Fix an off by one issue affecting number of iterations in multiple optimizers.

[ranjodhsingh1729]

Description

This PR fixes the following issue in multiple optimizers:

Example

If maxIterations is non-zero it will perform one less iteration.

ensmallen/include/ensmallen_bits/gradient_descent/gradient_descent_impl.hpp

Line 70 in 9aebc1d

for (size_t i = 1; i != maxIterations && !terminate; ++i)

[rcurtin]

Thanks; this is a really minor fix but I definitely don't mind adding it. There is a consistency issue worth thinking about here: there are a number of optimizers that use code exactly like this (before your fix), and there are also some that don't:

$ grep -r 'for (size_t i = 1' ~/src/ensmallen/include/
/home/ryan/src/ensmallen/include/ensmallen_bits/gradient_descent/gradient_descent_impl.hpp:  for (size_t i = 1; i != maxIterations && !terminate; ++i)
/home/ryan/src/ensmallen/include/ensmallen_bits/fw/frank_wolfe_impl.hpp:  for (size_t i = 1; i != maxIterations && !terminate; ++i)
/home/ryan/src/ensmallen/include/ensmallen_bits/fw/constr_structure_group.hpp:    for (size_t i = 1; i <= nGroups; ++i)
/home/ryan/src/ensmallen/include/ensmallen_bits/fista/fista_impl.hpp:  for (size_t i = 1; i != maxIterations && !terminate; ++i)
/home/ryan/src/ensmallen/include/ensmallen_bits/parallel_sgd/parallel_sgd_impl.hpp:  for (size_t i = 1; i != maxIterations && !terminate; ++i)
/home/ryan/src/ensmallen/include/ensmallen_bits/nsga2/nsga2_impl.hpp:    for (size_t i = 1; i < fSize - 1; i++)
/home/ryan/src/ensmallen/include/ensmallen_bits/iqn/iqn_impl.hpp:  for (size_t i = 1; i != maxIterations && !terminate; ++i)
/home/ryan/src/ensmallen/include/ensmallen_bits/cne/cne_impl.hpp:  for (size_t i = 1; i < populationSize; i++)
/home/ryan/src/ensmallen/include/ensmallen_bits/cmaes/active_cmaes_impl.hpp:  for (size_t i = 1; (i != maxIterations) && !terminate; ++i)
/home/ryan/src/ensmallen/include/ensmallen_bits/cmaes/cmaes_impl.hpp:  for (size_t i = 1; (i != maxIterations) && !terminate; ++i)
/home/ryan/src/ensmallen/include/ensmallen_bits/fasta/fasta_impl.hpp:  for (size_t i = 1; i != maxIterations && !terminate; ++i)
/home/ryan/src/ensmallen/include/ensmallen_bits/cd/cd_impl.hpp:  for (size_t i = 1; i != maxIterations && !terminate; ++i)
/home/ryan/src/ensmallen/include/ensmallen_bits/fbs/fbs_impl.hpp:  for (size_t i = 1; i != maxIterations && !terminate; ++i)

So it would make sense to fix all of them, but also, there is the inconsistency of iteration number: this gradient descent optimizer will print "iteration 1" on the first iteration, but, e.g., BigBatchSGD will print "iteration 0" on its first iteration. I think it would be easiest to just set the first iteration to 0 instead of 1.

Anyway, up to you whether you want to fix just this one or the other ones I grepped for too. Either way is an improvement. 👍

[ranjodhsingh1729]

Anyway, up to you whether you want to fix just this one or the other ones I grepped for too. Either way is an improvement.

No issue! I'll add the others.

[rcurtin]

Thanks for the fix. Do you want to add something to the changelog? Otherwise I will do it on merge. All my suggestions are just tiny style fixes.

[github-actions]

Second approval provided automatically after 24 hours. 👍