| """TF-IDF features of word 1-2-grams and character 3-5-grams (within word boundaries), stored without pickle.""" |
| from __future__ import annotations |
|
|
| import json |
| from pathlib import Path |
|
|
| import numpy as np |
|
|
|
|
| def vectorizers(): |
| from sklearn.feature_extraction.text import TfidfVectorizer |
| return (TfidfVectorizer(analyzer="word", ngram_range=(1, 2), min_df=1, sublinear_tf=True, lowercase=True), |
| TfidfVectorizer(analyzer="char_wb", ngram_range=(3, 5), min_df=2, sublinear_tf=True, lowercase=True)) |
|
|
|
|
| def save(vecs, path: Path) -> None: |
| path.mkdir(parents=True, exist_ok=True) |
| for i, v in enumerate(vecs): |
| (path / f"vocab{i}.json").write_text(json.dumps({k: int(x) for k, x in v.vocabulary_.items()}), encoding="utf-8") |
| np.save(path / f"idf{i}.npy", v.idf_.astype(np.float32)) |
|
|
|
|
| def load(path: Path): |
| vecs = vectorizers() |
| for i, v in enumerate(vecs): |
| v.vocabulary = json.loads((path / f"vocab{i}.json").read_text(encoding="utf-8")) |
| v.fit(["x"]) |
| v.idf_ = np.load(path / f"idf{i}.npy").astype(np.float64) |
| return vecs |
|
|
|
|
| def transform(vecs, texts): |
| from scipy.sparse import hstack |
| return hstack([v.transform(texts) for v in vecs]).tocsr() |
|
|
