Is it OK to change dataset during training?

#1

I use dataset and dataloader for my training script.

Is it OK to change dataset parameter during the traininng??

For example,

dataset = CustomDataset()
loader = DataLoader(dataset, num_worker=8)

for i, data in enumerate(loader):
     dataset.x = new_x

I saw that DataLoader use multithread and dataset is copied to other thread.

The change of dataset is reflected to the loader?

Thanks,

#2

No, if you are using multiple workers, the changes won’t be reflected.
What is your use case?
Maybe you could use shared arrays?

#3

My use case is for dataset to create sample using sample function.

During training, sampling strategy is changed.

I think I have to find alternative way… :frowning:

#4

Hello, what if I change dataloader.dataset.x like in this post?


Will that be reflected to multiple workers?

#5

I don’t think so as stated in my reply.