def gpu_decode(feat_list, gpu, cvlist=None, mcdpow_cvlist=None, mcdpowstd_cvlist=None, mcd_cvlist=None, \
mcdstd_cvlist=None, cvlist_dv=None, mcdpow_cvlist_dv=None, mcdpowstd_cvlist_dv=None, \
mcd_cvlist_dv=None, mcdstd_cvlist_dv=None, \
f0rmse_cvlist=None, f0corr_cvlist=None, caprmse_cvlist=None, \
f0rmse_cvlist_dv=None, f0corr_cvlist_dv=None, caprmse_cvlist_dv=None, \
cvlist_cyc=None, mcdpow_cvlist_cyc=None, mcdpowstd_cvlist_cyc=None, mcd_cvlist_cyc=None, \
mcdstd_cvlist_cyc=None, cvlist_cyc_dv=None, mcdpow_cvlist_cyc_dv=None, mcdpowstd_cvlist_cyc_dv=None, \
mcd_cvlist_cyc_dv=None, mcdstd_cvlist_cyc_dv=None, \
f0rmse_cvlist_cyc=None, f0corr_cvlist_cyc=None, caprmse_cvlist_cyc=None, \
f0rmse_cvlist_cyc_dv=None, f0corr_cvlist_cyc_dv=None, caprmse_cvlist_cyc_dv=None):
with torch.cuda.device(gpu):
# define model and load parameters
with torch.no_grad():
model_encoder_mcep = GRU_VAE_ENCODER(
in_dim=config.mcep_dim+config.excit_dim,
n_spk=n_spk,
lat_dim=config.lat_dim,
hidden_layers=config.hidden_layers_enc,
hidden_units=config.hidden_units_enc,
kernel_size=config.kernel_size_enc,
dilation_size=config.dilation_size_enc,
causal_conv=config.causal_conv_enc,
bi=config.bi_enc,
cont=False,
pad_first=True,
right_size=config.right_size,
ar=config.ar_enc)
logging.info(model_encoder_mcep)
model_decoder_mcep = GRU_SPEC_DECODER(
feat_dim=config.lat_dim,
out_dim=config.mcep_dim,
n_spk=n_spk,
hidden_layers=config.hidden_layers_dec,
hidden_units=config.hidden_units_dec,
kernel_size=config.kernel_size_dec,
dilation_size=config.dilation_size_dec,
causal_conv=config.causal_conv_dec,
bi=config.bi_dec,
spkidtr_dim=config.spkidtr_dim,
pad_first=True,
ar=config.ar_dec)
logging.info(model_decoder_mcep)
model_encoder_excit = GRU_VAE_ENCODER(
in_dim=config.mcep_dim+config.excit_dim,
n_spk=n_spk,
lat_dim=config.lat_dim_e,
hidden_layers=config.hidden_layers_enc,
hidden_units=config.hidden_units_enc,
kernel_size=config.kernel_size_enc,
dilation_size=config.dilation_size_enc,
causal_conv=config.causal_conv_enc,
bi=config.bi_enc,
cont=False,
pad_first=True,
right_size=config.right_size,
ar=config.ar_enc)
logging.info(model_encoder_excit)
model_decoder_excit = GRU_EXCIT_DECODER(
feat_dim=config.lat_dim_e,
cap_dim=config.cap_dim,
n_spk=n_spk,
hidden_layers=config.hidden_layers_lf0,
hidden_units=config.hidden_units_lf0,
kernel_size=config.kernel_size_lf0,
dilation_size=config.dilation_size_lf0,
causal_conv=config.causal_conv_lf0,
bi=config.bi_lf0,
spkidtr_dim=config.spkidtr_dim,
pad_first=True,
ar=config.ar_f0)
logging.info(model_decoder_excit)
model_vq = torch.nn.Embedding(config.ctr_size, config.lat_dim)
logging.info(model_vq)
model_encoder_mcep.load_state_dict(torch.load(args.model)["model_encoder_mcep"])
model_decoder_mcep.load_state_dict(torch.load(args.model)["model_decoder_mcep"])
