Member anonymous_namespace{sparse.cpp}::check_positive (long v, const char *error_msg): we have this code twice now.
Member anonymous_namespace{xmc.cpp}::count_features_per_example (std::istream &source, std::size_t num_examples=100 '000): also count number of labels based on ,, then we can reserve also the label vector
Member anonymous_namespace{xmc.cpp}::parse_xmc_header (const std::string &content): throw if there is more data on the line.
Member dismec::BinaryLabelVector: decide whether this should be a matrix or an array type
Class dismec::DiSMECTraining: Figure out why we do this and put a reference/explanation here.
Member dismec::io::model::PartialModelSaver::finalize (): Do we need this function? It seems like a nice idea to have a check if everything is done, but currently we don't use it.
Member dismec::io::prediction::read_sparse_prediction (std::istream &source): document this format and provide a link.
Member dismec::io::save_xmc_dataset (std::ostream &target, const MultiLabelData &data): insert proper checks that data is sparse
Class dismec::model::Model: what about per-label prediction? That would also be a valid axis for parallelizing.
Member dismec::objective::Objective::declare_vector_on_last_line (const HashVector &location, real_t t): improve this interface, together with project_to_line, to be less error prone!
Member dismec::stats::make_stat_from_json (const nlohmann::json &source): document the json format.
Class dismec::TrainingSpec: should we give the dataset to each operation, or maybe just set it in the beginning? I think maybe at some point convert all this dataset stuff to use shared_ptr everywhere.
Member TEST_CASE ("save dense npy"): test cases for mismatch
Member TEST_CASE ("parse labels"): we should also check that the returned pointer is valid.