_pickle.PicklingError, in training networks with num_workers positive

Question

I was training a big dataset, using pytorch Dataset and DataLoader, and this problem appeared. I deleted much of my code to ask this question, and the problem exists, these code can run independently:

import numpy as np
from torch.utils.data import DataLoader, Dataset

class PGDataset(Dataset):
    def __init__(self, X, y):
        self.data_x = X
        self.data_y = y

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

    def __getitem__(self, item):
        return self.data_x[item], self.data_y[item]


def train(dataloader):
    epochs = 5
    for epoch in range(epochs):
        for i, (data_x, data_y) in enumerate(dataloader):
            pass

    print('Train ended successfully.')


if __name__ == '__main__':
    X, y = np.random.rand(200, 4), np.random.randint(0, 5, 200, dtype='int64')
    train_set = PGDataset(X, y)
    dataloader = DataLoader(train_set, batch_size=50, shuffle=True, num_workers=4)

    train(dataloader)

These code seems very standerd for me, and I do not know the reason for this problem.

I am using python==3.9.7, numpy==1.21.5, pytorch==1.10.1. By the way, I just updated pycharm to 2022.1.

I have tryed this for all day, and searched in many websites, I found:

1. If I set num_workers in dataloader to 0, it's ok. But I need it not to be 0 for a big dataset.
2. In pycharm debug mode, there's also no mistakes.

Why this happen?