Create app.py

app.py

@@ -0,0 +1,51 @@
+import gradio as gr, numpy as np, librosa, soundfile as sf
+from perch_hoplite.zoo import model_configs
+
+# Charge Perch v2 (télécharge depuis Kaggle via hoplite)
+MODEL = model_configs.load_model_by_name("perch_v2")
+SR = 32000
+WIN = 5 * SR
+
+def _prep(wav, sr):
+    if wav.ndim > 1:
+        wav = np.mean(wav, axis=1)
+    if sr != SR:
+        wav = librosa.resample(wav.astype(np.float32), orig_sr=sr, target_sr=SR)
+    if len(wav) < WIN:
+        wav = np.pad(wav, (0, WIN - len(wav)))
+    else:
+        wav = wav[:WIN]
+    return wav.astype(np.float32)
+
+def infer(audio):
+    if audio is None:
+        return {"error": "no audio"}
+    wav, sr = audio
+    wav = _prep(wav, sr)
+
+    out = MODEL.embed(wav)
+    logits = out.logits["label"]
+    labels = out.label_names["label"] if hasattr(out, "label_names") else None
+
+    idx = np.argsort(logits)[::-1][:3]
+    topk = []
+    for i in idx:
+        name = labels[i] if labels is not None else f"class_{int(i)}"
+        prob = float(np.exp(logits[i]) / np.sum(np.exp(logits[idx])))
+        topk.append({"label": name, "score": round(prob, 4)})
+
+    return {
+        "topk": topk,
+        "embedding_dim": int(out.embeddings.shape[-1]),
+        "note": "scores non calibrés; régler un seuil selon votre usage"
+    }
+
+demo = gr.Interface(
+    fn=infer,
+    inputs=gr.Audio(type="numpy", sources=["microphone", "upload"]),
+    outputs=gr.JSON(label="Perch v2"),
+    title="Perch 2.0 — Bioacoustics",
+    allow_flagging="never"
+)
+
+demo.queue(api_open=True).launch()