Revisit random seed options in ANTs

[cookpa]

Following discussion in #1462

I believe reproducibility can be achieved by setting the env var ANTS_RANDOM_SEED and using single-threaded computation to avoid numerical issues in the mutual information metric.

But the usage of some scripts might be misleading, eg in antsBrainExtraction.sh, "-u 0" fixes the Atropos random seed, but doesn't set a random seed to be used by all the other processes.

[cookpa]

Hi @ntustison ,

Thinking about this some more, I think antsAI wasn't the cause of the issue you were seeing recently, because it defaults to a fixed seed:

ANTs/Examples/antsAI.cxx

Lines 1245 to 1265 in 74afcc5

	int antsRandomSeed = 1234;
	itk::ants::CommandLineParser::OptionType::Pointer randomSeedOption = parser->GetOption("random-seed");
	if (randomSeedOption && randomSeedOption->GetNumberOfFunctions())
	{
	antsRandomSeed = parser->Convert<int>(randomSeedOption->GetFunction(0)->GetName());
	}
	else
	{
	char * envSeed = getenv("ANTS_RANDOM_SEED");
	if (envSeed != nullptr)
	{
	antsRandomSeed = std::stoi(envSeed);
	}
	}
	if (antsRandomSeed != 0)
	{
	randomizer->SetSeed(antsRandomSeed);
	}

I think it was antsRegistration being variable. I'm thinking for now that we could set a fixed seed inside the script if "-u 0" is detected.

Longer term, it would be good to make all the programs consistent in how they handle random number generation. antsRegistration does, in decreasing order of priority,

1. Command line --random-seed uses
2. $ANTS_RANDOM_SEED
3. System time

Having this be consistent for all tools would be nice because it removes the need to add a random option explicitly to every executable call. I guess one criticism is that you could run a very complex pipeline with the same random numbers over and over, but I don't know if anyone cares to use ANTS_RANDOM_SEED outside of regression testing anyway.

[cookpa]

Solution in #1468 aims to maximize backwards compatibility and hopefully won't break anyone's regression tests

[TheChymera]

What's the situation with recent-ish ANTs (2.4.3)? is setting a global ANTS_RANDOM_SEED value analogous to using --random-seed for each antsRegistration call?

[cookpa]

@TheChymera yes, ANTS_RANDOM_SEED should be used unless --random-seed is specified at run time.

[TheChymera]

@cookpa thank you :)

[cookpa]

The last part of my to-do list for this issue was antsLongitudinalCorticalThickness.sh, but because it uses ANTS, it doesn't give identical results even if ANTS_RANDOM_SEED is set. I tried fixing the seed here

ANTs/ImageRegistration/ANTS_affine_registration.h

Lines 260 to 261 in 3fa6cfb

	unsigned int time_seed = (unsigned int)time(nullptr);
	srand(time_seed);

but it was insufficient.

Since deterministic results aren't a documented feature of the longitudinal script, I'm inclined to leave it alone.