app.py

import os
import torch
import pandas as pd
import gradio as gr
from collections import defaultdict
from transformers import AutoTokenizer, AutoModelForTokenClassification

# =========================================================================
# 1. Sabitler ve Model Yükleme
# =========================================================================

# Hugging Face Hub'daki modelinizin ID'si
HF_MODEL_ID = "LiProject/Bert-Turkish-POS-Trained-V2" 

# GPU/CPU kontrolü
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

try:
    # Model ve Tokenizer'ı HF Hub'dan yükle
    tok = AutoTokenizer.from_pretrained(HF_MODEL_ID, use_fast=True)
    mdl = AutoModelForTokenClassification.from_pretrained(HF_MODEL_ID).to(DEVICE).eval()
    print(f"Model yükleme başarılı: {HF_MODEL_ID} ({DEVICE} üzerinde)")

except Exception as e:
    print(f"Model veya Tokenizer yüklenirken kritik hata oluştu: {e}")
    exit(1)

# =========================================================================
# 2. Etiket Okuma Fonksiyonları
# =========================================================================

def build_id2label_from_config(cfg):
    # Modelin config dosyasından id2label'ı güvenilir bir şekilde okur
    n = getattr(cfg, "num_labels", None)
    if n is None:
        if isinstance(getattr(cfg, "id2label", None), dict): n = len(cfg.id2label)
        elif isinstance(getattr(cfg, "label2id", None), dict): n = len(cfg.label2id)
        else: raise ValueError("num_labels/id2label/label2id yok.")
    
    labels = [f"LABEL_{i}" for i in range(n)]
    id2label = getattr(cfg, "id2label", None)
    
    if id2label:
        if isinstance(id2label, dict):
            for k,v in id2label.items():
                try: i = int(k)
                except:
                    try: i = int(float(k))
                    except: continue
                if 0 <= i < n: labels[i] = str(v)
        elif isinstance(id2label, (list,tuple)) and len(id2label)==n:
            labels = [str(x) for x in id2label]
            
    l2i = getattr(cfg, "label2id", None)
    if isinstance(l2i, dict):
        for lbl, idx_ in l2i.items():
            try: i = int(idx_)
            except:
                try: i = int(float(idx_))
                except: continue
            if 0 <= i < n and labels[i].startswith("LABEL_"):
                labels[i] = str(lbl)
                
    for i,v in enumerate(labels):
        if v.startswith("LABEL_"): labels[i] = str(i)
        
    return labels

ID2LABEL = build_id2label_from_config(mdl.config)

# =========================================================================
# 3. Inference ve Çıktı Formatı
# =========================================================================

@torch.inference_mode()
def tag_rows(multiline_text: str):
    """Metni işler ve kelime bazlı etiketlenmiş DataFrame döndürür."""
    rows = []
    sentences = [s.strip() for s in multiline_text.splitlines() if s.strip()]
    
    if not sentences:
        return pd.DataFrame(rows)

    for sent in sentences:
        enc = tok(sent, return_tensors="pt", truncation=True, add_special_tokens=True).to(DEVICE)
        logits = mdl(**enc).logits[0] 

        fast = tok(sent, return_offsets_mapping=True, add_special_tokens=True)
        word_ids = fast.word_ids()
        offsets  = fast["offset_mapping"]

        idxs_by_word = defaultdict(list)
        for i, wid in enumerate(word_ids):
            if wid is not None:
                idxs_by_word[wid].append(i)

        for wid in sorted(idxs_by_word.keys()):
            sub_idxs = idxs_by_word[wid]

            start = offsets[sub_idxs[0]][0]
            end   = offsets[sub_idxs[-1]][1]
            surface = sent[start:end] if (start is not None and end is not None) else ""

            mean_logits = logits[sub_idxs].mean(dim=0)
            pid = int(mean_logits.argmax().item())
            
