def main():
parser = argparse.ArgumentParser()
parser.add_argument('--hp_file', type=str, default='hparams.py')
args = parser.parse_args()
hp.configure(args.hp_file)
fill_variables(hp)
log_config(hp)
os.makedirs(hp.save_dir, exist_ok=True)
n_gpus = torch.cuda.device_count()
args.__setattr__('n_gpus', n_gpus)
if n_gpus > 1:
run_distributed(run_training, args, hp)
else:
run_training(0, args, hp, None)
parser.add_argument('--log_params', action='store_true')
parser.add_argument('--calc_wer', action='store_true')
parser.add_argument('--segment', type=int, default=10000)
parser.add_argument('--silence_file', type=str, default=None)
parser.add_argument('--lm_type', type=str, default='LSTM')
args = parser.parse_args()
hp_file = args.hp_file
model_name = args.load_name # save dir name
model_path = os.path.dirname(model_name)
if hp_file is None:
hp_file = os.path.join(model_path, 'hparams.py')
hp.configure(hp_file)
fill_variables(hp)
setattr(hp, 'silence_file', args.silence_file)
if args.beam_width is not None:
print(f'beam width is set to {args.beam_width}')
hp.beam_width = args.beam_width
script_file = hp.eval_file
if args.test_script is not None:
script_file = args.test_script
if hp.lm_weight is not None:
args.lm_weight = hp.lm_weight
print(f'lm weight = {args.lm_weight}')
model = Transformer(hp)
hp.save_dir + "/network.optimizer.epoch{}".format(epoch + 1))
adjust_learning_rate(optimizer, epoch + 1)
if (epoch + 1) % hp.reset_optimizer_epoch == 0:
optimizer = torch.optim.Adam(model.parameters(), weight_decay=1e-5)
print("EPOCH {} end".format(epoch + 1))
print(f'elapsed time = {(time.time()-start_time)//60}m')
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--hp_file', metavar='FILE', default='hparams.py')
args = parser.parse_args()
#overwrite_hparams(args)
hp.configure(args.hp_file)
fill_variables()
hp.save_dir = os.path.join(hp.save_dir, 'LM')
os.makedirs(hp.save_dir, exist_ok=True)
if hp.debug_mode == 'tensorboard':
writer = SummaryWriter(f'{hp.save_dir}/logs/{hp.comment}')
log_config()
model = Model_lm(hp)
model.apply(init_weight)
if torch.cuda.device_count() > 1:
# multi-gpu configuration
ngpu = torch.cuda.device_count()
device_ids = list(range(ngpu))
parser.add_argument('--test_script', type=str, default=None)
parser.add_argument('--load_name_lm', type=str, default=None)
parser.add_argument('--lm_weight', type=float, default=None)
parser.add_argument('--log_params', action='store_true')
args = parser.parse_args()
load_name = args.load_name
load_dir = os.path.dirname(load_name)
if os.path.exists(os.path.join(load_dir, 'hparams.py')):
args.hp_file = os.path.join(load_dir, 'hparams.py')
#hp = utils.HParams()
hp.configure(args.hp_file)
fill_variables(hp, args.log_params)
overwrite_hparams(args, hp)
if hp.decoder_type == 'Attention':
model = AttModel(hp)
elif hp.decoder_type == 'CTC':
model = CTCModel(hp)
if hp.load_name_lm is not None:
hp_LM_path = os.path.join(os.path.dirname(hp.load_name_lm), 'hparams.py')
hp_LM = utils.HParams()
hp_LM.configure(hp_LM_path)
model_lm = Model_lm(hp_LM)
model_lm.to(DEVICE)
model_lm.load_state_dict(load_model(hp.load_name_lm))
model_lm.eval()