base_margin not copied by XGDMatrixSliceDMatrix. Leads to loss of base_margin when performing cross-validation with xgb.cv()

[osofr]

On latest version of the R package.

base_margin is not passed down to test / train data during cross-validation.

On the other hand other setinfo attributes such as weights and labels are passed down and handled properly (split into test / train sets).

Most likely the problem occurs when calling slice in xgb.cv, which can be traced down to these lines, where we can see that weights and labels are copied, but not base_margin. Presumabely this is a very quick fix, would it be possible to add a line to copy base_margin?

xgboost/src/c_api/c_api.cc

Lines 397 to 405 in 7078c41

	if (src.info.labels.size() != 0) {
	ret.info.labels.push_back(src.info.labels[ridx]);
	}
	if (src.info.weights.size() != 0) {
	ret.info.weights.push_back(src.info.weights[ridx]);
	}
	if (src.info.root_index.size() != 0) {
	ret.info.root_index.push_back(src.info.root_index[ridx]);
	}

[tqchen]

hmm, can you open a PR to fix this? Thanks!

[osofr]

I can try, but I hardly know any C++ :)

[owlas]

I just ran into this in the current version on pypi.

• Is this fix in the most recent release?
• Why was the issue reopened?

[khotilov]

It was fixed in the development but so far it got into neither pypi nor CRAN release.

[owlas]

@khotilov thanks. Is there a schedule for next pypi release?