app.py

import gradio as gr
from gradio import update
import stanza
import pandas as pd
import requests
import traceback
from pathlib import Path

# 1. MODEL VARIANTS & INITIALIZATION

LESBIAN_MODELS = {}
MODEL_VARIANTS = {
    "Lesbian-only (UD_Greek-Lesbian)":           "sbompolas/Lesbian-Greek-Morphosyntactic-Model",
    "Lesbian-augmented (UD_Greek-Lesbian+NGUD)": "sbompolas/NGUD-Lesbian-Morphosyntactic-Model",
    "Standard Modern Greek (UD_Greek-GUD)":      "sbompolas/GUD",
    "Cretan-only (UD_Greek-Cretan)":      "sbompolas/Cretan"
}

def download_model_file(url, filename):
    try:
        resp = requests.get(url, stream=True)
        resp.raise_for_status()
        with open(filename, "wb") as f:
            for chunk in resp.iter_content(8192):
                f.write(chunk)
        return True
    except Exception as e:
        print(f"Download failed {filename}: {e}")
        return False

def initialize_models():
    try:
        base = Path("./models")
        base.mkdir(exist_ok=True)
        for name, repo in MODEL_VARIANTS.items():
            out = base / name
            out.mkdir(exist_ok=True)
            files = {
                "tokenizer.pt":  f"https://huggingface.co/{repo}/resolve/main/tokenizer.pt",
                "lemmatizer.pt": f"https://huggingface.co/{repo}/resolve/main/lemmatizer.pt",
                "pos.pt":        f"https://huggingface.co/{repo}/resolve/main/pos.pt",
                "depparse.pt":   f"https://huggingface.co/{repo}/resolve/main/depparse.pt",
            }
            for fn, url in files.items():
                tgt = out / fn
                if not tgt.exists() and not download_model_file(url, str(tgt)):
                    return False, f"Failed to download {fn} for {name}"
            cfg = {
                'processors': 'tokenize,pos,lemma,depparse',
                'lang': 'el',
                'use_gpu': False,
                'verbose': False,
                'tokenize_model_path': str(out/"tokenizer.pt"),
                'pos_model_path':      str(out/"pos.pt"),
                'lemma_model_path':    str(out/"lemmatizer.pt"),
                'depparse_model_path': str(out/"depparse.pt")
            }
            LESBIAN_MODELS[name] = stanza.Pipeline(**cfg)
        return True, "Models loaded"
    except Exception as e:
        traceback.print_exc()
        return False, str(e)

loaded, load_status = initialize_models()


# 2. CONLL-U / DATAFRAME

def stanza_doc_to_conllu(doc) -> str:
    lines = []
    for sid, sent in enumerate(doc.sentences, 1):
        lines.append(f"# sent_id = {sid}")
        lines.append(f"# text = {sent.text}")
        for w in sent.words:
            fields = [
                str(w.id), w.text,
                w.lemma or "_", w.upos or "_",
                w.xpos or "_", w.feats or "_",
                str(w.head) if w.head is not None else "0",
                w.deprel or "_", "_", "_"
            ]
            lines.append("\t".join(fields))
        lines.append("")  # blank line after each sentence
    return "\n".join(lines)

def conllu_to_dataframe(conllu: str) -> pd.DataFrame:
    cols = ['ID','FORM','LEMMA','UPOS','XPOS','FEATS','HEAD','DEPREL','DEPS','MISC']
    rows = []
    for line in conllu.splitlines():
        if not line:
            # empty separator row
            rows.append({c: "" for c in cols})
            continue
        if line.startswith("#"):
            key, val = line[2:].split("=", 1)
            key, val = key.strip(), val.strip()
            if key == "sent_id":
                rows.append({'ID': f"# sent_id = {val}", 'FORM': ""})
            elif key == "text":
                rows.append({'ID': f"# text = {val}", 'FORM': ""})
            continue
        parts = line.split("\t")
        if len(parts) >= 10:
            rows.append(dict(zip(cols, parts)))
    return pd.DataFrame(rows, columns=cols).fillna("")


