updater='grow_gpu_hist' produces unstable results on equivalent but not identical sparce matrices

[mxxun]

Here's a notebook with reproduction: xgboost-unstable-grow-gpu-hist-reproduction.zip

Here's the data: reproduction_data.zip

Does not reproduce on 0.72, but does reproduce on 0.80 and some pretty recent commit that I lost id of; I assume this reproduces on current master, but will edit after verifying.

[hcho3]

Do you see the same problem with CPU hist?

[mxxun]

No; grow_histmaker, grow_fast_histmaker, grow_local_histmaker, grow_colmaker and grow_gpu each give me identical results at those two datasets, and their 0th epoch merrors all match.
grow_skmaker is not stable even on the same dataset, although occasionally its 0th epoch matches with those above.
Incidentally, is there a complete list of updaters somewhere? Docs list some but not all cpu updaters and none of the gpu updaters.

[thvasilo]

Incidentally, is there a complete list of updaters somewhere? Docs list some but not all cpu updaters and none of the gpu updaters.

Good point @mrgutkun, currently it's not clear what's the distinction between the updater and tree_method parameters.

[hcho3]

@thvasilo The tree_method parameter is supposed to cover 90% of use cases. The updater parameter is an "expert" parameter that's less well documented. The documentations says

This is an advanced parameter that is usually set automatically, depending on some other parameters.

So usually it is sufficient to set tree_method.

[hcho3]

Oops clicked wrong button.

[thvasilo]

Thanks for the clarification @hcho3. Out of curiosity are there sanity checks in place that would stop/inform users of passing contradicting values to these parameters?

[hcho3]

@thvasilo #3849 clarifies the behavior of tree_method and updater.

[trivialfis]

Closing in favour of #5023