def load_model_package(cls, package):
# TODO Added for backwards compatibility, should be remove for new release
if OmegaConf.get_type(package['audio_conf']) == dict:
audio_conf = package['audio_conf']
package['audio_conf'] = SpectConfig(sample_rate=audio_conf['sample_rate'],
window_size=audio_conf['window_size'],
window=SpectrogramWindow(audio_conf['window']))
model = cls(rnn_hidden_size=package['hidden_size'],
nb_layers=package['hidden_layers'],
labels=package['labels'],
audio_conf=package['audio_conf'],
rnn_type=supported_rnns[package['rnn_type']],
bidirectional=package.get('bidirectional', True))
model.load_state_dict(package['state_dict'])
return model
class OptimizerConfig:
model_path: str = ''
test_path: str = '' # Path to test manifest or csv
is_character_based: bool = True # Use CER or WER for finding optimal parameters
lm_path: str = ''
beam_width: int = 10
alpha_from: float = 0.0
alpha_to: float = 3.0
beta_from: float = 0.0
beta_to: float = 1.0
n_trials: int = 500 # Number of trials for optuna
n_jobs: int = 2 # Number of parallel jobs for optuna
precision: int = 16
batch_size: int = 1 # For dataloader
num_workers: int = 1 # For dataloader
spect_cfg: SpectConfig = SpectConfig()
def normalize_tensor(x):
return x / x.sum()
if __name__ == '__main__':
LABELS = ["_", "'", "A", "B", "C", "D", "E", "F", "G",
"H", "I", "J", "K", "L", "M", "N", "O", "P",
"Q", "R", "S", "T", "U", "V", "W", "X", "Y",
"Z", " "
]
path_input = '/home/coml/Documents/Victoria/noise_classifier/deepspeech_model/data/CommonVoice_dataset/train'
test_dataset = SpectrogramDataset(audio_conf=SpectConfig(), input_path=path_input, labels=LABELS)
test_loader = AudioDataLoader(dataset=test_dataset)
model = DeepSpeech(labels=LABELS, precision=32, spect_cfg=SpectConfig(),
optim_cfg=AdamConfig(), model_cfg=BiDirectionalConfig()) # args: 'labels', 'model_cfg', 'precision', 'optim_cfg', and 'spect_cfg'
# im = test_loader[0]
NUM_CLASSES = 29 # Corresponds to the length of the labels
# layer = nn.Linear()
for i, data in enumerate(test_loader, 0):
# print('DATA \n', data[1], '\n', data[2], '\n', data[3], '\n')
print('\n Sample {}: input length {}'.format(i, data[3]))
out, length = model.forward(data[0], data[3])
print(' Outputs shapes: {} {}'.format(out.shape, length))
print('Final output', out)
'Raindrop', 'Run', 'Scissors', 'Screaming', 'Shatter', 'Sigh',
'Sink_(filling_or_washing)', 'Skateboard', 'Slam', 'Sneeze', 'Squeak',
'Stream', 'Strum', 'Tap', 'Tick-tock', 'Toilet_flush',
'Traffic_noise_and_roadway_noise', 'Trickle_and_dribble',
'Walk_and_footsteps', 'Water_tap_and_faucet', 'Waves_and_surf',
'Whispering', 'Writing', 'Yell', 'Zipper_(clothing)'
}
LABELS = list(set.difference(ALL_LABELS, HUMAN_LABELS))
import matplotlib
matplotlib.use('TkAgg')
import matplotlib.pyplot as plt
path_input = '/home/coml/Documents/Victoria/noise_classifier/deepspeech_model/data/Freesound_dataset'
data = SpectrogramDataset(audio_conf=SpectConfig(),
input_path=path_input,
labels=LABELS)
print('First element \n', type(data[0][0]), data[0][0].shape,
len(data[0][1]))
print('Second element \n', type(data[1][0]), data[1][0].shape,
len(data[1][1]))
print('Third element \n', type(data[2][0]), data[2][0].shape,
len(data[2][1]))
print('Fourth element \n', type(data[3][0]), data[3][0].shape,
len(data[3][1]))
plt.imshow(data[0][0].detach().numpy())
plt.show()
"""