extract_mert_embedding.py

import os
import torch
import soundfile as sf
import librosa
import pandas as pd
import numpy as np
from transformers import AutoProcessor, AutoModel
from tqdm import tqdm
from datetime import datetime

# =============================
# CONFIG
# =============================
ROOT_DATA = "/lium/raid-b/mshamsi/FreeSound_Popularity/"
OUTPUT_PATH = "embeddings_mert_all_datasets.csv"
LOG_PATH = "errors_mert.log"

TARGET_SR = 24000
MAX_DURATION = 60  # secondes
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

AUDIO_EXTENSIONS = (".wav", ".WAV", ".mp3", ".flac", ".ogg", ".m4a")

# =============================
# INIT LOG
# =============================
with open(LOG_PATH, "w") as f:
    f.write("=== MERT EXTRACTION LOG ===\n")
    f.write(str(datetime.now()) + "\n\n")

def log_error(msg):
    with open(LOG_PATH, "a") as f:
        f.write(msg + "\n")

# =============================
# LOAD MODEL
# =============================
try:
    processor = AutoProcessor.from_pretrained(
        "m-a-p/MERT-v1-330M",
        trust_remote_code=True
    )

    model = AutoModel.from_pretrained(
        "m-a-p/MERT-v1-330M",
        trust_remote_code=True
    ).to(DEVICE)

    model.eval()

except Exception as e:
    log_error(f"[FATAL] Model loading failed: {e}")
    raise RuntimeError("Impossible de charger le modèle MERT")


# =============================
# LOOP OVER DATASETS
# =============================
datasets = [
    d for d in os.listdir(ROOT_DATA)
    if os.path.isdir(os.path.join(ROOT_DATA, d))
]

for dataset_name in datasets:
    dataset_path = os.path.join(ROOT_DATA, dataset_name)

    # =============================
    # STORAGE
    # =============================
    rows = []
    processed = 0
    skipped = 0

    for batch in ["batch_001", "batch_002"]:
        batch_path = os.path.join(dataset_path, batch)

        if not os.path.exists(batch_path):
            log_error(f"[INFO] Missing folder: {batch_path}")
            continue

        audio_files = [
            f for f in os.listdir(batch_path)
            if f.lower().endswith(AUDIO_EXTENSIONS)
        ]

        for audio_file in tqdm(audio_files, desc=f"{dataset_name}/{batch}"):
            audio_path = os.path.join(batch_path, audio_file)

            try:
                # =============================
                # LOAD AUDIO (SAFE)
                # =============================
                audio, sr = sf.read(audio_path, always_2d=False)

                if audio is None or len(audio) == 0:
                    raise ValueError("Empty audio file")

                # Stereo → mono
                if audio.ndim > 1:
                    audio = np.mean(audio, axis=1)

                # Convert to float32
                audio = audio.astype(np.float32)

                # Resample
                if sr != TARGET_SR:
                    audio = librosa.resample(
                        audio,
                        orig_sr=sr,
                        target_sr=TARGET_SR
                    )

                # Trim / pad duration
                max_len = TARGET_SR * MAX_DURATION
                audio = audio[:max_len]

                # =============================
                # PROCESS WITH MERT
                # =============================
                inputs = processor(
                    audio,
                    sampling_rate=TARGET_SR,
                    return_tensors="pt"
                )
                inputs = {k: v.to(DEVICE) for k, v in inputs.items()}

                with torch.no_grad():
                    outputs = model(**inputs)

                if not hasattr(outputs, "last_hidden_state"):
                    raise RuntimeError("Invalid model output")

                embedding = (
                    outputs.last_hidden_state
                    .mean(dim=1)
                    .squeeze()
                    .cpu()
                    .numpy()
                )

                # =============================
                # SAVE ROW
                # =============================
                row = {
                    "dataset": dataset_name,
                    "batch": batch,
                    "filename": audio_file
                }

                for i, val in enumerate(embedding):
                    row[f"mert_{i}"] = float(val)

                rows.append(row)
                processed += 1

            except Exception as e:
                skipped += 1
                log_error(f"[ERROR] {audio_path} -> {e}")

    # =============================
    # SAVE CSV
    # =============================
    df = pd.DataFrame(rows)
    df.to_csv(os.path.join(dataset_path, OUTPUT_PATH), index=False)

    print("\n=== EXTRACTION TERMINÉE ===")
    print(f"Dataset : {dataset_name}")
    print(f"Fichiers traités : {processed}")
    print(f"Fichiers ignorés : {skipped}")
    print(f"CSV sauvegardé : {OUTPUT_PATH}")
    print(f"Log erreurs : {LOG_PATH}")