File size: 40,088 Bytes
d45fdef |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 |
import os
import sys
import json
import pickle as pkl
import logging
import time
import copy
import random
from tqdm import tqdm
import re
import pdb
import string
from collections import Counter
from omegaconf import ListConfig
import multiprocessing
import numpy as np
import torch
from transformers import AutoTokenizer, AutoModel
from sentence_transformers import SentenceTransformer
try:
from pyserini.search.lucene import LuceneSearcher
except:
logging.warning(
"Failed to import pyserini! Please install it from https://github.com/castorini/pyserini/tree/master."
)
import contriever.src.index
import contriever.src.contriever
import contriever.src.utils
import contriever.src.slurm
from contriever.src.evaluation import calculate_matches
import contriever.src.normalize_text
from src.data import load_eval_data
from src.index import (
Indexer,
get_index_dir_and_passage_paths,
get_index_passages_and_id_map,
get_bm25_index_dir,
)
from src.decontamination import check_below_lexical_overlap_threshold
try:
from utils.deduplication import (
remove_duplicates_with_minhash,
multiprocess_deduplication,
)
except:
print("Cannot import from utils")
os.environ["TOKENIZERS_PARALLELISM"] = "true"
device = "cuda" if torch.cuda.is_available() else "cpu"
def embed_queries(args, queries, model, tokenizer, model_name_or_path):
if "sentence-transformers" in model_name_or_path:
all_question = []
for k, q in enumerate(queries):
if args.lowercase:
q = q.lower()
if args.normalize_text:
q = contriever.src.normalize_text.normalize(q)
all_question.append(q)
embeddings = model.encode(
all_question, batch_size=min(128, args.per_gpu_batch_size)
) # sentence-transformer has extra memory overhead and can only support a smaller batch size
else:
model.eval()
embeddings, batch_question = [], []
with torch.no_grad():
for k, q in tqdm(enumerate(queries)):
if args.lowercase:
q = q.lower()
if args.normalize_text:
q = contriever.src.normalize_text.normalize(q)
batch_question.append(q)
if (
len(batch_question) == args.per_gpu_batch_size
or k == len(queries) - 1
):
encoded_batch = tokenizer.batch_encode_plus(
batch_question,
return_tensors="pt",
max_length=args.question_maxlength,
padding=True,
truncation=True,
)
encoded_batch = {k: v.to(device) for k, v in encoded_batch.items()}
output = model(**encoded_batch)
if "contriever" not in model_name_or_path:
output = output.last_hidden_state[:, 0, :]
embeddings.append(output.cpu())
batch_question = []
embeddings = torch.cat(embeddings, dim=0).numpy()
print(f"Questions embeddings shape: {embeddings.shape}")
if args.get("cache_query_embedding", False):
with open(args.query_embedding_save_path, "wb") as fout:
pkl.dump(embeddings, fout)
return embeddings
def validate(data, workers_num):
match_stats = calculate_matches(data, workers_num)
top_k_hits = match_stats.top_k_hits
print("Validation results: top k documents hits %s", top_k_hits)
top_k_hits = [v / len(data) for v in top_k_hits]
message = ""
for k in [5, 10, 20, 100]:
if k <= len(top_k_hits):
message += f"R@{k}: {top_k_hits[k - 1]} "
print(message)
return match_stats.questions_doc_hits
def add_passages(data, passages, top_passages_and_scores, valid_query_idx, domain=None):
# add passages to original data
assert len(valid_query_idx) == len(top_passages_and_scores)
idx = 0
for i, d in enumerate(data):
if i in valid_query_idx:
results_and_scores = top_passages_and_scores[idx]
docs = [passages[doc_id] for doc_id in results_and_scores[0]]
next_docs = [
passages[str(int(doc_id) + 1)]
if int(doc_id) + 1 < len(passages)
else passages[doc_id]
for doc_id in results_and_scores[0]
]
scores = [str(score) for score in results_and_scores[1]]
ctxs_num = len(docs)
d["ctxs"] = [
{
"id": results_and_scores[0][c],
"source": domain,
# "retrieval title": docs[c]["title"],
"retrieval text": docs[c]["text"],
"retrieval next text": next_docs[c]["text"],
"retrieval score": scores[c],
}
for c in range(ctxs_num)
]
idx += 1
else:
d["ctxs"] = [None]
def add_hasanswer(data, hasanswer):
# add hasanswer to data
for i, ex in enumerate(data):
for k, d in enumerate(ex["ctxs"]):
d["hasanswer"] = hasanswer[i][k]
def get_search_output_path(cfg, index_shard_ids):
eval_args = cfg.evaluation
shards_postfix = "_".join([str(shard_id) for shard_id in index_shard_ids])
output_dir = os.path.join(eval_args.eval_output_dir, shards_postfix)
output_path = os.path.join(
output_dir,
os.path.basename(eval_args.data.eval_data).replace(
".jsonl", "_retrieved_results.jsonl"
),
)
return output_path
def get_merged_search_output_path(cfg):
index_args = cfg.datastore.index
eval_args = cfg.evaluation
if isinstance(index_args.index_shard_ids[0], ListConfig):
print(
f"Multi-index mode: building {len(index_args.index_shard_ids)} index for {index_args.index_shard_ids} sequentially..."
