| |
| """Generate a PROPS speaker-embedding GMM from a natural-language description.""" |
|
|
| from __future__ import annotations |
|
|
| import argparse |
| from pathlib import Path |
|
|
| import numpy as np |
| import torch |
| from sentence_transformers import SentenceTransformer |
|
|
| from model import ComposedGMMMDN |
|
|
|
|
| def main(): |
| parser = argparse.ArgumentParser() |
| parser.add_argument("description", help="Natural-language speaker-profile description") |
| parser.add_argument("--model", default=str(Path(__file__).resolve().parent)) |
| parser.add_argument("--output", type=Path, default=Path("generated_gmm.npz")) |
| parser.add_argument("--device", default="cuda" if torch.cuda.is_available() else "cpu") |
| args = parser.parse_args() |
|
|
| encoder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2", device=args.device) |
| embedding = encoder.encode([args.description], convert_to_tensor=True).to(args.device) |
| model = ComposedGMMMDN.from_pretrained(args.model, device=args.device) |
| with torch.inference_mode(): |
| pi_logits, mu, sigma = model(embedding) |
| pi = torch.softmax(pi_logits, dim=-1) |
| args.output.parent.mkdir(parents=True, exist_ok=True) |
| np.savez_compressed( |
| args.output, |
| pi_logits=pi_logits[0].cpu().numpy().astype(np.float32), |
| pi=pi[0].cpu().numpy().astype(np.float32), |
| mu=mu[0].cpu().numpy().astype(np.float32), |
| sigma=sigma[0].cpu().numpy().astype(np.float32), |
| description=np.asarray(args.description), |
| ) |
| print(args.output) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
