app.py

# -*- coding: utf-8 -*-
# app.py — CAMeL‑Lab GED→GEC (بدون camel_tools / بدون ged_tags في generate)

import gradio as gr
import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, BertForTokenClassification, MBartForConditionalGeneration

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

# 1) أسماء النماذج (يمكن ترقية qalb14 إلى qalb15 إن رغبت)
GED_MODEL_NAME = "CAMeL-Lab/camelbert-msa-qalb14-ged-13"
GEC_MODEL_NAME = "CAMeL-Lab/arabart-qalb14-gec-ged-13"

# 2) تحميل النماذج والمُرمّزات
ged_tokenizer = AutoTokenizer.from_pretrained(GED_MODEL_NAME)
ged_model = BertForTokenClassification.from_pretrained(GED_MODEL_NAME).to(DEVICE).eval()

gec_tokenizer = AutoTokenizer.from_pretrained(GEC_MODEL_NAME)
gec_model = MBartForConditionalGeneration.from_pretrained(GEC_MODEL_NAME).to(DEVICE).eval()

# 3) الدالة الأساسية
def camel_gec_correct(text: str) -> str:
    if not text or not text.strip():
        return "⚠️ يرجى إدخال نص."

    # (أ) GED مباشرة على النص
    inputs = ged_tokenizer([text], return_tensors="pt")
    inputs = {k: v.to(DEVICE) for k, v in inputs.items()}
    with torch.no_grad():
        logits = ged_model(**inputs).logits
        probs = F.softmax(logits, dim=-1).squeeze()[1:-1]  # تجاهل [CLS]/[SEP]
        pred_ids = torch.argmax(probs, -1)
    pred_ged_labels = [ged_model.config.id2label[i.item()] for i in pred_ids]

    # (ب) محاذاة وسوم GED مع تقطيع GEC
    tokens, ged_labels = [], []
    for word, label in zip(text.split(), pred_ged_labels):
        w_toks = gec_tokenizer.tokenize(word)
        if w_toks:
            tokens.extend(w_toks)
            ged_labels.extend([label] * len(w_toks))

    # (ج) نبني تسلسل الإدخال لـ GEC (بدون ged_tags)
    input_ids = [gec_tokenizer.bos_token_id] + \
                gec_tokenizer.convert_tokens_to_ids(tokens) + \
                [gec_tokenizer.eos_token_id]

    # (د) توليد التصحيح القياسي
    with torch.no_grad():
        generated = gec_model.generate(
            torch.tensor([input_ids], device=DEVICE),
            num_beams=5,
            max_length=max(128, len(input_ids) + 32),
            num_return_sequences=1,
        )
    corrected = gec_tokenizer.batch_decode(
        generated, skip_special_tokens=True, clean_up_tokenization_spaces=False
    )[0]
    return corrected

# 4) واجهة Gradio
def ui_correct(text): 
    return camel_gec_correct(text)

with gr.Blocks(title="المصحح النحوي عربي (CAMeL‑Lab GED→GEC)") as demo:
    gr.Markdown("### تصحيح نحوي عربي باستخدام نماذج CAMeL‑Lab (GED ثم GEC)")
    inp = gr.Textbox(label="النص", lines=8, placeholder="أدخل النص العربي هنا...")
    btn = gr.Button("تصحيح")
    out = gr.Textbox(label="النص المصحح", lines=8)
    btn.click(fn=ui_correct, inputs=inp, outputs=out)

if __name__ == "__main__":
    demo.launch()