def use_me_process(sources_list, output_file_names):
"""The usage process.
:param sources_list: The file names to be used.
:type sources_list: list[str]
:param output_file_names: The output file names to be used.
:type output_file_names: list[list[str]]
"""
print('\n-- Welcome to MaD TwinNet.')
if debug:
print(
'\n-- Cannot proceed in debug mode. Please set debug=False at the settings file.'
)
print('-- Exiting.')
exit(-1)
print(
'-- Now I will extract the voice and the background music from the provided files'
)
# Masker modules
rnn_enc = RNNEnc(hyper_parameters['reduced_dim'],
hyper_parameters['context_length'], debug)
rnn_dec = RNNDec(hyper_parameters['rnn_enc_output_dim'], debug)
fnn = FNNMasker(hyper_parameters['rnn_enc_output_dim'],
hyper_parameters['original_input_dim'],
hyper_parameters['context_length'])
# Denoiser modules
denoiser = FNNDenoiser(hyper_parameters['original_input_dim'])
rnn_enc.load_state_dict(torch.load(output_states_path['rnn_enc']))
rnn_dec.load_state_dict(torch.load(output_states_path['rnn_dec']))
fnn.load_state_dict(torch.load(output_states_path['fnn']))
denoiser.load_state_dict(torch.load(output_states_path['denoiser']))
if not debug and torch.has_cudnn:
rnn_enc = rnn_enc.cuda()
rnn_dec = rnn_dec.cuda()
fnn = fnn.cuda()
denoiser = denoiser.cuda()
testing_it = data_feeder_testing(
window_size=hyper_parameters['window_size'],
fft_size=hyper_parameters['fft_size'],
hop_size=hyper_parameters['hop_size'],
seq_length=hyper_parameters['seq_length'],
context_length=hyper_parameters['context_length'],
batch_size=1,
debug=debug,
sources_list=sources_list)
print('-- Let\'s go!\n')
total_time = 0
for index, data in enumerate(testing_it()):
s_time = time.time()
mix, mix_magnitude, mix_phase, voice_true, bg_true = data
voice_predicted = np.zeros(
(mix_magnitude.shape[0], hyper_parameters['seq_length'] -
hyper_parameters['context_length'] * 2,
hyper_parameters['window_size']),
dtype=np.float32)
for batch in range(
int(mix_magnitude.shape[0] /
training_constants['batch_size'])):
b_start = batch * training_constants['batch_size']
b_end = (batch + 1) * training_constants['batch_size']
v_in = Variable(
torch.from_numpy(mix_magnitude[b_start:b_end, :, :]))
if not debug and torch.has_cudnn:
v_in = v_in.cuda()
tmp_voice_predicted = rnn_enc(v_in)
tmp_voice_predicted = rnn_dec(tmp_voice_predicted)
tmp_voice_predicted = fnn(tmp_voice_predicted, v_in)
tmp_voice_predicted = denoiser(tmp_voice_predicted)
voice_predicted[
b_start:b_end, :, :] = tmp_voice_predicted.data.cpu().numpy()
data_process_results_testing(
index=index,
voice_true=voice_true,
bg_true=bg_true,
voice_predicted=voice_predicted,
window_size=hyper_parameters['window_size'],
mix=mix,
mix_magnitude=mix_magnitude,
mix_phase=mix_phase,
hop=hyper_parameters['hop_size'],
context_length=hyper_parameters['context_length'],
output_file_name=output_file_names[index])
e_time = time.time()
print(
usage_output_string_per_example.format(f=sources_list[index],
t=e_time - s_time))
total_time += e_time - s_time
print('\n-- Testing finished\n')
print(usage_output_string_total.format(t=total_time))
print('-- That\'s all folks!')
def use_me_process(sources_list, output_file_names):
"""The usage process.
:param sources_list: The file names to be used.
:type sources_list: list[pathlib.Path]
:param output_file_names: The output file names to be used.
:type output_file_names: list[list[str]]
"""
printing.print_msg('Welcome to MaD TwinNet.', end='\n\n')
if debug:
printing.print_msg('Cannot proceed in debug mode. '
'Please set `debug=False` at the settings '
'file.')
printing.print_msg('Exiting.')
exit(-1)
printing.print_msg('Now I will extract the voice and the '
'background music from the provided files')
device = 'cuda' if not debug and torch.cuda.is_available() else 'cpu'
# MaD setting up
mad = MaD(rnn_enc_input_dim=hyper_parameters['reduced_dim'],
rnn_dec_input_dim=hyper_parameters['rnn_enc_output_dim'],
original_input_dim=hyper_parameters['original_input_dim'],
context_length=hyper_parameters['context_length'])
mad.load_state_dict(torch.load(output_states_path['mad']))
mad = mad.to(device).eval()
testing_it = data_feeder.data_feeder_testing(
window_size=hyper_parameters['window_size'],
fft_size=hyper_parameters['fft_size'],
hop_size=hyper_parameters['hop_size'],
seq_length=hyper_parameters['seq_length'],
context_length=hyper_parameters['context_length'],
batch_size=1,
debug=debug,
sources_list=sources_list)
printing.print_msg('Let\'s go!', end='\n\n')
total_time = 0
for index, data in enumerate(testing_it()):
s_time = time.time()
mix, mix_magnitude, mix_phase, voice_true, bg_true = data
voice_predicted = np.zeros(
(mix_magnitude.shape[0], hyper_parameters['seq_length'] -
hyper_parameters['context_length'] * 2,
hyper_parameters['window_size']),
dtype=np.float32)
for batch in range(
int(mix_magnitude.shape[0] /
training_constants['batch_size'])):
b_start = batch * training_constants['batch_size']
b_end = (batch + 1) * training_constants['batch_size']
v_in = torch.from_numpy(
mix_magnitude[b_start:b_end, :, :]).to(device)
voice_predicted[b_start:b_end, :, :] = mad(
v_in).v_j_filt.cpu().numpy()
data_feeder.data_process_results_testing(
index=index,
voice_true=voice_true,
bg_true=bg_true,
voice_predicted=voice_predicted,
window_size=hyper_parameters['window_size'],
mix=mix,
mix_magnitude=mix_magnitude,
mix_phase=mix_phase,
hop=hyper_parameters['hop_size'],
context_length=hyper_parameters['context_length'],
output_file_name=output_file_names[index])
e_time = time.time()
printing.print_msg(
usage_output_string_per_example.format(f=sources_list[index],
t=e_time - s_time))
total_time += e_time - s_time
printing.print_msg('MaDTwinNet finished')
printing.print_msg(usage_output_string_total.format(t=total_time))
printing.print_msg('That\'s all folks!')