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 _testing_process(data, index, mad, device, seq_length,
context_length, window_size, batch_size,
hop_size, outputPath):
"""The testing process over testing data.
:param data: The testing data.
:type data: numpy.ndarray
:param index: The index of the testing data (used for\
calculating scores).
:type index: int
:param mad: The MaD system.
:type mad: torch.nn.Module
:param device: The device to be used.
:type device: str
:param seq_length: The sequence length used.
:type seq_length: int
:param context_length: The context length used.
:type context_length: int
:param window_size: The window size used.
:type window_size: int
:param batch_size: The batch size used.
:type batch_size: int
:param hop_size: The hop size used.
:type hop_size: int
:return: The SDR and SIR scores, and the time elapsed for\
the process.
:rtype: (numpy.ndarray, numpy.ndarray, float)
"""
s_time = time.time()
mix, mix_magnitude, mix_phase, voice_true, bg_true = data
voice_predicted = np.zeros((
mix_magnitude.shape[0],
seq_length - context_length * 2,
window_size), dtype=np.float32)
for batch in range(int(mix_magnitude.shape[0] / batch_size)):
b_start = batch * batch_size
b_end = (batch + 1) * batch_size
v_in = from_numpy(
mix_magnitude[b_start:b_end, :, :]).to(device)
voice_predicted[b_start:b_end, :, :] = mad(
v_in.unsqueeze(1)).v_j_filt.cpu().numpy()
tmp_sdr, tmp_sir, tmp_sar = data_feeder.data_process_results_testing(
index=index, voice_true=voice_true,
bg_true=bg_true, voice_predicted=voice_predicted,
window_size=window_size, mix=mix,
mix_magnitude=mix_magnitude,
mix_phase=mix_phase, hop=hop_size,
context_length=context_length, outputPath=outputPath)
time_elapsed = time.time() - s_time
printing.print_msg(testing_output_string_per_example.format(
e=index,
sdr=np.median([i for i in tmp_sdr[0] if not np.isnan(i)]),
sir=np.median([i for i in tmp_sir[0] if not np.isnan(i)]),
sar=np.median([i for i in tmp_sar[0] if not np.isnan(i)]),
t=time_elapsed
))
return tmp_sdr, tmp_sir,tmp_sar, time_elapsed
def testing_process():
"""The testing process.
"""
device = 'cuda' if not debug and torch.cuda.is_available() else 'cpu'
print('\n-- Starting testing process. Debug mode: {}'.format(debug))
print('-- Process on: {}'.format(device), end='\n\n')
print('-- Setting up modules... ', end='')
# 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'])).to(device)
rnn_dec.load_state_dict(torch.load(
output_states_path['rnn_dec'])).to(device)
fnn.load_state_dict(torch.load(output_states_path['fnn'])).to(device)
denoiser.load_state_dict(torch.load(
output_states_path['denoiser'])).to(device)
print('done.')
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)
print('-- Testing starts\n')
sdr = []
sir = []
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)
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()
tmp_sdr, tmp_sir = 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'])
e_time = time.time()
print(
testing_output_string_per_example.format(
e=index,
sdr=np.median([i for i in tmp_sdr[0] if not np.isnan(i)]),
sir=np.median([i for i in tmp_sir[0] if not np.isnan(i)]),
t=e_time - s_time))
total_time += e_time - s_time
sdr.append(tmp_sdr)
sir.append(tmp_sir)
print('\n-- Testing finished\n')
print(
testing_output_string_all.format(
sdr=np.median([ii for i in sdr for ii in i[0]
if not np.isnan(ii)]),
sir=np.median([ii for i in sir for ii in i[0]
if not np.isnan(ii)]),
t=total_time))
print('\n-- Saving results... ', end='')
with open(metrics_paths['sdr'], 'wb') as f:
pickle.dump(sdr, f, protocol=2)
with open(metrics_paths['sir'], 'wb') as f:
pickle.dump(sir, f, protocol=2)
print('done!')
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!')