Spaces:
Sleeping
Sleeping
import argparse | |
import os | |
from Preprocessing.multilinguality.create_distance_lookups import CacheCreator | |
from Preprocessing.multilinguality.create_lang_dist_dataset import LangDistDatasetCreator | |
from Preprocessing.multilinguality.generate_zero_shot_lang_embs import approximate_and_inject_language_embeddings | |
from Utility.storage_config import MODELS_DIR | |
if __name__ == "__main__": | |
default_model_path = os.path.join(MODELS_DIR, "ToucanTTS_Meta", "best.pt") | |
parser = argparse.ArgumentParser() | |
parser.add_argument("--model_path", "-m", type=str, default=default_model_path, help="model path from which to obtain pretrained language embeddings") | |
parser.add_argument("--distance_type", "-d", type=str, choices=["map", "tree", "asp", "learned", "combined"], default="learned", | |
help="which type of distance to use for finding nearest languages") | |
parser.add_argument("--n_closest", "-k", type=int, default=50, help="how many nearest languages to select for language embedding approximation") | |
args = parser.parse_args() | |
# make sure that cache files exist | |
cc = CacheCreator(cache_root="Preprocessing/multilinguality") | |
cc.create_required_files(model_path=os.path.join(MODELS_DIR, "ToucanTTS_Meta", "best.pt")) | |
# create distance dataset | |
dc = LangDistDatasetCreator(args.model_path, cache_root="Preprocessing/multilinguality") | |
distance_dataset = dc.create_dataset(args.distance_type, n_closest=args.n_closest, zero_shot=True) | |
# generate zero-shot lang embs and inject into pretrained model, then save to modified model path | |
approximate_and_inject_language_embeddings(model_path=args.model_path, | |
df=distance_dataset, | |
iso_lookup=dc.iso_lookup) | |