finetune_xtts_hf.py

import os
import subprocess
from huggingface_hub import HfApi, HfFolder, upload_folder

# === Configuración ===
HF_MODEL_ID = "tu_usuario/xtts-v2-finetuned"   # <--- cambia con tu repo en HF
HF_TOKEN = os.environ.get("HF_TOKEN")          # Debe estar definido en tu Space/entorno
DATASET_PATH = "/home/user/app/dataset"        # Ruta a tu dataset
OUTPUT_PATH = "/tmp/output_model"
BASE_MODEL = "coqui/XTTS-v2"

# === 1. Descargar modelo base de HF ===
print("=== Descargando modelo base desde Hugging Face ===")
subprocess.run([
    "huggingface-cli", "download", BASE_MODEL, "--local-dir", "./xtts_model"
], check=True)

CONFIG_PATH = "./xtts_model/config.json"
RESTORE_PATH = "./xtts_model/model.pth"

# === 2. Editar configuración para tu dataset VoxPopuli ===
print("=== Editando configuración para fine-tuning con VoxPopuli ===")
import json
with open(CONFIG_PATH, "r") as f:
    config = json.load(f)

config["output_path"] = OUTPUT_PATH
config["datasets"] = [
    {
        "formatter": "voxpopuli",
        "path": DATASET_PATH,
        "meta_file_train": "metadata.json"
    }
]
config["run_name"] = "xtts-finetune-voxpopuli"
config["lr"] = 1e-5  # más bajo para fine-tuning

with open(CONFIG_PATH, "w") as f:
    json.dump(config, f, indent=2)

# === 3. Lanzar entrenamiento ===
print("=== Iniciando fine-tuning de XTTS-v2 ===")
subprocess.run([
    "python", "TTS/bin/train_tts.py",
    "--config_path", CONFIG_PATH,
    "--restore_path", RESTORE_PATH
], check=True)

# === 4. Subir modelo resultante a HF ===
print("=== Subiendo modelo fine-tuneado a Hugging Face Hub ===")
api = HfApi()
HfFolder.save_token(HF_TOKEN)

upload_folder(
    repo_id=HF_MODEL_ID,
    repo_type="model",
    folder_path=OUTPUT_PATH,
    token=HF_TOKEN
)

print("✅ Fine-tuning completado y modelo subido a Hugging Face.")