File size: 11,306 Bytes
74b1bac |
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 |
import json
import os
import time
import unittest
from pathlib import Path
import warnings
import numpy as np
import bm25s
# Make sure to import or define the functions/classes you're going to use,
# such as bm25s.skl_tokenize and the bm25s.BM25 class, among others.
def save_scores(scores, artifact_dir="tests/artifacts"):
if os.getenv("ARTIFACTS_DIR"):
artifacts_dir = Path(os.getenv("BM25_ARTIFACTS_DIR"))
elif artifact_dir is not None:
artifacts_dir = Path(artifact_dir)
else:
artifacts_dir = Path(__file__).parent / "artifacts"
if "dataset" not in scores:
raise ValueError("scores must contain a 'dataset' key.")
if "model" not in scores:
raise ValueError("scores must contain a 'model' key.")
artifacts_dir = artifacts_dir / scores["model"]
artifacts_dir.mkdir(exist_ok=True, parents=True)
filename = f"{scores['dataset']}-{os.urandom(8).hex()}.json"
with open(artifacts_dir / filename, "w") as f:
json.dump(scores, f, indent=2)
class BM25TestCase(unittest.TestCase):
def compare_with_rank_bm25(
self,
dataset,
artifact_dir="tests/artifacts",
rel_save_dir="datasets",
corpus_subsample=None,
queries_subsample=None,
method="rank",
):
from beir.datasets.data_loader import GenericDataLoader
from beir.util import download_and_unzip
import rank_bm25
import Stemmer
warnings.filterwarnings("ignore", category=ResourceWarning)
if method not in ["rank", "bm25+", "bm25l"]:
raise ValueError("method must be either 'rank' or 'bm25+'.")
# Download and prepare dataset
base_url = (
"https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/{}.zip"
)
url = base_url.format(dataset)
out_dir = Path(__file__).parent / rel_save_dir
data_path = download_and_unzip(url, str(out_dir))
corpus, queries, qrels = GenericDataLoader(data_folder=data_path).load(
split="test"
)
# Convert corpus and queries to lists
corpus_lst = [val["title"] + " " + val["text"] for val in corpus.values()]
queries_lst = list(queries.values())
if corpus_subsample is not None:
corpus_lst = corpus_lst[:corpus_subsample]
if queries_subsample is not None:
queries_lst = queries_lst[:queries_subsample]
# Tokenize using sklearn-style tokenizer + PyStemmer
stemmer = Stemmer.Stemmer("english")
corpus_token_strs = bm25s.tokenize(
corpus_lst, stopwords="en", stemmer=stemmer, return_ids=False
)
queries_token_strs = bm25s.tokenize(
queries_lst, stopwords="en", stemmer=stemmer, return_ids=False
)
print()
print(f"Dataset: {dataset}\n")
# print corpus and queries size
print(f"Corpus size: {len(corpus_lst)}")
print(f"Queries size: {len(queries_lst)}")
print()
# Initialize and index bm25s with atire + robertson idf (to match rank-bm25)
if method == "rank":
bm25_sparse = bm25s.BM25(k1=1.5, b=0.75, method="atire", idf_method="robertson")
elif method in ["bm25+", "bm25l"]:
bm25_sparse = bm25s.BM25(k1=1.5, b=0.75, delta=0.5, method=method)
else:
raise ValueError("invalid method")
start_time = time.monotonic()
bm25_sparse.index(corpus_token_strs)
bm25_sparse_index_time = time.monotonic() - start_time
print(f"bm25s index time: {bm25_sparse_index_time:.4f}s")
# Scoring with bm25-sparse
start_time = time.monotonic()
bm25_sparse_scores = [bm25_sparse.get_scores(q) for q in queries_token_strs]
bm25_sparse_score_time = time.monotonic() - start_time
print(f"bm25s score time: {bm25_sparse_score_time:.4f}s")
# Initialize and index rank-bm25
start_time = time.monotonic()
if method == "rank":
bm25_rank = rank_bm25.BM25Okapi(corpus_token_strs, k1=1.5, b=0.75, epsilon=0.0)
elif method == "bm25+":
bm25_rank = rank_bm25.BM25Plus(corpus_token_strs, k1=1.5, b=0.75, delta=0.5)
elif method == "bm25l":
bm25_rank = rank_bm25.BM25L(corpus_token_strs, k1=1.5, b=0.75, delta=0.5)
else:
raise ValueError("invalid method")
bm25_rank_index_time = time.monotonic() - start_time
print(f"rank-bm25 index time: {bm25_rank_index_time:.4f}s")
# Scoring with rank-bm25
start_time = time.monotonic()
bm25_rank_scores = [bm25_rank.get_scores(q) for q in queries_token_strs]
bm25_rank_score_time = time.monotonic() - start_time
print(f"rank-bm25 score time: {bm25_rank_score_time:.4f}s")
# print difference in time
print(
f"Index Time: BM25S is {bm25_rank_index_time / bm25_sparse_index_time:.2f}x faster than rank-bm25."
)
print(
f"Score Time: BM25S is {bm25_rank_score_time / bm25_sparse_score_time:.2f}x faster than rank-bm25."