)
index_shard_ids_list = index_args.index_shard_ids
else:
print(
f"Single-index mode: building a single index over {index_args.index_shard_ids} shards..."
)
index_shard_ids_list = [index_args.index_shard_ids]
merged_postfix = ""
for index_shard_ids in sorted(index_shard_ids_list, key=lambda x: int(x[0])):
shards_postfix = "_".join([str(shard_id) for shard_id in index_shard_ids])
merged_postfix += "-" + shards_postfix
merged_postfix = merged_postfix.strip("-")
output_dir = os.path.join(eval_args.eval_output_dir, merged_postfix)
output_path = os.path.join(
output_dir,
os.path.basename(eval_args.data.eval_data).replace(
".jsonl", "_retrieved_results.jsonl"
),
)
return output_path
def get_merged_subsampled_search_output_path(cfg):
index_args = cfg.datastore.index
eval_args = cfg.evaluation
if isinstance(index_args.index_shard_ids[0], ListConfig):
print(
f"Multi-index mode: building {len(index_args.index_shard_ids)} index for {index_args.index_shard_ids} sequentially..."
)
index_shard_ids_list = index_args.index_shard_ids
else:
print(
f"Single-index mode: building a single index over {index_args.index_shard_ids} shards..."
)
index_shard_ids_list = [index_args.index_shard_ids]
merged_postfix = ""
for index_shard_ids in sorted(index_shard_ids_list, key=lambda x: int(x[0])):
shards_postfix = "_".join([str(shard_id) for shard_id in index_shard_ids])
merged_postfix += "-" + shards_postfix
merged_postfix = merged_postfix.strip("-")
if cfg.evaluation.search.get("topk_subsample_p", None):
seed = cfg.evaluation.search.get("subsample_seed", 1000)
output_dir = os.path.join(
eval_args.eval_output_dir,
os.path.join(
f"subsampled_{cfg.evaluation.search.topk_subsample_p}_seed_{seed}",
merged_postfix,
),
)
else:
output_dir = os.path.join(eval_args.eval_output_dir, merged_postfix)
output_path = os.path.join(
output_dir,
os.path.basename(eval_args.data.eval_data).replace(
".jsonl", "_retrieved_results.jsonl"
),
)
return output_path
def calculate_recall(pred_ids_and_scores, gt_ids_and_scores):
recalls = []
for pred, gt in zip(pred_ids_and_scores, gt_ids_and_scores):
pred_ids = set(pred[0]) # ['717', '1288', '1283']
gt_ids = set(gt[0]) # ['3029', '3283', '2584']
# Calculate intersection
correct = len(pred_ids.intersection(gt_ids))
# Recall = correct / total ground truth
recall = correct / len(gt_ids)
recalls.append(recall)
# Average recall across all samples
avg_recall = sum(recalls) / len(recalls)
return avg_recall
def search_dense_topk(cfg):
index_args = cfg.datastore.index
eval_args = cfg.evaluation
ds_domain = cfg.datastore.domain
if isinstance(index_args.index_shard_ids[0], ListConfig):
print(
f"Multi-index mode: building {len(index_args.index_shard_ids)} index for {index_args.index_shard_ids} sequentially..."
)
index_shard_ids_list = index_args.index_shard_ids
else:
print(
f"Single-index mode: building a single index over {index_args.index_shard_ids} shards..."
)
index_shard_ids_list = [index_args.index_shard_ids]
all_exist = True
for index_shard_ids in index_shard_ids_list:
# check if all search results exist
output_path = get_search_output_path(cfg, index_shard_ids)
all_exist = all_exist and os.path.exists(output_path)
if all_exist and not eval_args.search.overwrite:
logging.info(
f"All search results for {index_args.index_shard_ids} exist, skipping searching."
