o }oi@sTddlmZddlZddlZddlmZdd d ZdddZdddZ dddZ dS)) annotationsN)compute_hypervolumerank_i_loss_vals np.ndarrayrank_i_indices subset_sizeintreference_pointreturncCs:|jddkr||jdksJ|jd}t|jd}|}tj|tjddf|dd}tj|td}t|D]_} tj ||dd} t | } ||| || <|| } tj || t d} d| | <|| }|| }|| }t | d|d| df|d| df<t | d|| ddf|| ddf<q;|S)NraxisdtypeF)shapenparangecopyrepeatnewaxiszerosrrangeprodargmaxonesboolminimum)rrrr n_trialssorted_indicessorted_loss_vals rect_diagsselected_indicesicontribs max_index loss_valskeepr)L/home/ubuntu/.local/lib/python3.10/site-packages/optuna/_hypervolume/hssp.py_solve_hssp_2d s&     (*r+r%pareto_loss_values selected_vecs hv_selectedfloatc Cs&t|r t|tjSt|ddtjf|dd}tj||dd}t|}t||tj||dddfdd}d}|j ddk} t | D]A} || r]tj}|| <qO|| |krdqO| r|||  |d<t ||dd} | ||| <n || t || ||| <t || |}qO|S) aLazy update the hypervolume contributions. (1) Lazy update of the hypervolume contributions S=selected_indices - {indices[max_index]}, T=selected_indices, and S' is a subset of S. As we would like to know argmax H(T v {i}) in the next iteration, we can skip HV calculations for j if H(T v {i}) - H(T) > H(S' v {j}) - H(S') >= H(T v {j}) - H(T). We used the submodularity for the inequality above. As the upper bound of contribs[i] is H(S' v {j}) - H(S'), we start to update from i with a higher upper bound so that we can skip more HV calculations. (2) A simple cheap-to-evaluate contribution upper bound The HV difference only using the latest selected point and a candidate is a simple, yet obvious, contribution upper bound. Denote t as the latest selected index and j as an unselected index. Then, H(T v {j}) - H(T) <= H({t} v {j}) - H({t}) holds where the inequality comes from submodularity. We use the inclusion-exclusion principle to calculate the RHS. Nr rr gT) assume_pareto)mathisinfr full_likeinfmaximumrrrrargsortrrmax) r%r,r-r r.intersec inclusive_hvsis_contrib_inf max_contribis_hv_calc_fastr$hv_plusr)r)r*_lazy_contribs_update-s. " " r?cCsJt|s |d|S|j|kr|S|jddkr"t||||S||jks)J||}|j\}}tj|dd}tj|td}t ||f} t |} d} t |D]N} tt |} | || 7} | | || <||  | | <tj|jtd}d|| <||}| |} ||}| |dkr||St||| d| d|| }qR||S)Nr r r rrFr)risfiniteallsizerr+rrremptyrrrrrrr?)rrrr diff_of_loss_vals_and_ref_point n_solutions n_objectivesr%r#r-indiceshvkr&r(r)r)r*_solve_hssp_on_unique_loss_valsds>        rJc Cs||jkr|Stj|ddd\}}|j}||kr;tj|jtd}d||<t|j|}d||d||<||St||||} || S)ahSolve a hypervolume subset selection problem (HSSP) via a greedy algorithm. This method is a 1-1/e approximation algorithm to solve HSSP. For further information about algorithms to solve HSSP, please refer to the following paper: - `Greedy Hypervolume Subset Selection in Low Dimensions `__ Tr) return_indexrrN)rBruniquerrrrJ) rrrr rank_i_unique_loss_valsindices_of_unique_loss_valsn_uniquechosenduplicated_indices$selected_indices_of_unique_loss_valsr)r)r* _solve_hssps  rS) rrrrrrr rr r) r%rr,rr-rr rr.r/r r) __future__rr2numpyroptuna._hypervolume.wfgrr+r?rJrSr)r)r)r*s    # 7+