)
# Check if scores are exactly the same
sparse_scores = np.array(bm25_sparse_scores)
rank_scores = np.array(bm25_rank_scores)
error_msg = f"\nScores between bm25-sparse and rank-bm25 are not exactly the same on dataset {dataset}."
almost_equal = np.allclose(sparse_scores, rank_scores)
self.assertTrue(almost_equal, error_msg)
general_info = {
"date": time.strftime("%Y-%m-%d %H:%M:%S"),
"num_jobs": 1,
"dataset": dataset,
"corpus_size": len(corpus_lst),
"queries_size": len(queries_lst),
"corpus_subsampled": corpus_subsample is not None,
"queries_subsampled": queries_subsample is not None,
}
# Save metrics
res = {
"model": "bm25s",
"index_time": bm25_sparse_index_time,
"score_time": bm25_sparse_score_time,
}
res.update(general_info)
save_scores(res, artifact_dir=artifact_dir)
res = {
"model": "rank-bm25",
"score_time": bm25_rank_score_time,
"index_time": bm25_rank_index_time,
}
res.update(general_info)
save_scores(res, artifact_dir=artifact_dir)
def compare_with_bm25_pt(
self,
dataset,
artifact_dir="tests/artifacts",
rel_save_dir="datasets",
corpus_subsample=None,
queries_subsample=None,
):
from beir.datasets.data_loader import GenericDataLoader
from beir.util import download_and_unzip
import bm25_pt
import bm25s.hf
from transformers import AutoTokenizer
warnings.filterwarnings("ignore", category=ResourceWarning)
warnings.filterwarnings("ignore", category=UserWarning)
# Download and prepare dataset
base_url = (
"https://public.ukp.informatik.tu-darmstadt.de/thakur/BEIR/datasets/{}.zip"
)
url = base_url.format(dataset)
out_dir = Path(__file__).parent / rel_save_dir
data_path = download_and_unzip(url, str(out_dir))
corpus, queries, qrels = GenericDataLoader(data_folder=data_path).load(
split="test"
)
# Convert corpus and queries to lists
corpus_lst = [val["title"] + " " + val["text"] for val in corpus.values()]
queries_lst = list(queries.values())
if corpus_subsample is not None:
corpus_lst = corpus_lst[:corpus_subsample]
if queries_subsample is not None:
queries_lst = queries_lst[:queries_subsample]
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
t0 = time.monotonic()
tokenized_corpus = bm25s.hf.batch_tokenize(tokenizer, corpus_lst)
time_corpus_tok = time.monotonic() - t0
t0 = time.monotonic()
queries_tokenized = bm25s.hf.batch_tokenize(tokenizer, queries_lst)
time_query_tok = time.monotonic() - t0
print()
print(f"Dataset: {dataset}\n")
# print corpus and queries size
print(f"Corpus size: {len(corpus_lst)}")
print(f"Queries size: {len(queries_lst)}")
print()
# Initialize and index bm25-sparse
bm25_sparse = bm25s.BM25(k1=1.5, b=0.75, method="atire", idf_method="lucene")
start_time = time.monotonic()
bm25_sparse.index(tokenized_corpus)
bm25s_index_time = time.monotonic() - start_time
print(f"bm25s index time: {bm25s_index_time:.4f}s")
# Scoring with bm25-sparse
start_time = time.monotonic()
bm25_sparse_scores = [bm25_sparse.get_scores(q) for q in queries_tokenized]
bm25s_score_time = time.monotonic() - start_time
print(f"bm25s score time: {bm25s_score_time:.4f}s")
# Initialize and index rank-bm25
start_time = time.monotonic()
model_pt = bm25_pt.BM25(tokenizer=tokenizer, device="cpu", k1=1.5, b=0.75)
model_pt.index(corpus_lst)
bm25_pt_index_time = time.monotonic() - start_time
bm25_pt_index_time -= time_corpus_tok
print(f"bm25-pt index time: {bm25_pt_index_time:.4f}s")
# Scoring with rank-bm25
start_time = time.monotonic()
bm25_pt_scores = model_pt.score_batch(queries_lst)
bm25_pt_scores = bm25_pt_scores.cpu().numpy()
bm25_pt_score_time = time.monotonic() - start_time
bm25_pt_score_time -= time_query_tok
print(f"bm25-pt score time: {bm25_pt_score_time:.4f}s")
# print difference in time
print(
f"Index Time: BM25S is {bm25_pt_index_time / bm25s_index_time:.2f}x faster than bm25-pt."
)
print(
f"Score Time: BM25S is {bm25_pt_score_time / bm25s_score_time:.2f}x faster than bm25-pt."
)
# Check if scores are exactly the same
bm25_sparse_scores = np.array(bm25_sparse_scores)
bm25_pt_scores = np.array(bm25_pt_scores)
error_msg = f"\nScores between bm25-sparse and rank-bm25 are not exactly the same on dataset {dataset}."
almost_equal = np.allclose(bm25_sparse_scores, bm25_pt_scores, atol=1e-4)
self.assertTrue(almost_equal, error_msg)
general_info = {
"date": time.strftime("%Y-%m-%d %H:%M:%S"),
"num_jobs": 1,
"dataset": dataset,
"corpus_size": len(corpus_lst),
"queries_size": len(queries_lst),
"corpus_was_subsampled": corpus_subsample is not None,
"queries_was_subsampled": queries_subsample is not None,
}
# Save metrics
res = {
"model": "bm25s",
"index_time": bm25s_index_time,
"score_time": bm25s_score_time,
}
res.update(general_info)
save_scores(res, artifact_dir=artifact_dir)
res = {
"model": "bm25-pt",
"score_time": bm25_pt_score_time,
"index_time": bm25_pt_index_time,
}
res.update(general_info)
save_scores(res, artifact_dir=artifact_dir)
|