from surrogate import train from utils import get_args from influence_score import get_influence_score from effort_score import get_effort_score import torch import math import random if __name__ == '__main__': args = get_args() trainer = train(args) influence = get_influence_score(args, trainer) effort = get_effort_score(args) # normalization influence_norm = (influence-torch.min(influence))/(torch.max(influence)-torch.min(influence)) effort_norm = (effort-torch.min(effort))/(torch.max(effort)-torch.min(effort)) # overall score overall = influence_norm + args.lamda * effort_norm scores_sorted, indices = torch.sort(overall, descending=True) # coverage-enhanced sample selection n_prune = math.floor(args.hard_prune * len(scores_sorted)) scores_sorted = scores_sorted[n_prune:] indices = indices[n_prune:] print(f"** after hard prune with {args.hard_prune*100}% data:", len(scores_sorted)) # split scores into k ranges s_max = torch.max(scores_sorted) s_min = torch.min(scores_sorted) print("== max socre:", s_max) print("== min score:", s_min) interval = (s_max - s_min) / args.k s_split = [min(s_min + (interval * _), s_max)for _ in range(1, args.k+1)] score_split = [[] for _ in range(args.k)] for idxx, s in enumerate(scores_sorted): for idx, ref in enumerate(s_split): if s.item() <= ref: score_split[idx].append({indices[idxx].item():s.item()}) break coreset = [] m = args.n_fewshot while len(score_split): # select the group with fewest samples group = sorted(score_split, key=lambda x:len(x)) if len(group) > 3: print("** sorted group length:", len(group[0]), len(group[1]), len(group[2]), len(group[3]),"...") group = [strat for strat in group if len(strat)] if len(group) > 3: print("** sorted group length after removing empty ones:", len(group[0]), len(group[1]), len(group[2]), len(group[3]),"...") budget = min(len(group[0]), math.floor(m/len(group))) print("** budget for current fewest group:", budget) # random select and add to the fewshot indices list fewest = group[0] selected_idx = random.sample([list(_.keys())[0] for _ in fewest], budget) coreset.extend(selected_idx) # remove the fewest group score_split = group[1:] m = m - len(selected_idx) print(f"** finish selecting {len(coreset)} samples.") torch.save(coreset, f"selected/{args.data}_{args.n_fewshot}.pt") # Alternatively, you may wish to save the lambda, group number, and hard prune ratio for hyper-parameter tuning.