assert np.array_equal(
test_labels_acoustic, test_labels_linguistic
), "Labels for acoustic and linguistic datasets are not the same!"
"""Choosing hardware"""
device = 'cuda' if torch.cuda.is_available() else 'cpu'
if device == "cuda":
print(
"Using GPU. Setting default tensor type to torch.cuda.FloatTensor")
torch.set_default_tensor_type("torch.cuda.FloatTensor")
else:
print("Using CPU. Setting default tensor type to torch.FloatTensor")
torch.set_default_tensor_type("torch.FloatTensor")
"""Converting model to specified hardware and format"""
acoustic_cfg_json = json.load(
open(args.acoustic_model.replace(".torch", ".json"), "r"))
acoustic_cfg = AcousticConfig.from_json(acoustic_cfg_json)
acoustic_model = CNN(acoustic_cfg)
acoustic_model.float().to(device)
try:
acoustic_model.load_state_dict(torch.load(args.acoustic_model))
except:
print(
"Failed to load model from {} without device mapping. Trying to load with mapping to {}"
.format(args.acoustic_model, device))
acoustic_model.load_state_dict(
torch.load(args.acoustic_model, map_location=device))
linguistic_cfg_json = json.load(
open(args.linguistic_model.replace(".torch", ".json"), "r"))
linguistic_cfg = LinguisticConfig.from_json(linguistic_cfg_json)
params["n_layers"] = np.random.randint(1, 4)
params["hidden_dim"] = np.random.randint(10, 100)
params["dropout"] = 0.5 + np.random.rand() * 0.4
params["dropout2"] = 0.5 + np.random.rand() * 0.45
params["reg_ratio"] = np.random.rand() * 0.0015
params["batch_size"] = np.random.randint(26, 256)
params["bidirectional"] = bool(np.random.randint(0, 2))
cfg = AcousticLLDConfig(**params)
model = RNN(cfg)
elif args.model_type == "acoustic-spectrogram":
test_features, test_labels, val_features, val_labels, train_features, train_labels = load_spectrogram_dataset(
)
params["fc_size"] = np.random.randint(10, 200)
params["dropout"] = 0.3 + np.random.rand() * 0.6
cfg = AcousticSpectrogramConfig(**params)
model = CNN(cfg)
else:
raise Exception(
"model_type parameter has to be one of [linguistic|acoustic-lld|acoustic-spectrogram]"
)
print(
"Subsets sizes: test_features:{}, test_labels:{}, val_features:{}, val_labels:{}, train_features:{}, train_labels:{}"
.format(test_features.shape[0], test_labels.shape[0],
val_features.shape[0], val_labels.shape[0],
train_features.shape[0], train_labels.shape[0]))
"""Converting model to specified hardware and format"""
model.float()
model = model.to(get_device())
parser = argparse.ArgumentParser()
parser.add_argument("-m", "--model_type", type=str, default="linguistic")
args = parser.parse_args()
if args.model_type == "linguistic":
cfg = LinguisticConfig()
test_features, test_labels, val_features, val_labels, train_features, train_labels = load_linguistic_dataset(
)
model = RNN(cfg)
elif args.model_type == "acoustic-lld":
cfg = AcousticLLDConfig()
test_features, test_labels, val_features, val_labels, train_features, train_labels = load_acoustic_features_dataset(
)
model = RNN(cfg)
elif args.model_type == "acoustic-spectrogram":
cfg = AcousticSpectrogramConfig()
test_features, test_labels, val_features, val_labels, train_features, train_labels = load_spectrogram_dataset(
)
model = CNN(cfg)
else:
raise Exception(
"model_type parameter has to be one of [linguistic|acoustic-lld|acoustic-spectrogram]"
)
print(
"Subsets sizes: test_features:{}, test_labels:{}, val_features:{}, val_labels:{}, train_features:{}, train_labels:{}"
.format(test_features.shape[0], test_labels.shape[0],
val_features.shape[0], val_labels.shape[0],
train_features.shape[0], train_labels.shape[0]))
"""Running training"""
run_training(model, cfg, test_features, test_labels, train_features,
dtype=np.int16)
sd.wait() # Wait until recording is finished
print("Recording finished")
write(TMP_FILENAME, SAMPLE_RATE, myrecording) # Save as WAV file
from deepspeech_generator import speech_to_text
transcription = speech_to_text(join(args.deepspeech, "output_graph.pbmm"),
join(args.deepspeech, "alphabet.txt"),
join(args.deepspeech, "lm.binary"),
join(args.deepspeech, "trie"), TMP_FILENAME)
print(transcription)
"""Converting model to specified hardware and format"""
acoustic_cfg_json = json.load(
open(args.acoustic_model.replace(".torch", ".json"), "r"))
acoustic_cfg = AcousticSpectrogramConfig.from_json(acoustic_cfg_json)
acoustic_model = CNN(acoustic_cfg)
acoustic_model.float().to("cpu")
try:
acoustic_model.load_state_dict(torch.load(args.acoustic_model))
except:
print(
"Failed to load model from {} without device mapping. Trying to load with mapping to {}"
.format(args.acoustic_model, "cpu"))
acoustic_model.load_state_dict(
torch.load(args.acoustic_model, map_location="cpu"))
acoustic_model.eval()