from __future__ import annotations from dataclasses import dataclass from typing import Generator import numpy as np from sentence_transformers import SentenceTransformer from tqdm import tqdm import torch import argparse from pathlib import Path from datasets import load_dataset parser = argparse.ArgumentParser(description="Convert datasets to embeddings") parser.add_argument( "-t", "--target", type=str, required=True, choices=["data", "chunked"], help="target dataset, data or chunked", ) parser.add_argument( "-d", "--debug", action="store_true", help="debug mode, use small dataset", ) # model_name parser.add_argument( "-m", "--model_name", type=str, required=True, help="huggingface model name", ) # input_prefix parser.add_argument( "-i", "--input_prefix", type=str, required=False, default="", help="input prefix", ) # max_seq_length parser.add_argument( "-l", "--max_seq_length", type=int, required=False, default=512, help="max sequence length", ) # output_name parser.add_argument( "-o", "--output_name", type=str, required=True, help="output dir", ) args = parser.parse_args() @dataclass class EmbConfig: model_name: str input_prefix: str max_seq_length: int args = parser.parse_args() target_local_ds = args.target EMB_CONFIG = EmbConfig( model_name=args.model_name, input_prefix=args.input_prefix, max_seq_length=args.max_seq_length, ) embs_dir = f"embs{'_debug' if args.debug else ''}" output_embs_path = Path("/".join([embs_dir, args.output_name, target_local_ds])) output_embs_path.mkdir(parents=True, exist_ok=True) print("output path:", output_embs_path) MODEL = SentenceTransformer(EMB_CONFIG.model_name) MODEL.max_seq_length = EMB_CONFIG.max_seq_length def to_embs(texts: list[str], group_size=1024) -> Generator[np.ndarray, None, None]: group = [] for text in texts: group.append(text) if len(group) == group_size: embeddings = MODEL.encode( group, normalize_embeddings=True, show_progress_bar=False, ) yield embeddings # type: ignore group = [] if len(group) > 0: embeddings = MODEL.encode( group, normalize_embeddings=True, show_progress_bar=False ) yield embeddings # type: ignore def _to_data_text( data, prefix=EMB_CONFIG.input_prefix, max_len=int(EMB_CONFIG.max_seq_length * 1.5) ): return (prefix + data["title"] + "\n" + data["text"])[0:max_len] def _to_chunk_text( data, prefix=EMB_CONFIG.input_prefix, max_len=int(EMB_CONFIG.max_seq_length * 1.5) ): return (prefix + data["title"] + "\n" + data["overlap_text"] + data["text"])[ :max_len ] def ds_to_embs( ds, text_fn, group_size: int, ): texts = [] total = len(ds) pbar = tqdm(total=total) # text は group_size 件ごとに処理する for i in range(0, total, group_size): texts = [] for data in ds.select(range(i, min(i + group_size, total))): data: dict = data text = text_fn(data) texts.append(text) embs = [] for group_embs in to_embs(texts): embs.append(group_embs) pbar.update(len(group_embs)) embs = np.concatenate(embs) yield embs, i, pbar if torch.cuda.is_available(): print("use cuda") MODEL.to("cuda") else: print("!! Warning: use cpu") ds = load_dataset(args.target)["train"] # type: ignore to_text = _to_data_text if args.target == "data" else _to_chunk_text if args.debug: print("debug mode") ds = ds.select(range(19998)) # type: ignore print("small dataset len:", len(ds)) group_size = 10000 else: print("dataset len:", len(ds)) group_size = 100_000 for embs, idx, pbar in ds_to_embs(ds, to_text, group_size=group_size): filename = f"{idx}.npz" filepath = output_embs_path / filename pbar.desc = f"saving...: {str(filepath)}" np.savez_compressed(filepath, embs=embs.astype(np.float16)) pbar.desc = ""