model_encoder_excit.load_state_dict(torch.load(args.model)["model_encoder_excit"])
model_decoder_excit.load_state_dict(torch.load(args.model)["model_decoder_excit"])
model_vq.load_state_dict(torch.load(args.model)["model_vq"])
model_encoder_mcep.cuda()
model_decoder_mcep.cuda()
model_encoder_excit.cuda()
model_decoder_excit.cuda()
model_vq.cuda()
model_encoder_mcep.eval()
model_decoder_mcep.eval()
model_encoder_excit.eval()
model_decoder_excit.eval()
model_vq.eval()
for param in model_encoder_mcep.parameters():
param.requires_grad = False
for param in model_decoder_mcep.parameters():
param.requires_grad = False
for param in model_encoder_excit.parameters():
param.requires_grad = False
for param in model_decoder_excit.parameters():
param.requires_grad = False
for param in model_vq.parameters():
param.requires_grad = False
if config.ar_enc:
yz_in = torch.zeros((1, 1, n_spk+config.lat_dim)).cuda()
yz_in_e = torch.zeros((1, 1, n_spk+config.lat_dim_e)).cuda()
if config.ar_dec or config.ar_f0:
mean_stats = torch.FloatTensor(read_hdf5(config.stats, "/mean_"+config.string_path.replace("/","")))
scale_stats = torch.FloatTensor(read_hdf5(config.stats, "/scale_"+config.string_path.replace("/","")))
if config.ar_dec:
x_in = ((torch.zeros((1, 1, config.mcep_dim))-mean_stats[config.excit_dim:])/scale_stats[config.excit_dim:]).cuda()
if config.ar_f0:
e_in = torch.cat((torch.zeros(1,1,1), (torch.zeros(1,1,1)-mean_stats[1:2])/scale_stats[1:2], \
torch.zeros(1,1,1), (torch.zeros(1,1,config.cap_dim)-mean_stats[3:config.excit_dim])/scale_stats[3:config.excit_dim]), 2).cuda()
count = 0
pad_left = (model_encoder_mcep.pad_left + model_decoder_mcep.pad_left)*2
pad_right = (model_encoder_mcep.pad_right + model_decoder_mcep.pad_right)*2
outpad_lefts = [None]*3
outpad_rights = [None]*3
outpad_lefts[0] = pad_left-model_encoder_mcep.pad_left
outpad_rights[0] = pad_right-model_encoder_mcep.pad_right
outpad_lefts[1] = outpad_lefts[0]-model_decoder_mcep.pad_left
outpad_rights[1] = outpad_rights[0]-model_decoder_mcep.pad_right
outpad_lefts[2] = outpad_lefts[1]-model_encoder_mcep.pad_left
outpad_rights[2] = outpad_rights[1]-model_encoder_mcep.pad_right
for feat_file in feat_list:
# convert mcep
logging.info("recmcep " + feat_file)
feat_org = read_hdf5(feat_file, "/feat_mceplf0cap")
logging.info(feat_org.shape)
with torch.no_grad():
feat = F.pad(torch.FloatTensor(feat_org).cuda().unsqueeze(0).transpose(1,2), (pad_left,pad_right), "replicate").transpose(1,2)
if config.ar_enc:
spk_logits, lat_src, _, _ = model_encoder_mcep(feat, yz_in=yz_in)
spk_logits_e, lat_src_e, _, _ = model_encoder_excit(feat, yz_in=yz_in)
else:
spk_logits, lat_src, _ = model_encoder_mcep(feat)
spk_logits_e, lat_src_e, _ = model_encoder_excit(feat)
