Why global value is reset in __getitem__?

#1

I have a global value in init function. I want to increase it as the number of epoch. However, it reset to zero every epoch. How should I fix it

import torch.utils.data as data

class My_Dataset(data.Dataset):

    def __init__(self, data_list):
        self.rand_num=0
    def __getitem__(self, index):
         self.rand_num +=1

    def __len__(self):
        return len(self.data_list)
#2

Are you talking about self.rand_num? I tried iterating through the dataset using a data_loader, and the value of rand_num keeps increasing:

import torch
import torch.utils.data as data
from torch.utils.data import DataLoader


class My_Dataset(data.Dataset):

    def __init__(self, data_list):
        self.data_list = data_list
        self.rand_num=0

    def __getitem__(self, index):
         self.rand_num +=1
         return torch.tensor([self.data_list[index]])

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

vals = [i for i in range(20)]
ds = My_Dataset(vals)
data_loader = DataLoader(ds, batch_size=10)
print(ds.rand_num)
next(iter(data_loader))
print(ds.rand_num)

for epoch in range(5):
    for batch in data_loader:
        print(epoch, ds.rand_num)

which will print out the following epoch and rand_num:

0
10
0 20
0 30
1 40
1 50
2 60
2 70
3 80
3 90
4 100
4 110
#3

How about print in the getitem?

def __getitem__(self, index):
         self.rand_num +=1
         print (self.rand_num)
#4

Even in __getitem__ it is still fine for me as long as I use num_workers=0 in DataLoader.

But when I use num_workers=1, then it gets reset to zero as you have mentioned. Do you have to use num_workers > 0?

#5

Yes. I used number of worker is 1