def infer_random(self, gender=False):
#using the first sample from in_test
if gender:
source_sp = random.sample(self.speakers_man, 1)[0]
target_sp = random.sample(self.speakers_woman, 1)[0]
else:
source_sp, target_sp = random.sample(self.speakers, 2)
source_utt = random.sample(self.indexes[source_sp], 1)[0]
target_utt = random.sample(self.indexes[target_sp], 1)[0]
source_sp = '3436'
target_sp = '6064'
source_utt = '172162_000012_000003'
target_utt = '300880_000029_000002'
# source_sp = '226'
# target_sp = '238'
# source_utt = '001'
# target_utt = '001'
display = False
output_image_path = os.path.join(
self.args.output_image_path, self.args.val_set,
self.args.load_iteration,
's{}_{}---t{}_{}'.format(source_sp, source_utt, target_sp,
target_utt))
output_audio_path = os.path.join(
self.args.output_audio_path, self.args.val_set,
self.args.load_iteration,
's{}_{}---t{}_{}'.format(source_sp, source_utt, target_sp,
target_utt))
os.makedirs(output_image_path, exist_ok=True)
os.makedirs(output_audio_path, exist_ok=True)
source_spec_ori = self.dataset[f'{source_sp}/{source_utt}/mel'][:]
target_spec_ori = self.dataset[f'{target_sp}/{target_utt}/mel'][:]
print('source time {} | target time {}'.format(source_spec_ori.shape,
target_spec_ori.shape))
b = 8
source_spec, source_len_pad = self.pad_seq(source_spec_ori, base=b)
target_spec, target_len_pad = self.pad_seq(target_spec_ori, base=b)
source_spec = cc(torch.from_numpy(source_spec))
target_spec = cc(torch.from_numpy(target_spec))
source_dspec = source_spec
target_dspec = target_spec
wav2wav(self.speaker2filenames[source_sp][source_utt],
os.path.join(output_audio_path, 'src_ori.wav'))
wav2wav(self.speaker2filenames[target_sp][target_utt],
os.path.join(output_audio_path, 'tar_ori.wav'))
# sync
self.plot_spectrograms(source_spec_ori,
os.path.join(output_image_path, 'src_sync.png'))
self.wave_generate(source_spec_ori,
os.path.join(output_audio_path, 'src_sync'))
self.plot_spectrograms(target_spec_ori,
os.path.join(output_image_path, 'tar_sync.png'))
self.wave_generate(target_spec_ori,
os.path.join(output_audio_path, 'tar_sync'))
# reconstruction
source_spec_rec = self.inference_one_utterance(source_spec,
source_dspec,
source_len_pad,
pic_path=os.path.join(
output_image_path,
's_indices.txt'))
self.plot_spectrograms(source_spec_rec,
os.path.join(output_image_path, 'src_rec.png'))
self.wave_generate(source_spec_rec,
os.path.join(output_audio_path, 'src_rec'))
target_spec_rec = self.inference_one_utterance(target_spec,
target_dspec,
target_len_pad,
pic_path=os.path.join(
output_image_path,
't_indices.txt'))
self.plot_spectrograms(target_spec_rec,
os.path.join(output_image_path, 'tar_rec.png'))
self.wave_generate(target_spec_rec,
os.path.join(output_audio_path, 'tar_rec'))
criterion = nn.L1Loss()
loss_src_rec = criterion(torch.from_numpy(source_spec_rec),
torch.from_numpy(source_spec_ori))
loss_trg_rec = criterion(torch.from_numpy(target_spec_rec),
torch.from_numpy(target_spec_ori))
print('Source Rec Loss: {} | Target Rec Loss: {}'.format(
loss_src_rec, loss_trg_rec))
# conversion speaker
s2t_spec = self.inference_one_utterance(source_spec,
target_dspec,
len_pad=source_len_pad)
self.plot_spectrograms(s2t_spec,
os.path.join(output_image_path, 's2t.png'))
self.wave_generate(s2t_spec, os.path.join(output_audio_path, 's2t'))
t2s_spec = self.inference_one_utterance(target_spec,
source_dspec,
len_pad=target_len_pad)
self.plot_spectrograms(t2s_spec,
os.path.join(output_image_path, 't2s.png'))
self.wave_generate(t2s_spec, os.path.join(output_audio_path, 't2s'))
print('Complete...')