)
else:
# load model and evaluation data
logging.info(f"Loading model from: {cfg.model.datastore_encoder}")
model_name_or_path = cfg.model.query_encoder
tokenizer_name_or_path = cfg.model.query_tokenizer
if "contriever" in model_name_or_path:
query_encoder, query_tokenizer, _ = (
contriever.src.contriever.load_retriever(model_name_or_path)
)
elif "dragon" in model_name_or_path:
query_tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path)
query_encoder = AutoModel.from_pretrained(model_name_or_path)
elif "sentence-transformers" in model_name_or_path:
query_tokenizer = None
query_encoder = SentenceTransformer(model_name_or_path)
else:
print(f"{model_name_or_path} is not supported!")
raise AttributeError
query_encoder.eval()
query_encoder = query_encoder.to(device)
if not index_args.no_fp16:
query_encoder = query_encoder.half()
# load eval data
data = load_eval_data(cfg)
# if eval_args.data.num_eval_samples is not None:
# random.seed(eval_args.data.seed)
# data = random.sample(data, int(eval_args.data.num_eval_samples))
queries = []
valid_query_idx = []
for idx, ex in enumerate(data):
raw_query = ex["raw_query"]
if raw_query:
queries.append(ex["raw_query"])
valid_query_idx.append(idx)
logging.info(
f"Searching for {len(queries)} queries from {len(data)} total evaluation samples..."
)
if eval_args.search.get("cache_query_embedding", False) and os.path.exists(
eval_args.search.get("query_embedding_save_path", "")
):
logging.info(
f"Loading query embeddings from {eval_args.search.query_embedding_save_path}"
)
with open(eval_args.search.query_embedding_save_path, "rb") as fin:
questions_embedding = pkl.load(fin)
else:
questions_embedding = embed_queries(
eval_args.search,
queries,
query_encoder,
query_tokenizer,
model_name_or_path,
)
if eval_args.search.get("cache_query_embedding_only", False):
return
# load index
for index_shard_ids in index_shard_ids_list:
output_path = get_search_output_path(cfg, index_shard_ids)
if os.path.exists(output_path) and not eval_args.search.overwrite:
logging.info(f"{output_path} exists, skipping searching.")
else:
copied_data = copy.deepcopy(data)
index_dir, _ = get_index_dir_and_passage_paths(cfg, index_shard_ids)
index = Indexer(
index_args.projection_size,
index_args.n_subquantizers,
index_args.n_bits,
)
index.deserialize_from(index_dir)
# load passages and id mapping corresponding to the index
passages, passage_id_map = get_index_passages_and_id_map(
cfg, index_shard_ids
)
assert len(passages) == index.index.ntotal, (
f"number of documents {len(passages)} and number of embeddings {index.index.ntotal} mismatch"
)
# get top k results
start_time_retrieval = time.time()
top_ids_and_scores = index.search_knn(
questions_embedding, eval_args.search.n_docs
)
logging.info(
f"Search time: {time.time() - start_time_retrieval:.1f} s."
)
# todo: double check valid_query_idx
logging.info(f"Adding documents to eval data...")
add_passages(
copied_data,
passage_id_map,
top_ids_and_scores,
valid_query_idx,
domain=ds_domain,
)
os.makedirs(os.path.dirname(output_path), exist_ok=True)
safe_write_jsonl(copied_data, output_path)
## TODO: check here
if cfg.datastore.index.index_type == "IVF_FLAT":
# replace index_dir from scaling_out/embeddings/facebook/contriever-msmarco/fineweb_edu_1m/1-shards/index/0 to scaling_out/embeddings/facebook/contriever-msmarco/fineweb_edu_1m/1-shards/index_ivf_flat_0
index_dir_ = index_dir.replace("index", "index_ivf_flat")
# remove the last /0
index_dir_ = index_dir[:-2]
from src.indicies.ivf_flat import IVFFlatIndexer
from api.build_ivf_index import build_ivf_flat_index
index_ = build_ivf_flat_index("fineweb_edu_1m", 1, 0)
searched_scores, searched_passages, db_ids = index_.search(
questions_embedding, eval_args.search.n_docs
)
copied_data2 = copy.deepcopy(data)
top_ids_and_scores_ = index_.search_knn(
questions_embedding, eval_args.search.n_docs
)
add_passages(
copied_data2,
passage_id_map,
top_ids_and_scores_,
valid_query_idx,
domain=ds_domain,
)
output_path_ivf = output_path.replace(".jsonl", "_ivf.jsonl")
safe_write_jsonl(copied_data2, output_path_ivf)
# calculate the recall of ivf by compareing top_ids_and_scores and top_ids_and_scores_
recall = calculate_recall(top_ids_and_scores_, top_ids_and_scores)
print(f"recall of ivf is {recall}")
if cfg.evaluation.search.get(
"merge_multi_source_results", False
) and cfg.evaluation.search.get("topk_subsample_p", None):
post_hoc_merge_topk_multi_domain(cfg)
elif cfg.evaluation.search.get("merge_multi_index_results", True):
post_hoc_merge_topk(cfg)
def post_hoc_merge_topk(cfg):
"""
Post hoc merge the searched results obtained by multiple indices.
