app.py

import gradio as gr
import numpy as np
import librosa
from perch_hoplite.zoo import model_configs

# Utiliser le nom de modèle "perch_8"
MODEL = model_configs.load_model_by_name("perch_8")
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"}
    sr, wav = audio
    wav = _prep(wav, sr)

    out = MODEL.embed(wav)
    logits = out.logits["label"]
    labels = out.label_names.get("label")

    idx = np.argsort(logits)[::-1][:3]
    topk = []
    
    top_logits = logits[idx]
    exp_logits = np.exp(top_logits - np.max(top_logits))
    sum_exp_logits = np.sum(exp_logits)
    
    for i in range(len(idx)):
        class_index = idx[i]
        name = labels[class_index] if labels is not None and class_index < len(labels) else f"classe_{int(class_index)}"
        prob = float(exp_logits[i] / sum_exp_logits)
        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."
    }

# CORRECTION : Retrait du paramètre obsolète "allow_flagging"
demo = gr.Interface(
    fn=infer,
    inputs=gr.Audio(type="numpy", sources=["microphone", "upload"]),
    outputs=gr.JSON(label="Perch 8 Inference"),
    title="Perch Bioacoustics"
)

# CORRECTION : Retrait de .queue() et du paramètre api_open
# L'API est activée par défaut dans cette version de Gradio.
demo.launch()