import torch
from transformers import AutoModelForTokenClassification, AutoTokenizer

model_name = "cointegrated/rubert-tiny2-sentence-compression"
model = AutoModelForTokenClassification.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)


def compress(text, threshold=0.5, keep_ratio=None):
    """Compress a sentence by removing the least important words.
    Parameters:
        threshold: cutoff for predicted probabilities of word removal
        keep_ratio: proportion of words to preserve
    By default, threshold of 0.5 is used.
    """
    with torch.inference_mode():
        tok = tokenizer(text, return_tensors="pt").to(model.device)
        proba = torch.softmax(model(**tok).logits, -1).cpu().numpy()[0, :, 1]
    if keep_ratio is not None:
        threshold = sorted(proba)[int(len(proba) * keep_ratio)]
    kept_toks = []
    keep = False
    prev_word_id = None
    for word_id, score, token in zip(tok.word_ids(), proba, tok.input_ids[0]):
        if word_id is None:
            keep = True
        elif word_id != prev_word_id:
            keep = score < threshold
        if keep:
            kept_toks.append(token)
        prev_word_id = word_id
    return tokenizer.decode(kept_toks, skip_special_tokens=True)