"""
index_args = cfg.datastore.index
output_path = get_merged_search_output_path(cfg)
if os.path.exists(output_path) and not cfg.evaluation.search.overwrite:
print(f"The merged path exists, skipping...\n{output_path}")
return
if (
isinstance(index_args.index_shard_ids[0], ListConfig)
and len(index_args.index_shard_ids) > 1
):
print(
f"Multi-index mode: building {len(index_args.index_shard_ids)} index for {index_args.index_shard_ids} sequentially..."
)
index_shard_ids_list = index_args.index_shard_ids
else:
print(f"Single-index mode: no need to merge")
return
merged_data = []
for i, index_shard_ids in enumerate(index_shard_ids_list):
path_to_merge = get_search_output_path(cfg, index_shard_ids)
print(f"Adding {path_to_merge}")
data_to_merge = []
with open(path_to_merge, "r") as file:
idx = 0
for line in file:
try:
_ex = json.loads(line)
except:
print(f"Line read error when reading {path_to_merge}")
continue
if not _ex["ctxs"] or _ex["ctxs"][0] is None:
assert idx == 0 # the first example in ppl eval does not have query
ctxs = []
else:
ctxs = _ex["ctxs"]
_ex["ctxs"] = ctxs
data_to_merge.append(_ex)
if i == 0:
merged_data = data_to_merge
else:
for id_, (_, _ex) in enumerate(zip(merged_data, data_to_merge)):
assert merged_data[id_]["raw_query"] == _ex["raw_query"]
merged_data[id_]["ctxs"].extend(_ex["ctxs"])
# Rerank based on score and only keep the top n_docs to avoid memory explosion
if merged_data[id_]["ctxs"] and merged_data[id_]["ctxs"][0] is not None:
merged_data[id_]["ctxs"] = sorted(
merged_data[id_]["ctxs"],
key=lambda x: float(x["retrieval score"]),
reverse=True,
)
merged_data[id_]["ctxs"] = merged_data[id_]["ctxs"][
: cfg.evaluation.search.n_docs
]
# make sure we still have n_docs documents
assert len(merged_data[id_]["ctxs"]) == cfg.evaluation.search.n_docs
else:
assert (
id_ == 0 or id_ == 983
) # the 983rd example in RPJ has an empty query
# Write merged and reranked data to a new JSONL file
os.makedirs(os.path.dirname(output_path), exist_ok=True)
safe_write_jsonl(merged_data, output_path)
def subsample_by_coin_flip(items, probability):
subsampled_list = []
for item in items:
# Perform a coin flip with probability p of being True (keep the item)
if random.random() < probability:
subsampled_list.append(item)
return subsampled_list
def post_hoc_merge_topk_multi_domain(cfg):
"""
Post hoc merge the searched results obtained by multiple domains/sources. Each source may have multiple indices.
