def load_model(device, model_path, use_half):
model = DeepSpeech.load_model(model_path)
model.eval()
model = model.to(device)
if use_half:
model = model.half()
return model
parser.add_argument('--lm-num-alphas',
default=45,
type=float,
help='Number of alpha candidates for tuning')
parser.add_argument('--lm-num-betas',
default=8,
type=float,
help='Number of beta candidates for tuning')
parser = add_decoder_args(parser)
args = parser.parse_args()
if args.lm_path is None:
print("error: LM must be provided for tuning")
sys.exit(1)
model = DeepSpeech.load_model(args.model_path)
saved_output = np.load(args.saved_output)
def init(beam_width, blank_index, lm_path):
global decoder
decoder = BeamCTCDecoder(model.labels,
lm_path=lm_path,
beam_width=beam_width,
num_processes=args.lm_workers,
blank_index=blank_index)
def decode_dataset(params):
lm_alpha, lm_beta = params
import argparse
import numpy as np
from deepspeech.model import DeepSpeech
from deepspeech.data.data_loader import SpectrogramParser
from noswear.model import load_model
parser = argparse.ArgumentParser()
parser.add_argument('audio_file',
type=argparse.FileType('r'),
help='File to classify')
args = parser.parse_args()
base_model = DeepSpeech.load_model('models/librispeech_pretrained.pth')
audio_conf = DeepSpeech.get_audio_conf(base_model)
parser = SpectrogramParser(audio_conf, normalize=True)
net = load_model(base_model, {'f_pickle': 'models/binary_clf.pkl'})
print(net)
fpath = args.audio_file.name
audio = parser.parse_audio(fpath)
X = {'lens': np.array([audio.shape[1]]), 'X': np.array(audio)[None]}
y_pred = net.predict(X)
print(y_pred[0] and 'swear! :(' or 'noswear :)')