"""Export the inference model from a full training checkpoint (last.pt or step-*.pt). The export holds the speech model only: no optimizer state, no auxiliary mel head and no discriminators. """ import argparse from pathlib import Path import torch from little_tts.config import CHECKPOINT_FORMAT, Config from little_tts.model import LittleTTS from little_tts.text import frontend_fingerprint from little_tts.train import export_model def main(): parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("--checkpoint", type=Path, required=True) parser.add_argument("--output", type=Path, required=True) args = parser.parse_args() state = torch.load(args.checkpoint, map_location="cpu", weights_only=True) if state.get("format_version") != CHECKPOINT_FORMAT or state["frontend"] != frontend_fingerprint(): raise ValueError("checkpoint format or front end does not match this code") model = LittleTTS(Config.from_dict(state["config"])) model.load_state_dict({k.removeprefix("model."): v for k, v in state["system"].items() if k.startswith("model.")}, strict=True) export_model(model, args.output, state["step"], state["annotation"]) print(f"exported step {state['step']} with {sum(p.numel() for p in model.parameters()):,} parameters to {args.output}") if __name__ == "__main__": main()