Required inputs:
1. A list of searched results to be merged defined by `cfg.evaluation.search.paths_to_merge`
2. A path to save merged results defined by `cfg.evaluation.search.merged_path`
"""
txt_file_with_paths_to_merge = cfg.evaluation.search.paths_to_merge
base_merged_path = cfg.evaluation.search.merged_path
merged_path = os.path.join(
os.path.dirname(base_merged_path),
os.path.basename(base_merged_path).strip("dedup_"),
)
if (
not os.path.exists(base_merged_path)
or not cfg.evaluation.search.use_saved_dedup_data
):
if cfg.evaluation.search.get("topk_subsample_p", 1) < 1:
# Set a random seed for subsampling
seed = cfg.evaluation.search.get("subsample_seed", 1000)
random.seed(seed)
if not os.path.exists(merged_path):
# Read .txt file containing all files of searched results to merge
paths_to_merge = []
with open(txt_file_with_paths_to_merge, "r") as file:
for line in file:
path = line.strip()
paths_to_merge.append(path)
assert os.path.exists(path), f"{path}"
print(f"Merging files:\n{paths_to_merge}")
datastore_domain_pattern = re.compile(r"/([^/]+)_datastore")
merged_data = []
for domain_idx, path_to_merge in tqdm(enumerate(paths_to_merge)):
print(f"Adding {path_to_merge}")
data_to_merge = []
with open(path_to_merge, "r") as file:
# annotate datastore domain for analysis
matches = datastore_domain_pattern.findall(path_to_merge)
ds_domain = matches[0] if matches else None
idx = 0
for line in file:
try:
_ex = json.loads(line)
except:
print(f"Line read error when reading {path_to_merge}")
raise AttributeError
if not _ex["ctxs"] or _ex["ctxs"][0] is None:
assert (
idx == 0
) # the first example in ppl eval does not have query
ctxs = []
else:
if (
not "source" in _ex["ctxs"][0].keys()
or not _ex["ctxs"][0]["source"]
):
for ctx_idx in range(len(_ex["ctxs"])):
_ex["ctxs"][ctx_idx]["source"] = ds_domain
ctxs = _ex["ctxs"]
_ex["ctxs"] = ctxs
data_to_merge.append(_ex)
if domain_idx == 0:
merged_data = data_to_merge
else:
for id_, (_, _ex) in enumerate(zip(merged_data, data_to_merge)):
assert merged_data[id_]["raw_query"] == _ex["raw_query"]
merged_data[id_]["ctxs"].extend(_ex["ctxs"])
# Rerank based on score and only keep the top n_docs to avoid memory explosion
if (
merged_data[id_]["ctxs"]
and merged_data[id_]["ctxs"][0] is not None
):
merged_data[id_]["ctxs"] = sorted(
merged_data[id_]["ctxs"],
key=lambda x: x["retrieval score"],
reverse=True,
)
merged_data[id_]["ctxs"] = merged_data[id_]["ctxs"][
: cfg.evaluation.search.n_docs
]
# make sure we still have n_docs documents
assert (
len(merged_data[id_]["ctxs"])
== cfg.evaluation.search.n_docs
)
else:
assert (
id_ == 0 or id_ == 983
) # the 983rd example in RPJ has an empty query
safe_write_jsonl(merged_data, merged_path)
else:
merged_data = []
with open(merged_path, "r") as fin:
for line in fin:
ex = json.loads(line)
merged_data.append(ex)
# Post-process to remove duplication using multithreading
use_multi_process = True
if use_multi_process:
merged_data = multiprocess_deduplication(merged_data)
else:
for id_, ex in enumerate(merged_data):
merged_data[id_]["ctxs"] = remove_duplicates_with_minhash(
merged_data[id_]["ctxs"],
string_for_decontamination=merged_data[id_]["raw_query"],
)
# merged_data[id_]['ctxs'] = remove_duplicates_with_minhash(merged_data[id_]['ctxs'], string_for_decontamination=None)
# pass
if os.path.exists(base_merged_path) and cfg.evaluation.search.use_saved_dedup_data:
merged_data = []
with open(base_merged_path, "r") as fin:
for line in fin:
ex = json.loads(line)
merged_data.append(ex)
else:
# Write merged and reranked data to a new JSONL file
os.makedirs(os.path.dirname(base_merged_path), exist_ok=True)
safe_write_jsonl(merged_data, base_merged_path)
# Subsample document from B(n_docs, p)
seed = cfg.evaluation.search.get("subsample_seed", 1000)
if cfg.evaluation.search.topk_subsample_p < 1:
# Set a random seed for subsampling
random.seed(seed)
for id_, _ in enumerate(merged_data):
subsampled_ctxs = subsample_by_coin_flip(
merged_data[id_]["ctxs"], cfg.evaluation.search.topk_subsample_p
)
merged_data[id_]["ctxs"] = subsampled_ctxs
# Post-process to rerank
if cfg.evaluation.search.get("rerank_method", None):
rerank_n_docs = cfg.evaluation.search.get("rerank_n_docs", None)
no_enough_rerank_data_cout = 0
for id_, ex in enumerate(merged_data):
merged_data[id_]["ctxs"], no_enough_rerank_data = extract_rerank_docs(
merged_data[id_]["ctxs"], rerank_n_docs
)
no_enough_rerank_data_cout += no_enough_rerank_data
if no_enough_rerank_data_cout:
print(
f"WARNING: there are {no_enough_rerank_data_cout} example having no enough data for reranking!"
