Synthèse filtrée ≥0,5 % et barres horizontales

app.py

@@ -2,23 +2,35 @@
 La Fréquence du Vivant – Bio acoustic découverte
 Comédie des Mondes Hybrides · Gaspard Boréal
 ------------------------------------------------
-• Upload .wav/.mp3 (≤ 30 s) OU colle un lien http(s) (Dictaphone iCloud, Dropbox…)
-• Prédit les 3 tags AudioSet (AST 527) + spectrogramme
+• 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 os, pathlib, tempfile, requests, matplotlib.pyplot as plt
+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(
-    task="audio-classification",
-    model="MIT/ast-finetuned-audioset-10-10-0.4593",
-    top_k=3,
-    device=0 if torch.cuda.is_available() else -1,
-)
+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):
@@ -37,89 +49,106 @@ def load_audio(path, target_sr=16000, max_sec=30):
     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] if v_ > 0}
+    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 = {d["label"]: round(float(d["score"]), 4) for d in res}
+    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)
-    fig, ax = plt.subplots(figsize=(6, 2)); ax.specgram(wav.numpy()[0], Fs=sr, NFFT=1024, noverlap=512)
+
+    # 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
-
-def analyse_link(link, expanded):
-    if not link.startswith(("http://", "https://")):
-        raise gr.Error("Colle un lien commençant par http:// ou https://")
-    try:
-        r = requests.get(link, timeout=10); r.raise_for_status()
-    except Exception as e:
-        raise gr.Error(f"Téléchargement impossible : {e}")
-    with tempfile.NamedTemporaryFile(delete=False, suffix=".tmp") as tmp:
-        tmp.write(r.content); tmp_path = tmp.name
-    try:
-        return analyse(tmp_path, expanded)
-    finally:
-        os.remove(tmp_path)
+
+    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 – Bio acoustic découverte")
-    gr.Markdown("### Outils pour marche techno-sensible entre vivant, humain et machine")
+    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())
 
-    # Champ lien + bouton analysé uniquement si URL http(s)
-    with gr.Row():
-        url_in  = gr.Textbox(lines=1,
-            placeholder="�� Colle un lien http(s) (Dictaphone iCloud, Dropbox, etc.)",
-            label="Analyse via lien web")
-        url_btn = gr.Button("Analyser le lien", size="sm", variant="primary", interactive=False)
-
-    # Upload local
     audio_in = gr.Audio(sources=["upload"], type="filepath",
-                        label="🎙️  Ou charge un fichier .wav / .mp3 (≤ 30 s)")
+                        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 (%)")
 
-    # Tags + boutons
+    # 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 ----------
+    # 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): return _txt(exp), (full if exp else top_k_dict(full, 5))
+    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, {}, False, "Voir tous les tags", gr.update(interactive=False)
-
-    # Activer bouton si URL http(s)
-    def enable_btn(link): return gr.update(interactive=link.startswith(("http://", "https://")))
-    url_in.change(enable_btn, url_in, url_btn)
+        return None, {}, fig, pd.DataFrame(), {}, False, "Voir tous les tags"
 
-    # Bouton Analyse lien
-    url_btn.click(analyse_link, [url_in, expanded], [tags_out, spec_out, full_tags])
+    audio_in.upload(analyse,
+                    [audio_in, expanded],
+                    [tags_out, spec_out, full_tags, synth_df])
 
-    # Upload local
-    audio_in.upload(analyse, [audio_in, expanded], [tags_out, spec_out, full_tags])
-
-    # Toggle
     toggle_btn.click(flip, expanded, expanded)\
-              .then(refresh, [expanded, full_tags], [toggle_btn, tags_out])
+              .then(refresh,
+                    [expanded, full_tags, synth_df],
+                    [toggle_btn, tags_out, synth_out])
 
-    # Reset
     reset_btn.click(
         reset_ui, None,
-        [audio_in, url_in, tags_out, spec_out, full_tags, expanded, toggle_btn]
+        [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 hybrydes</center>")
+    gr.Markdown("<center>2025 : Gaspard Boréal / La comédie des mondes hybrides</center>")
 
 # ---------------------------- 5. Run ----------------------------------
 if __name__ == "__main__":