return
def infer_random(self,gender=False):
#using the first sample from in_test
if gender:
source_sp = random.sample(self.speakers_man, 1)[0]
target_sp = random.sample(self.speakers_woman, 1)[0]
else:
source_sp, target_sp = random.sample(self.speakers,2)
source_utt = random.sample(self.indexes[source_sp],1)[0]
target_utt = random.sample(self.indexes[target_sp],1)[0]
if self.args.val_set == 'out_test':
source_sp = '237'
target_sp = '225'
source_utt = '007'
target_utt = '008'
# source_sp = '227'
# target_sp = '312'
# source_utt = '099'
# target_utt = '285'
display = False
output_image_path = os.path.join(self.args.output_image_path,self.args.val_set,self.args.load_iteration,
's{}_{}---t{}_{}'.format(source_sp, source_utt, target_sp, target_utt))
output_audio_path = os.path.join(self.args.output_audio_path,self.args.val_set,self.args.load_iteration,
's{}_{}---t{}_{}'.format(source_sp, source_utt, target_sp, target_utt))
os.makedirs(output_image_path,exist_ok=True)
os.makedirs(output_audio_path,exist_ok=True)
source_spec_ori = self.dataset[f'{source_sp}/{source_utt}/mel'][:]
target_spec_ori = self.dataset[f'{target_sp}/{target_utt}/mel'][:]
print('source time {} | target time {}'.format(source_spec_ori.shape, target_spec_ori.shape))
b = 32 if self.args.model_type=='AutoVC' else 8
source_spec, source_len_pad = self.pad_seq(source_spec_ori, base=b)
target_spec, target_len_pad = self.pad_seq(target_spec_ori, base=b)
source_spec = cc(torch.from_numpy(source_spec))
target_spec = cc(torch.from_numpy(target_spec))
if self.args.model_type=='AdaVAEd':
target_slices = self.dataset[f'{target_sp}/{target_utt}/dmels'][:]
source_slices = self.dataset[f'{source_sp}/{source_utt}/dmels'][:]
target_dspec = self.dataset[f'{target_sp}/{target_utt}/dmel'][:]
source_dspec = self.dataset[f'{source_sp}/{source_utt}/dmel'][:]
target_dspec = cc(torch.from_numpy(np.array([target_dspec[t0:t1] for t0, t1 in target_slices])))
source_dspec = cc(torch.from_numpy(np.array([source_dspec[t0:t1] for t0, t1 in source_slices])))
else:
source_dspec = source_spec
target_dspec = target_spec
wav2wav(self.speaker2filenames[source_sp][source_utt],os.path.join(output_audio_path,'src_ori.wav'))
wav2wav(self.speaker2filenames[target_sp][target_utt],os.path.join(output_audio_path,'tar_ori.wav'))
# sync
self.plot_spectrograms(source_spec_ori, os.path.join(output_image_path,'src_sync.png'))
self.wave_generate(source_spec_ori, os.path.join(output_audio_path,'src_sync'))
self.plot_spectrograms(target_spec_ori, os.path.join(output_image_path,'tar_sync.png'))
self.wave_generate(target_spec_ori, os.path.join(output_audio_path,'tar_sync'))
# reconstruction
source_spec_rec = self.inference_one_utterance(source_spec, source_dspec, source_len_pad)
self.plot_spectrograms(source_spec_rec, os.path.join(output_image_path,'src_rec.png'))
self.wave_generate(source_spec_rec, os.path.join(output_audio_path,'src_rec'))
target_spec_rec = self.inference_one_utterance(target_spec, target_dspec ,target_len_pad)
self.plot_spectrograms(target_spec_rec, os.path.join(output_image_path,'tar_rec.png'))
self.wave_generate(target_spec_rec, os.path.join(output_audio_path,'tar_rec'))
criterion = nn.L1Loss()
loss_src_rec = criterion(torch.from_numpy(source_spec_rec), torch.from_numpy(source_spec_ori))
loss_trg_rec = criterion(torch.from_numpy(target_spec_rec), torch.from_numpy(target_spec_ori))
print('Source Rec Loss: {} | Target Rec Loss: {}'.format(loss_src_rec, loss_trg_rec))
# conversion speaker
s2t_spec = self.inference_one_utterance(source_spec, target_dspec, len_pad=source_len_pad)
self.plot_spectrograms(s2t_spec, os.path.join(output_image_path,'s2t.png'))
self.wave_generate(s2t_spec, os.path.join(output_audio_path,'s2t'))
t2s_spec = self.inference_one_utterance(target_spec, source_dspec, len_pad=target_len_pad)
self.plot_spectrograms(t2s_spec, os.path.join(output_image_path,'t2s.png'))
self.wave_generate(t2s_spec, os.path.join(output_audio_path,'t2s'))
print('Complete...')
return