sts.py

import time
import pandas as pd
import polars as pl
import torch
from datasets import Dataset
from sentence_transformers import SentenceTransformer


def sts(modelname, data1, data2, score):
    st = time.time()

    device = "cuda" if torch.cuda.is_available() else "cpu"
    model = SentenceTransformer(
        modelname,
        device=device,
        trust_remote_code=True,
    )

    sentences1 = Dataset.from_pandas(pd.read_csv(data1, on_bad_lines='skip', header=0, sep="\t"))
    sentences2 = Dataset.from_pandas(pd.read_csv(data2, on_bad_lines='skip', header=0, sep="\t"))

    embeddings1 = model.encode(sentences1["text"], normalize_embeddings=True, batch_size=1024,
                               show_progress_bar=True)
    embeddings2 = model.encode(sentences2["text"], normalize_embeddings=True, batch_size=1024,
                               show_progress_bar=True)

    similarity_matrix = model.similarity(embeddings1, embeddings2)

    df_pd = pd.DataFrame(similarity_matrix)
    dfi = df_pd.__dataframe__()
    df = pl.from_dataframe(dfi)

    df_matrix_with_index = df.with_row_index(name="row_index").with_columns(pl.col("row_index").cast(pl.UInt64))
    df_long = df_matrix_with_index.unpivot(index="row_index", variable_name="column_index",
                                           value_name="score").with_columns(pl.col("column_index").cast(pl.UInt64))
    df_sentences1 = pl.DataFrame(sentences1.to_pandas()).with_row_index(name="row_index").with_columns(
        pl.col("row_index").cast(pl.UInt64))
    df_sentences2 = pl.DataFrame(sentences2.to_pandas()).with_row_index(name="column_index").with_columns(
        pl.col("column_index").cast(pl.UInt64))

    df_long = (df_long
               .with_columns([pl.col("score").round(4).cast(pl.Float32)])  # Ensure column_index is UInt32
               .join(df_sentences1, on="row_index")
               .join(df_sentences2, on="column_index"))

    df_long = df_long.rename({
        "text": "sentences1",
        "text_right": "sentences2",
    }).drop(["row_index", "column_index"])

    elapsed_time = time.time() - st
    print('Execution time:', time.strftime("%H:%M:%S", time.gmtime(elapsed_time)))

    return df_long.filter(pl.col("score") > score).sort(["score"],
                                                        descending=True)