idx_vq = nn_search_batch(lat_src, model_vq.weight)
lat_src = model_vq(idx_vq)
if outpad_rights[0] > 0:
unique, counts = np.unique(idx_vq[:,outpad_lefts[0]:-outpad_rights[0]].cpu().data.numpy(), return_counts=True)
else:
unique, counts = np.unique(idx_vq[:,outpad_lefts[0]:].cpu().data.numpy(), return_counts=True)
logging.info("input vq")
logging.info(dict(zip(unique, counts)))
idx_vq_e = nn_search_batch(lat_src_e, model_vq.weight)
lat_src_e = model_vq(idx_vq_e)
if outpad_rights[0] > 0:
unique, counts = np.unique(idx_vq_e[:,outpad_lefts[0]:-outpad_rights[0]].cpu().data.numpy(), return_counts=True)
else:
unique, counts = np.unique(idx_vq_e[:,outpad_lefts[0]:].cpu().data.numpy(), return_counts=True)
logging.info("input vq_e")
logging.info(dict(zip(unique, counts)))
logging.info('input spkpost')
if outpad_rights[0] > 0:
logging.info(torch.mean(F.softmax(spk_logits[:,outpad_lefts[0]:-outpad_rights[0]], dim=-1), 1))
else:
logging.info(torch.mean(F.softmax(spk_logits[:,outpad_lefts[0]:], dim=-1), 1))
logging.info('input spkpost_e')
if outpad_rights[0] > 0:
logging.info(torch.mean(F.softmax(spk_logits_e[:,outpad_lefts[0]:-outpad_rights[0]], dim=-1), 1))
else:
logging.info(torch.mean(F.softmax(spk_logits_e[:,outpad_lefts[0]:], dim=-1), 1))
src_code = (torch.ones((1, lat_src.shape[1]))*spk_idx).cuda().long()
if config.ar_dec:
cvmcep_src, _, _ = model_decoder_mcep(src_code, lat_src, x_in=x_in)
else:
cvmcep_src, _ = model_decoder_mcep(src_code, lat_src)
if config.ar_f0:
cvlf0_src, _, _ = model_decoder_excit(src_code, lat_src_e, e_in=e_in)
else:
cvlf0_src, _ = model_decoder_excit(src_code, lat_src_e)
cv_feat = torch.cat((cvlf0_src, cvmcep_src), 2)
if config.ar_enc:
spk_logits, lat_rec, _, _ = model_encoder_mcep(cv_feat, yz_in=yz_in)
spk_logits_e, lat_rec_e, _, _ = model_encoder_excit(cv_feat, yz_in=yz_in)
else:
spk_logits, lat_rec, _ = model_encoder_mcep(cv_feat)
spk_logits_e, lat_rec_e, _ = model_encoder_excit(cv_feat)
idx_vq = nn_search_batch(lat_rec, model_vq.weight)
lat_rec = model_vq(idx_vq)
if outpad_rights[2] > 0:
unique, counts = np.unique(idx_vq[:,outpad_lefts[2]:-outpad_rights[2]].cpu().data.numpy(), return_counts=True)
else:
unique, counts = np.unique(idx_vq[:,outpad_lefts[2]:].cpu().data.numpy(), return_counts=True)
logging.info("input vq")
logging.info(dict(zip(unique, counts)))
idx_vq_e = nn_search_batch(lat_rec_e, model_vq.weight)
lat_rec_e = model_vq(idx_vq_e)
if outpad_rights[2] > 0:
unique, counts = np.unique(idx_vq_e[:,outpad_lefts[2]:-outpad_rights[2]].cpu().data.numpy(), return_counts=True)
else:
unique, counts = np.unique(idx_vq_e[:,outpad_lefts[2]:].cpu().data.numpy(), return_counts=True)
logging.info("input vq_e")
logging.info(dict(zip(unique, counts)))
logging.info('rec spkpost')
if outpad_rights[2] > 0:
