Problem with DataLoader

[suchithtuple]

🐛 Bug

To Reproduce

I have training , validation and test dataset. I have created dataloaders for three of them with training shuffle = True and for valdiation and test shuffle = False . So in one of the code epoch loop contains iterator over train and validation and in another one , epoch loop contains iterator over train , validation and test.

Check the output of the code 1 and 2

1. https://www.kaggle.com/suchith0312/pytorch-dataloader-testing?scriptVersionId=14425205
2. https://www.kaggle.com/suchith0312/pytorch-dataloader-testing?scriptVersionId=14425182

Expected behavior

The train id's from output of files must be same but they are different.

Environment

• PyTorch Version : 1.0.1.post2
• Python version: 3.6.6
• CUDA version : 10.0
• Cudnn version 7.4.2

[ssnl]

Your iteration over test advances rng state.

[suchithtuple]

Your iteration over test advances rng state.

How can I prevent that or there is no way to get the outputs of two codes same.

[InnovArul]

A possible solution could be: #20749