)
print(f"Reranking with method: {cfg.evaluation.search.rerank_method}")
if cfg.evaluation.search.rerank_method in [
"lexical",
"unigram_f1",
"inclusion",
]:
all_answers = get_answers(cfg)
for id_, ex in tqdm(enumerate(merged_data)):
query = ex["raw_query"]
merged_data[id_]["ctxs"] = post_rerank_ctxs(
ex["ctxs"], all_answers[query], cfg
)
# Additional decontamination for ablation study
ablation_study = False
if ablation_study:
for id_, _ in enumerate(merged_data):
merged_data[id_]["ctxs"] = additional_decon(merged_data[id_])
# Additional short chunk removal
for id_, _ in enumerate(merged_data):
merged_data[id_]["ctxs"] = additional_remove_short_chunk(
merged_data[id_]["ctxs"]
)
# Check the number of remaining documents
no_enough_data_count = 0
for id_, _ in enumerate(merged_data):
if len(merged_data[id_]["ctxs"]) < 3:
no_enough_data_count += 1
print(
f"WARNING: the subsampled documents only have {len(merged_data[id_]['ctxs'])} left!"
)
# Write merged and reranked data to a new JSONL file
output_path = f"full_subsampled_{str(cfg.evaluation.search.topk_subsample_p)}_{seed}_{os.path.basename(base_merged_path)}"
output_path = os.path.join(os.path.dirname(base_merged_path), output_path)
if cfg.evaluation.search.get("rerank_method", None):
output_path = output_path.replace(
".jsonl", f"_rerank_{cfg.evaluation.search.rerank_method}.jsonl"
)
elif ablation_study:
output_path = output_path.replace(".jsonl", f"_standard_decon.jsonl")
os.makedirs(os.path.dirname(output_path), exist_ok=True)
safe_write_jsonl(merged_data, output_path)
print(
f"Saved merged results to {output_path} with {no_enough_data_count} documents having less than 5 documents."
)
def additional_decon(example):
answer = example["raw_query"]
num_doc_before = len(example["ctxs"])
clean_ctxs = []
for ctx in example["ctxs"]:
# if check_below_lexical_overlap_threshold(ctx['retrieval text'], answer, 8, 'longest'):
if check_below_lexical_overlap_threshold(
ctx["retrieval text"], answer, 0.8, "jaccard"
):
clean_ctxs.append(ctx)
num_doc_after = len(clean_ctxs)
print(f"Additional decon: {num_doc_before - num_doc_after} documents are removed")
return clean_ctxs
def additional_remove_short_chunk(ctxs):
new_ctxs = []
for ctx in ctxs:
if len(ctx["retrieval text"].split(" ")) > 12:
new_ctxs.append(ctx)
return new_ctxs
def extract_rerank_docs(ctxs, rerank_n_docs):
filtered_ctxs = [ctx for ctx in ctxs if ctx["quality score"]]
if rerank_n_docs is None or len(filtered_ctxs) >= rerank_n_docs:
return filtered_ctxs[:rerank_n_docs], 0
else:
return filtered_ctxs, 1
def post_process_ctxs(ctxs):
# + remove ctx that is shorter than 5 words
# + deduplicate ctx with >80% 13-gram overlap
if ctxs[0] is None:
return ctxs
def remove_short_ctx(ctxs):
new_ctxs = []
for ctx in ctxs:
# remove chunks that have less than 5 words
if len(ctx["retrieval text"].split(" ")) > 5:
new_ctxs.append(ctx)
if len(new_ctxs) < 5:
new_ctxs = ctxs[:5]
return new_ctxs
def remove_duplication(ctxs, first_k=5):
new_ctxs = []
num_passed = 0
ctx_idx = 0
while ctx_idx < len(ctxs) and num_passed < first_k:
ctx = ctxs[ctx_idx]
ctx_idx += 1
can_add = True
for added_ctx in new_ctxs:
can_add = check_below_lexical_overlap_threshold(
ctx["retrieval text"],
added_ctx["retrieval text"],
threshold=0.8,
mode="jaccard",
)
if not can_add:
# with open('count_intra_and_inter.txt', 'a') as fout:
# if ctx['source'] == added_ctx['source']:
# fout.write('intra\n')
# else:
# fout.write('inter\n')
break
if can_add:
new_ctxs.append(ctx)
