to_extracted_epochs = [
"21"
] # It is model's name, such as 10.params or final.params (suffix is w.r.t package).
nj = 10
force = False
use_gpu = True
gpu_id = ""
sleep_time = 10
# Run a batch extracting process.
try:
for position in to_extracted_positions:
# Generate the extracting config from nnet config where
# which position to extract depends on the 'extracted_embedding' parameter of model_creation (by my design).
model_blueprint, model_creation = utils.read_nnet_config(
"{0}/config/nnet.config".format(model_dir))
model_creation = model_creation.replace(
"training=True", "training=False"
) # To save memory without loading some independent components.
model_creation = model_creation.replace(
model_params["extracted_embedding"], position)
extract_config = "{0}.extract.config".format(position)
utils.write_nnet_config(
model_blueprint, model_creation,
"{0}/config/{1}".format(model_dir, extract_config))
for epoch in to_extracted_epochs:
model_file = "{0}.{1}".format(epoch, suffix)
point_name = "{0}_epoch_{1}".format(position, epoch)
# If run a trainer with background thread (do not be supported now) or run this launcher extrally with stage=4
# (it means another process), then this while-listen is useful to start extracting immediately (but require more gpu-memory).
parser.add_argument("feats_rspecifier", metavar="feats-rspecifier",
type=str, help="")
parser.add_argument("vectors_wspecifier", metavar="vectors-wspecifier",
type=str, help="")
print(' '.join(sys.argv))
args = parser.parse_args()
# Start
try:
# nnet_config include model_blueprint and model_creation
if args.nnet_config != "":
model_blueprint, model_creation = utils.read_nnet_config(args.nnet_config)
elif args.model_blueprint is not None and args.model_creation is not None:
model_blueprint = args.model_blueprint
model_creation = args.model_creation
else:
raise ValueError("Expected nnet_config or (model_blueprint, model_creation) to exist.")
model = utils.create_model_from_py(model_blueprint, model_creation)
model.load_state_dict(torch.load(args.model_path, map_location='cpu'), strict=False)
# Select device
model = utils.select_model_device(model, args.use_gpu, gpu_id=args.gpu_id)
model.eval()
with kaldi_io.open_or_fd(args.feats_rspecifier, "rb") as r, \