# 3. FULL SVG BUILDER (crop top + bottom padding, arrows at start)

def create_single_sentence_svg(sentence_data, sentence_num=1, total_sentences=1):
    try:
        df = pd.DataFrame(sentence_data) if isinstance(sentence_data, list) else sentence_data
        n = len(df)
        base_w, min_sp = 100, 30
        spacing = max(base_w, (n*base_w + (n-1)*min_sp)/n)
        width = max(800, n*spacing + 100)
        orig_height = 500
        crop_top   = 30  # px to remove from top
        bottom_pad = 30  # px to add at bottom
        height     = orig_height - crop_top + bottom_pad

        word_y   = height - 120
        feats_y  = word_y + 35

        colors = {
            'root':'#000000','nsubj':'#2980b9','obj':'#27ae60','det':'#e67e22',
            'amod':'#8e44ad','nmod':'#16a085','case':'#34495e','punct':'#7f8c8d',
            'cc':'#d35400','conj':'#2c3e50','cop':'#e74c3c','mark':'#9b59b6',
            'csubj':'#3498db','xcomp':'#1abc9c','ccomp':'#f39c12','advcl':'#e91e63',
            'advmod':'#9c27b0','obl':'#795548','iobj':'#607d8b','fixed':'#ff5722',
            'aux':'#ff9800','acl':'#4caf50','appos':'#673ab7','compound':'#009688'
        }

        svg = [
            f'<svg width="{width}" height="{height}" viewBox="0 {crop_top} {width} {orig_height}" '
            'xmlns="http://www.w3.org/2000/svg" style="background:white;border:1px solid #eee;"><defs>'
        ]
        for rel, c in colors.items():
            svg.append(
                f'<marker id="m_{rel}" markerWidth="4" markerHeight="4" '
                'markerUnits="userSpaceOnUse" orient="auto-start-reverse" refX="3.5" refY="2">'
                f'<path d="M0,0 L4,2 L0,4Z" fill="{c}"/></marker>'
            )
        svg.append('</defs><g>')

        # compute x positions
        xpos = {
            int(r['ID']): 50 + (int(r['ID']) - 1) * spacing
            for _, r in df.iterrows() if str(r['ID']).isdigit()
        }

        used_spans = []
        for _, r in df.iterrows():
            if not str(r['ID']).isdigit(): 
                continue
            i, h = int(r['ID']), int(r['HEAD'])
            rel, c = r['DEPREL'], colors.get(r['DEPREL'], '#000')
            x1 = xpos[i]
            if h == 0:
                # ROOT line
                svg.append(
                    f'<line x1="{x1}" y1="{word_y-15}" x2="{x1}" y2="50" '
                    f'stroke="{c}" stroke-width="1.5"/>'
                )
                mid = (word_y-15 + 50) / 2
                svg.append(
                    f'<rect x="{x1-15}" y="{mid-8}" width="30" height="14" '
                    f'fill="white" stroke="{c}" rx="2"/>'
                )
                svg.append(
                    f'<text x="{x1}" y="{mid+2}" text-anchor="middle" '
                    f'fill="{c}" font-family="Arial" font-size="8" font-weight="bold">ROOT</text>'
                )
            else:
                x2 = xpos.get(h, x1)
                span = (min(i, h), max(i, h))
                lvl = 0
                conflict = True
                while conflict:
                    conflict = False
                    for (es, el), used_lvl in used_spans:
                        if used_lvl == lvl and not (span[1] < es or span[0] > el):
                            lvl += 1
                            conflict = True
                            break
                used_spans.append((span, lvl))
                dist = abs(x2 - x1)
                arc_h = min(40 + dist * 0.15, 100) + lvl * 35
                midx, cty = (x1 + x2) / 2, word_y - arc_h
                path_d = f'M{x1} {word_y-15} Q{midx} {cty} {x2} {word_y-15}'
                svg.append(
                    f'<path d="{path_d}" stroke="{c}" fill="none" stroke-width="1.5" '
                    f'marker-start="url(#m_{rel})"/>'
                )
                amx = 0.25*x1 + 0.5*midx + 0.25*x2
                amy = 0.25*(word_y-15) + 0.5*cty + 0.25*(word_y-15)
                lw = len(rel)*6 + 8
                svg.append(
                    f'<rect x="{amx-lw/2}" y="{amy-8}" width="{lw}" height="14" '
                    f'fill="white" stroke="{c}" rx="2"/>'
                )
                svg.append(
                    f'<text x="{amx}" y="{amy+2}" text-anchor="middle" '
                    f'fill="{c}" font-family="Arial" font-size="8" font-weight="bold">{rel}</text>'
                )