num_passed += 1
new_ctxs = new_ctxs + ctxs[ctx_idx:]
# if len(new_ctxs) < 5:
# pdb.set_trace()
return new_ctxs
return remove_duplication(remove_short_ctx(ctxs))
def post_rerank_ctxs(ctxs, answers, cfg):
rerank_method = cfg.evaluation.search.rerank_method
good_ctxs = [ctx for ctx in ctxs if ctx["quality score"]]
bad_ctxs = [ctx for ctx in ctxs if not ctx["quality score"]]
assert len(good_ctxs) + len(bad_ctxs) == len(ctxs)
if rerank_method == "lexical":
good_ctxs = lexical_rerank(good_ctxs, answers)
elif rerank_method == "inclusion":
good_ctxs = inclusion_rerank(good_ctxs, answers)
elif rerank_method == "unigram_f1":
good_ctxs = unigram_f1_rerank(good_ctxs, answers)
return good_ctxs + bad_ctxs
def get_answers(cfg):
if cfg.tasks.eval.task_name == "perplexity":
eval_data = load_eval_data(cfg)
all_answers = []
for ex in eval_data:
answer = extract_ppl_answer(ex["raw_inputs"], ex["raw_query"])
all_answers.append([answer])
elif cfg.tasks.eval.task_name == "lm-eval":
answer_path = cfg.evaluation.search.answer_path
all_answers = {}
with open(answer_path, "r") as fin:
for line in fin:
ex = json.loads(line)
if "triviaqa" in answer_path:
answer = {ex["query"]: ex["answer"]["normalized_aliases"]}
elif "nq_open" in answer_path:
answer = {ex["query"]: ex["answer"]}
else:
answer = {ex["query"]: ex["answer"]}
all_answers.update(answer)
return all_answers
def extract_ppl_answer(raw_input, raw_query):
inputs = raw_input.replace("<|endoftext|>", "")
query = raw_query.replace("<|endoftext|>", "")
try:
answer = inputs.replace(query, "")
except:
try:
answer = inputs.replace(query[:-1], "")
except:
answer = inputs[-len(inputs) // 2 :]
return answer
def inclusion_rerank(ctxs, answers):
inclusion_scores = [
inclusion_metric(ctx["retrieval text"], answers) for ctx in ctxs
]
ctxs = sort_ctxs_with_1_scores(ctxs, inclusion_scores)
return ctxs
def unigram_f1_rerank(ctxs, answers):
unigram_f1_scores = [
unigram_f1_metric(ctx["retrieval text"], answers) for ctx in ctxs
]
ctxs = sort_ctxs_with_1_scores(ctxs, unigram_f1_scores)
return ctxs
def lexical_rerank(ctxs, answers):
if not ctxs or ctxs[0] is None:
return ctxs
inclusion_scores = [
inclusion_metric(ctx["retrieval text"], answers) for ctx in ctxs
]
unigram_f1_scores = [
unigram_f1_metric(ctx["retrieval text"], answers) for ctx in ctxs
]
retrieval_scores = [ctx["retrieval score"] for ctx in ctxs]
ctxs = sort_ctxs_with_3_scores(
ctxs, inclusion_scores, unigram_f1_scores, retrieval_scores
)
return ctxs
def inclusion_metric(ctx, answers):
if not ctx or not answers:
return 0
score_list = []
for answer in answers:
score = 1 if normalize_text(answer) in normalize_text(ctx) else 0
score_list.append(score)
return max(score_list)
def unigram_f1_metric(ctx, answers):
if not ctx or not answers:
return 0
norm_answers = [normalize_text(ans) for ans in answers]
norm_ctx = normalize_text(ctx)
common_tokens = [
Counter(norm_ctx.split()) & Counter(norm_ans.split())
for norm_ans in norm_answers
]
num_same = [sum(common.values()) for common in common_tokens]
score_list = []
for i, num in enumerate(num_same):
if num == 0:
score_list.append(0.0)
else:
p = 1.0 * num / len(norm_ctx.split())
r = 1.0 * num / len(norm_answers[i].split())
f1 = 2 * p * r / (p + r)
score_list.append(f1)
return max(score_list)
def sort_ctxs_with_1_scores(ctxs, scores_1):
combined_list = list(zip(scores_1, ctxs))
combined_list.sort(key=lambda x: x[0], reverse=True)
sorted_ctxs = [ctx for _, ctx in combined_list]
return sorted_ctxs
