The Random Seed

1

Today i use such method to set random seed
The args.seed is default to 1

torch.manual_seed(args.seed)
if args.cuda:
    torch.cuda.manual_seed(args.seed)

But everytime the result is slightly different, why?
I use PyTorch0.4

1 Like
2

Are you using the GPU with cuDNN?
If so, you could set the cuDNN behavior to be deterministic, which would unfortunately trade performance for determinism.

torch.backends.cudnn.deterministic = True

Also, are you using some other libraries sampling random numbers?
If so, you should also seed them.

2 Likes
3

No, i use none of them.
If i haven’t use

import cudnn

means no use of the module of cuDNN,i only use torch and torchvision

Does Xavier Initialization matter?I use such way to init my network.

4

The initializations should yield the same random numbers, if the seed was set.
To see, if you are using cuDNN, use print(torch.backends.cudnn.enabled).

5

Why can’t you write meaningful titles? It makes everyone’s life harder. Please write meaningful titles.

related: