app.py

"""
La Fréquence du Vivant – Bio acoustic découverte
Comédie des Mondes Hybrides · Gaspard Boréal
------------------------------------------------
• Upload .wav/.mp3 (≤ 30 s)
• Prédit les tags AudioSet (AST 527) + spectrogramme
• Synthèse bio-acoustique : barres horizontales (≥ 0,5 %)
• Boutons : « Nouveau fichier » & « Voir tags (condensé / complet) »
"""

# ---------------------------- Imports ---------------------------------
import json, pathlib, matplotlib.pyplot as plt, pandas as pd
import torch, torchaudio, librosa, gradio as gr
from transformers import pipeline

# ---------------------------- 0. Ontology -----------------------------
ONTO = json.loads(pathlib.Path(__file__).with_name("ontology.json").read_text())
by_id   = {n["id"]: n for n in ONTO}
name2id = {n["name"]: n["id"] for n in ONTO}
parents = {}
for n in ONTO:
    for c in n.get("child_ids", []):
        parents.setdefault(c, []).append(n["id"])
def root_parent(mid):            # ID racine
    while parents.get(mid):
        mid = parents[mid][0]
    return mid

# ---------------------------- 1. Modèle -------------------------------
clf = pipeline("audio-classification",
               model="MIT/ast-finetuned-audioset-10-10-0.4593",
               top_k=None,
               device=0 if torch.cuda.is_available() else -1)

# ---------------------------- 2. Utils audio --------------------------
def load_audio(path, target_sr=16000, max_sec=30):
    ext = pathlib.Path(path).suffix.lower()
    if ext == ".wav":
        wav, sr = torchaudio.load(path)
    else:
        y, sr = librosa.load(path, sr=None, mono=False)
        wav = torch.tensor(y).unsqueeze(0) if y.ndim == 1 else torch.tensor(y)
    wav = wav[:, : max_sec * sr]
    if wav.shape[0] > 1:
        wav = wav.mean(0, keepdim=True)
    if sr != target_sr:
        wav = torchaudio.functional.resample(wav, sr, target_sr)
        sr = target_sr
    return wav, sr

def top_k_dict(d, k=5):
    return {k_: v_ for k_, v_ in
            sorted(d.items(), key=lambda x: x[1], reverse=True)[:k]}

# ---------------------------- 3. Analyse ------------------------------
THRESH = 0.005  # 0,5 %  (score >= 0.005)

def analyse(audio_path, expanded):
    wav, sr = load_audio(audio_path)
    res = clf(wav.numpy().squeeze(), sampling_rate=sr)

    full = {}
    for d in res:
        s = float(d["score"]) / 100 if d["score"] > 1 else float(d["score"])
        if s >= THRESH:
            full[d["label"]] = round(s, 4)

    # Synthèse racine (max) puis normalisation
    synth_raw = {}
    for label, sc in full.items():
        mid = name2id.get(label)
        if mid:
            root = by_id[root_parent(mid)]["name"]
            synth_raw[root] = max(synth_raw.get(root, 0), sc)
    tot = sum(synth_raw.values()) or 1
    synth_norm = {k: sc / tot for k, sc in synth_raw.items()}

    disp = full if expanded else top_k_dict(full, 5)

    # DataFrame filtré pour BarPlot (> 0,5 %)
    df = (pd.DataFrame({"Racine": synth_norm.keys(),
                        "Pourcent": [round(v * 100, 1) for v in synth_norm.values()]})
          .query("Pourcent >= 0.5")
          .sort_values("Pourcent", ascending=False))

    fig, ax = plt.subplots(figsize=(6, 2))
    ax.specgram(wav.numpy()[0], Fs=sr, NFFT=1024, noverlap=512)
    ax.set_axis_off(); plt.tight_layout()

    return disp, fig, full, df

# ---------------------------- 4. Interface ----------------------------
with gr.Blocks(title="La Fréquence du Vivant – Bio acoustic découverte") as demo:

    gr.Markdown("# La Fréquence du Vivant")
    gr.Markdown("### Écoute bio-acoustique : marche techno-sensible entre vivant, humain et machine")

    expanded   = gr.State(False)
    full_tags  = gr.State({})
    synth_df   = gr.State(pd.DataFrame())

    audio_in = gr.Audio(sources=["upload"], type="filepath",
                        label="🎙️  Charger un fichier .wav / .mp3 (≤ 30 s)")

    # Synthèse
    gr.Markdown("**Synthèse bio-acoustique (racines)**")
    synth_out = gr.BarPlot(x="Racine", y="Pourcent",
                           y_lim=(0, 100),
                           height=260,
                           title="Répartition par racine (%)")

    # Détails tags
    gr.Markdown("**Tags AudioSet**")
    with gr.Row():
        reset_btn  = gr.Button("🔄 Nouveau fichier / Réinitialiser", size="sm")
        toggle_btn = gr.Button("Voir tous les tags", size="sm", variant="primary")

    tags_out = gr.Label()
    spec_out = gr.Plot(label="Spectrogramme")

    # Helpers
    def _txt(exp): return "Uniquement les principaux tags" if exp else "Voir tous les tags"
    def flip(b): return not b
    def refresh(exp, full, df):
        disp = full if exp else top_k_dict(full, 5)
        return _txt(exp), disp, df
    def reset_ui():
        fig = plt.figure(figsize=(6, 2)); plt.axis("off")
        return None, {}, fig, pd.DataFrame(), {}, False, "Voir tous les tags"

    audio_in.upload(analyse,
                    [audio_in, expanded],
                    [tags_out, spec_out, full_tags, synth_df])

    toggle_btn.click(flip, expanded, expanded)\
              .then(refresh,
                    [expanded, full_tags, synth_df],
                    [toggle_btn, tags_out, synth_out])

    reset_btn.click(
        reset_ui, None,
        [audio_in, tags_out, spec_out,
         synth_out, full_tags, expanded, toggle_btn]
    )

    synth_df.change(
        lambda d: d.query("Pourcent >= 0.5").sort_values("Pourcent", ascending=False),
        synth_df, synth_out
    )

    gr.Markdown("<center>2025 : Gaspard Boréal / La comédie des mondes hybrides</center>")

# ---------------------------- 5. Run ----------------------------------
if __name__ == "__main__":
    demo.launch(debug=True)