def sort_ctxs_with_3_scores(ctxs, scores_1, scores_2, scores_3):
combined_list = list(zip(scores_1, scores_2, scores_3, ctxs))
combined_list.sort(key=lambda x: x[2], reverse=True)
combined_list.sort(key=lambda x: x[1], reverse=True)
combined_list.sort(key=lambda x: x[0], reverse=True)
sorted_ctxs = [ctx for _, _, _, ctx in combined_list]
return sorted_ctxs
def normalize_text(text):
def remove_articles(text):
return re.sub(r"\b(a|an|the)\b", " ", text)
def white_space_fix(text):
return " ".join(text.split())
def lower(text):
return text.lower()
return white_space_fix(remove_articles(lower(text)))
def search_sparse_topk(cfg):
index_args = cfg.datastore.index
eval_args = cfg.evaluation
if isinstance(index_args.index_shard_ids[0], ListConfig):
print(
f"Multi-index mode: building a BM25 index over {len(index_args.index_shard_ids)} shards..."
)
index_shard_ids_list = [
i for index_shards in index_args.index_shard_ids for i in index_shards
]
else:
print(
f"Single-index mode: building a BM25 index over {index_args.index_shard_ids} shards..."
)
index_shard_ids_list = index_args.index_shard_ids
# check if all search results exist
output_path = get_search_output_path(cfg, index_shard_ids_list)
all_exist = os.path.exists(output_path)
if all_exist and not eval_args.search.overwrite:
logging.info(
f"All search results for {index_args.index_shard_ids} exist, skipping searching."
)
else:
# load eval data
data = load_eval_data(cfg)
logging.info(f"Searching for {len(data)} total evaluation samples...")
# load index
bm25_index_path = os.path.join(
get_bm25_index_dir(cfg, index_shard_ids_list), "index"
)
assert os.path.exists(bm25_index_path), (
f"The index path does not exist, please build the index first\nMissing: {bm25_index_path}"
)
logging.info(f"Loading BM25 index from {bm25_index_path}")
searcher = LuceneSearcher(bm25_index_path)
for ex in tqdm(data):
query = ex["raw_query"]
if query:
hits = searcher.search(query, cfg.evaluation.search.n_docs)
# ctxs = []
# for i in range(len(hits)):
# raw = searcher.doc(hits[i].docid).raw()
# ex = json.loads(raw)
# ctxs.append(
# {
# "id": int(ex["id"]),
# "retrieval text": ex["contents"],
# "retrieval score": hits[i].score,
# } for i in range(len(hits))
# )
# if len(hits) < cfg.evaluation.search.n_docs: # will there be any case where len(hits) < n_docs?
# dummy_ctx = {"id": None, "retrieval text": '', "retrieval score": float('-inf')}
# ctxs += [dummy_ctx] * (cfg.evaluation.search.n_docs - len(hits))
# print(f"The number of retrieved documents is less than n_docs: {len(hits)} < {cfg.evaluation.search.n_docs}")
ex["ctxs"] = [
{
# "id": int(ex["id"]),
"retrieval text": json.loads(searcher.doc(hits[i].docid).raw())[
"contents"
],
"retrieval score": hits[i].score,
}
for i in range(len(hits))
]
else:
ex["ctxs"] = [None]
os.makedirs(os.path.dirname(output_path), exist_ok=True)
safe_write_jsonl(data, output_path)
def safe_write_jsonl(data, output_file):
success = False
try:
with open(output_file, "w") as fout:
for ex in data:
fout.write(json.dumps(ex) + "\n")
success = True
logging.info(f"Saved results to {output_file}")
except Exception as e:
print(f"An error occurred: {e}")
finally:
# If an error was raised, and success is still False, delete the file
if not success and os.path.exists(output_file):
os.remove(output_file)
print(f"File '{output_file}' has been deleted due to an error.")
def search_topk(cfg):
if cfg.model.get("sparse_retriever", None):
search_sparse_topk(cfg)
else:
search_dense_topk(cfg)
|