        # draw tokens + annotations
        for _, r in df.iterrows():
            if not str(r['ID']).isdigit():
                continue
            x = xpos[int(r['ID'])]
            svg.append(
                f'<text x="{x}" y="{word_y}" text-anchor="middle" '
                f'font-family="Arial" font-size="13" font-weight="bold">{r["FORM"]}</text>'
            )
            ann = []
            if r['UPOS'] and r['UPOS'] != '_': ann.append(f"upos={r['UPOS']}")
            if r['LEMMA'] not in ('_', r['FORM']): ann.append(f"lemma={r['LEMMA']}")
            if r['FEATS'] and r['FEATS'] not in ('', '_'):
                for f in r['FEATS'].split('|'):
                    if '=' in f:
                        ann.append(f)
            for i, a in enumerate(ann):
                svg.append(
                    f'<text x="{x}" y="{feats_y + i*12}" text-anchor="middle" '
                    f'font-family="Arial" font-size="7" fill="#666">{a}</text>'
                )

        svg.append('</g></svg>')
        return "".join(svg)
    except Exception as e:
        return f"<p>Error creating SVG: {e}</p>"


# 4. PROCESS & DROPDOWN-UPDATES

def process_text(text, variant):
    if not text.strip():
        return (
            gr.HTML.update(value="<p>No data</p>"),
            gr.Dropdown.update(choices=[], value=None),
            [], "", pd.DataFrame()
        )
    pipe = LESBIAN_MODELS.get(variant)
    if pipe is None:
        return (
            gr.HTML.update(value="<p>Error: model not loaded</p>"),
            gr.Dropdown.update(choices=[], value=None),
            [], "", pd.DataFrame()
        )
    doc = pipe(text)
    conllu = stanza_doc_to_conllu(doc)
    df = conllu_to_dataframe(conllu)

    sentences = []
    for sent in doc.sentences:
        payload = [{
            'ID':     w.id,
            'FORM':   w.text,
            'LEMMA':  w.lemma or "_",
            'UPOS':   w.upos or "_",
            'XPOS':   w.xpos or "_",
            'FEATS':  w.feats or "_",
            'HEAD':   w.head or 0,
            'DEPREL': w.deprel or "_"
        } for w in sent.words]
        sentences.append(payload)

    sent_ids = [str(i+1) for i in range(len(sentences))]
    dd_upd = update(choices=sent_ids, value=sent_ids[0] if sent_ids else None)
    init_svg = create_single_sentence_svg(sentences[0]) if sentences else "<p>No data</p>"

    return init_svg, dd_upd, sentences, conllu, df

def update_svg(selected_id, sentences):
    try:
        idx = int(selected_id) - 1
        return create_single_sentence_svg(sentences[idx])
    except:
        return "<p>Invalid selection</p>"


# 5. BUILD GRADIO UI

def create_app():
    with gr.Blocks(title="Parser for MG Dialects") as app:
        gr.Markdown("# Morphosyntactic Parser for MG Dialects")

        if not loaded:
            gr.Markdown(f"❌ Load error: {load_status}")

        with gr.Row():
            with gr.Column():
                txt = gr.Textbox(label="Input Text", lines=4, placeholder="Εισάγετε κείμενο…")
                mdl = gr.Radio(choices=list(MODEL_VARIANTS.keys()), value="Lesbian-only", label="Model Variant")
                btn = gr.Button("Parse", variant="primary")

        with gr.Row():
            with gr.Column():
                svg_out      = gr.HTML("<p>No visualization</p>")
                sentence_dd  = gr.Dropdown(label="Choose sentence", choices=[])
                sentences_st = gr.State([])

        with gr.Row():
            with gr.Column():
                conllu_out = gr.Textbox(label="CoNLL-U", lines=10, show_copy_button=True)
                table_out  = gr.Dataframe(label="Token Table")

        btn.click(
            fn=process_text,
            inputs=[txt, mdl],
            outputs=[svg_out, sentence_dd, sentences_st, conllu_out, table_out]
        )
        sentence_dd.change(fn=update_svg, inputs=[sentence_dd, sentences_st], outputs=svg_out)

    return app

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