            # Modelin doğrudan tahmin ettiği orijinal etiketi al
            tag = ID2LABEL[pid] if pid < len(ID2LABEL) else str(pid)

            rows.append({"Full_Sentence": sent, "Word": surface, "Tag": tag})

    return pd.DataFrame(rows)

def add_sentence_separators(df: pd.DataFrame, char: str = "-", repeat: int = 10) -> pd.DataFrame:
    """Görünürlük için cümleler arasına ayraç satırları ekler."""
    rows, prev = [], None
    for _, r in df.iterrows():
        if prev is not None and r["Full_Sentence"] != prev:
            sep = char * repeat
            rows.append({"Full_Sentence": sep, "Word": sep, "Tag": sep})
        rows.append(r.to_dict())
        prev = r["Full_Sentence"]
    return pd.DataFrame(rows)

def run_and_save(text):
    """Ana çalıştırma fonksiyonu, tabloyu ve indirilebilir CSV'yi hazırlar."""
    df = tag_rows(text)
    df_view = add_sentence_separators(df, char="-", repeat=10)

    out_path = "pos_output.csv" 
    df.to_csv(out_path, index=False) 

    return df_view, out_path

examples = [
    "hızlıca koştu",
    "O her zaman güler.",
    "At hızlıca koştu."
]

# =========================================================================
# 4. Gradio Arayüzü
# =========================================================================

# Tema ve CSS (Sizin orijinal stiliniz korundu)
theme = gr.themes.Soft(primary_hue="slate", neutral_hue="slate")
custom_css = """
/* Sayfa ve temel renkler */
.gradio-container { background: #000000 !important; color: #FFE8DB !important; font-family: Inter, ui-sans-serif, system-ui, -apple-system, Segoe UI, Roboto, "Helvetica Neue", Arial, sans-serif; }
.prose h1, .prose h2, .prose h3, .prose p, label { color: #FFE8DB !important; }
.gr-box, .gr-panel, .border, .container { background: #0b0b0b !important; border: 1.5px solid #739EC9 !important; border-radius: 14px !important; }
textarea, input, .gr-textbox, .gr-file, .gr-form input, .gr-form textarea { background: #0f1a26 !important; color: #FFE8DB !important; border: 2px solid #5682B1 !important; border-radius: 12px !important; }
button { transition: background 0.15s ease, filter 0.15s ease, box-shadow 0.15s ease; }
button.primary, .btn-primary { background: #FFE8DB !important; color: #000000 !important; }
button.primary:hover, .btn-primary:hover { filter: brightness(0.92); }
button.secondary, .btn-secondary { background: rgba(86,130,177,0.15) !important; color: #FFE8DB !important; }
button.secondary:hover, .btn-secondary:hover { background: rgba(86,130,177,0.38) !important; border-color: #5682B1 !important; }
table { border-collapse: separate !important; border-spacing: 0 !important; }
th { background: #5682B1 !important; color: #FFE8DB !important; }
td { background: #0f1a26 !important; color: #FFE8DB !important; }
tbody tr:nth-child(2n) td { background: #122434 !important; }
#results_table { max-height: 360px !important; overflow: auto !important; }
#results_table table { table-layout: fixed !important; width: 100% !important; }
#results_table th, #results_table td { white-space: normal !important; word-break: break-word !important; }
#input_text textarea { min-height: 150px !important; }
"""

with gr.Blocks(title="TR POS Tagger", theme=theme, css=custom_css, fill_height=True) as demo:
    gr.Markdown("# 🇹🇷 Türkçe POS Tagger")
    gr.Markdown(f"Model: `{HF_MODEL_ID.split('/')[-1]}`. Metni satır satır gir (her satır = 1 cümle). Çıktı: **Full_Sentence, Word, Tag**.")

    with gr.Row():
        with gr.Column(scale=3):
            inp = gr.Textbox(
                lines=6,
                placeholder="Örn:\nAt hızlıca koştu.\nO her zaman güler.",
                show_label=False,
                elem_id="input_text"
            )
        with gr.Column(scale=1):
            btn = gr.Button("Etiketle ve CSV indir", variant="primary", elem_id="run_btn")
            clr = gr.Button("Temizle", variant="secondary", elem_id="clear_btn")

    out_tbl = gr.Dataframe(
        headers=["Full_Sentence","Word","Tag"],
        label="Önizleme",
        interactive=False,
        elem_id="results_table"
    )

    out_file = gr.File(label="Çıktı CSV")

    gr.Examples(examples=[[e] for e in examples], inputs=inp)

    btn.click(run_and_save, inputs=inp, outputs=[out_tbl, out_file])
    inp.submit(run_and_save, inputs=inp, outputs=[out_tbl, out_file])
    clr.click(lambda: ("", None, None), outputs=[inp, out_tbl, out_file])

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