app.py

import torch
import numpy as np
import pickle
from transformers import AutoTokenizer, AutoModel
from normalizer import normalize
import gradio as gr

# --- Device ---
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# --- Load tokenizers and models ---
bert_tokenizer = AutoTokenizer.from_pretrained("csebuetnlp/banglabert")
titu_tokenizer = AutoTokenizer.from_pretrained("hishab/titulm-llama-3.2-1b-v1.0")
bert_model = AutoModel.from_pretrained("csebuetnlp/banglabert").to(device).eval()
titu_model = AutoModel.from_pretrained("hishab/titulm-llama-3.2-1b-v1.0").to(device).eval()

# --- Load trained LightGBM model and preprocessing info ---
with open("multiclass_lightgbm_bert_titu.pkl", "rb") as f:
    classifier = pickle.load(f)

with open("preprocessing_info.pkl", "rb") as f:
    info = pickle.load(f)
class_names = info['class_names']
bert_max_len = 45
titu_max_len = 148

# --- Helper functions ---
def preprocess(text):
    return normalize(text)

def get_embedding(text, tokenizer, model, max_len):
    enc = tokenizer([text], return_tensors="pt", padding=True, truncation=True, max_length=max_len).to(device)
    with torch.inference_mode():
        out = model(**enc)
    last_hidden = out.last_hidden_state
    attn = enc.get("attention_mask", None)
    if attn is not None:
        attn = attn.unsqueeze(-1)
        emb = (last_hidden * attn).sum(dim=1) / attn.sum(dim=1).clamp(min=1e-6)
    else:
        emb = last_hidden.mean(dim=1)
    return emb.detach().cpu().numpy()

def predict(text):
    text = preprocess(text)
    bert_emb = get_embedding(text, bert_tokenizer, bert_model, bert_max_len)
    titu_emb = get_embedding(text, titu_tokenizer, titu_model, titu_max_len)
    features = np.concatenate([bert_emb, titu_emb], axis=1)
    pred_idx = classifier.predict(features)[0]
    return class_names[pred_idx]

# --- Gradio interface ---
iface = gr.Interface(
    fn=predict, 
    inputs=gr.Textbox(label="Enter Bangla text"),
    outputs=gr.Textbox(label="Predicted Trait"),
    title="Bangla Personality Trait Predictor",
    description="Enter Bangla text and get the predicted personality trait."
)

iface.launch()