Why does the random seed need many ones?

[fepegar]

The docs for monai.utils.misc.set_determinism say:

It is recommended to set a large seed, i.e. a number that has a good balance of 0 and 1 bits. Avoid having many 0 bits in the seed.

Two things confuse me here:

1. Why is a good balance between 0s and 1s better? Why is having many 0s bad? I haven't found any sources to support these claims. Moreover, 0 is the most common random seed (source) and some important libraries use it happily in their examples.
2. A large seed doesn't imply a good balance, nor a lot of ones. The current default is np.iinfo(np.uint32).max, which is 4294967295 or 11111111111111111111111111111111. Another large number is 2147483648 or 10000000000000000000000000000000. The default kwarg is not consistent with the docstring.

[Nic-Ma]

Hi @wyli ,

Could you please help share some comments about this question？

Thanks in advance.

[wyli]

Thanks, for item 1, the docstring is probably from pytorch https://pytorch.org/docs/stable/generated/torch.Generator.html#torch.Generator.manual_seed
and this "Good Practice in (Pseudo) Random Number Generation for
Bioinformatics Applications" suggests:

This is important if your seed
values have very low entropy (or more simply they have too many runs of 0 or 1 bits in their binary
representation) – some otherwise good RNGs (especially shift-register-based generators such as MT)
are not very random from their initial state when certain seed values are used. 

I think item 2 is a valid issue and I don't know the reason of the current default... following the good practice doc, we might want to 'warm up' the RNGs. cc @ericspod @rijobro

[rijobro]

Interesting read, thanks @wyli! Looks like warming up might be the simplest solution, we could an extra argument to the set_determinism/set_random_state functions. The only tricky thing I see initially is that we'd have to warm up all of our random libraries equally (python, np and torch).

[fepegar]

Thanks, Wenqi. I think PyTorch docs on this are not good, then. Given that no one seems to care much about this (even Soumith Chintala), it might be an unnecessary worry. Shall we ask the PyTorch devs about this issue?