=compose_trans, input_columns=["gt"])  
train_dataset = train_dataset.batch(opt.bs, drop_remainder=True)  
    
if __name__ == '__main__':  
    for i in range(2):  
        print(i)  
        for batch in train_dataset.create_dict_iterator():  
            # print(batch)  
            # hazy = Tensor(batch["hazy"], dtype=mindspore.float32)  
            # clear = Tensor(batch["gt"], dtype=mindspore.float32)  
            print(batch["hazy"].shape, batch["gt"].shape)