predict_utils.py

import torch
import numpy as np
from transformers import AutoTokenizer
from model import CyberRoBERTa

LABELS = ['HS', 'Abusive', 'HS_Individual', 'HS_Group', 'HS_Religion', 'HS_Race',
          'HS_Physical', 'HS_Gender', 'HS_Other', 'HS_Weak', 'HS_Moderate', 'HS_Strong']
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

def load_model_and_thresholds():
    model = CyberRoBERTa()
    model.load_state_dict(torch.load("indoroberta_multilabel_model84%.bin", map_location=DEVICE))
    model.to(DEVICE)
    model.eval()

    tokenizer = AutoTokenizer.from_pretrained("cahya/roberta-base-indonesian-522M")
    thresholds = np.load("optimal_thresholds.npy")
    return model, tokenizer, thresholds

def predict(text, model, tokenizer, thresholds):
    encoding = tokenizer(text, return_tensors='pt', padding='max_length', truncation=True, max_length=128)
    input_ids = encoding['input_ids'].to(DEVICE)
    attention_mask = encoding['attention_mask'].to(DEVICE)

    with torch.no_grad():
        outputs = model(input_ids, attention_mask)
    probs = outputs[0].cpu().numpy()
    preds_bin = (probs > thresholds).astype(int)

    return {label: float(prob) for label, prob in zip(LABELS, probs)}