CPU evaluation for cat data. #7393
Conversation
* Implementation for one hot based. * Implementation for partition based.
| @@ -5,6 +5,13 @@ | |||
| #include "../../../src/common/quantile.cuh" | |||
|
|
|||
There was a problem hiding this comment.
Some issues in the quantile test were found after changing the SimpleLCG, so the changes in this file are related.
| @@ -91,49 +96,118 @@ template <typename GradientSumT, typename ExpandEntry> class HistEvaluator { | |||
| iend = static_cast<int32_t>(cut_ptr[fidx]) - 1; | |||
| } | |||
|
|
|||
| auto calc_bin_value = [&](auto i) { | |||
There was a problem hiding this comment.
This can be split into multiple functions, but then we will have lots of duplicated code.
| std::vector<size_t> sorted_idx(n_bins); | ||
| std::iota(sorted_idx.begin(), sorted_idx.end(), 0); | ||
| auto feat_hist = histogram.subspan(cut_ptr[fidx], n_bins); | ||
| std::stable_sort(sorted_idx.begin(), sorted_idx.end(), [&](size_t l, size_t r) { |
There was a problem hiding this comment.
Can't use argsort as we don't have a cpu transform iter.
| @@ -39,7 +36,7 @@ template <typename GradientSumT> void TestEvaluateSplits() { | |||
| std::iota(row_indices.begin(), row_indices.end(), 0); | |||
| row_set_collection.Init(); | |||
|
|
|||
| auto hist_builder = GHistBuilder<GradientSumT>(n_threads, gmat.cut.Ptrs().back()); | |||
| auto hist_builder = GHistBuilder<GradientSumT>(omp_get_max_threads(), gmat.cut.Ptrs().back()); | |||
There was a problem hiding this comment.
Shouldn't we call function OmpGetThreadLimit() to query OMP_THREAD_LIMIT?
There was a problem hiding this comment.
It's a c++ unittest, I think we don't have to worry about that too much. But on the other hand, we need more thorough integration tests for that env, probably in another PR.
| size_t n_cats{8}; | ||
|
|
||
| auto dmat = | ||
| RandomDataGenerator(kRows, kCols, 0).Seed(3).Type(ft).MaxCategory(n_cats).GenerateDMatrix(); |
There was a problem hiding this comment.
I really like how we can generate random data with a fluent interface.
Currently, the evaluation function is using parameters from numerical split. We should set up a new set of training parameters later. But before that, I want to have a working implementation and some experiment results first.
Extracted from #7214 .