Repository navigation
Replies: 2 comments 2 replies
|
On your follow-up about the final checkpoint: it can preserve the final model, but it cannot reconstruct the training path that produced it. Setting and recording a seed now controls future random choices; it does not reveal the earlier initialization, batch order, augmentation, or dropout choices. There are three different goals:
The PyTorch checkpoint guide explains why resuming requires more than the model's |
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
I have a few questions regarding seed and reproducibility. It will be very helpful if someone gives me the answers regarding this matter.
Can I assign any random number when I am creating a seed? Like, can I use seed = 1 or seed = 456? Will it make any difference due to different numbers? And if a model is trained without assigning any seed, is it possible to reproduce the same result without a seed? If it is possible, how to do that? If it is not possible, how can it be possible to reproduce the same results since the model is not trained without assigning any seed?
All reactions