logging.info(torch.mean(F.softmax(spk_logits[:,outpad_lefts[2]:-outpad_rights[2]], dim=-1), 1))
else:
logging.info(torch.mean(F.softmax(spk_logits[:,outpad_lefts[2]:], dim=-1), 1))
logging.info('rec spkpost_e')
if outpad_rights[2] > 0:
logging.info(torch.mean(F.softmax(spk_logits_e[:,outpad_lefts[2]:-outpad_rights[2]], dim=-1), 1))
else:
logging.info(torch.mean(F.softmax(spk_logits_e[:,outpad_lefts[2]:], dim=-1), 1))
src_code = (torch.ones((1, lat_rec.shape[1]))*spk_idx).cuda().long()
if config.ar_dec:
cvmcep_cyc, _, _ = model_decoder_mcep(src_code, lat_rec, x_in=x_in)
else:
cvmcep_cyc, _ = model_decoder_mcep(src_code, lat_rec)
if config.ar_f0:
cvlf0_cyc, _, _ = model_decoder_excit(src_code, lat_rec_e, e_in=e_in)
else:
cvlf0_cyc, _ = model_decoder_excit(src_code, lat_rec_e)
if outpad_rights[1] > 0:
cvmcep_src = cvmcep_src[:,outpad_lefts[1]:-outpad_rights[1]]
cvlf0_src = cvlf0_src[:,outpad_lefts[1]:-outpad_rights[1]]
else:
cvmcep_src = cvmcep_src[:,outpad_lefts[1]:]
cvlf0_src = cvlf0_src[:,outpad_lefts[1]:]
feat_rec = torch.cat((torch.round(cvlf0_src[:,:,:1]), cvlf0_src[:,:,1:2], \
torch.round(cvlf0_src[:,:,2:3]), cvlf0_src[:,:,3:], cvmcep_src), \
2)[0].cpu().data.numpy()
feat_cyc = torch.cat((torch.round(cvlf0_cyc[:,:,:1]), cvlf0_cyc[:,:,1:2], \
torch.round(cvlf0_cyc[:,:,2:3]), cvlf0_cyc[:,:,3:], cvmcep_cyc), \
2)[0].cpu().data.numpy()
cvmcep_src = np.array(cvmcep_src[0].cpu().data.numpy(), dtype=np.float64)
cvlf0_src = np.array(cvlf0_src[0].cpu().data.numpy(), dtype=np.float64)
cvmcep_cyc = np.array(cvmcep_cyc[0].cpu().data.numpy(), dtype=np.float64)
cvlf0_cyc = np.array(cvlf0_cyc[0].cpu().data.numpy(), dtype=np.float64)
logging.info(cvlf0_src.shape)
logging.info(cvmcep_src.shape)
logging.info(cvlf0_cyc.shape)
logging.info(cvmcep_cyc.shape)
mcep = np.array(feat_org[:,-model_decoder_mcep.out_dim:])
f0 = np.array(np.rint(feat_org[:,0])*np.exp(feat_org[:,1]))
codeap = np.array(np.rint(feat_org[:,2:3])*(-np.exp(feat_org[:,3:feat_org.shape[-1]-model_decoder_mcep.out_dim])))
cvf0_src = np.array(np.rint(cvlf0_src[:,0])*np.exp(cvlf0_src[:,1]))
cvcodeap_src = np.array(np.rint(cvlf0_src[:,2:3])*(-np.exp(cvlf0_src[:,3:])))
f0_rmse = np.sqrt(np.mean((cvf0_src-f0)**2))
logging.info('F0_rmse_rec: %lf Hz' % (f0_rmse))
cvf0_src_mean = np.mean(cvf0_src)
f0_mean = np.mean(f0)
f0_corr = np.sum((cvf0_src-cvf0_src_mean)*(f0-f0_mean))/\
(np.sqrt(np.sum((cvf0_src-cvf0_src_mean)**2))*np.sqrt(np.sum((f0-f0_mean)**2)))
logging.info('F0_corr_rec: %lf' % (f0_corr))
codeap_rmse = np.sqrt(np.mean((cvcodeap_src-codeap)**2, axis=0))
for i in range(codeap_rmse.shape[-1]):
logging.info('codeap-%d_rmse_rec: %lf dB' % (i+1, codeap_rmse[i]))
cvf0_cyc = np.array(np.rint(cvlf0_cyc[:,0])*np.exp(cvlf0_cyc[:,1]))
cvcodeap_cyc = np.array(np.rint(cvlf0_cyc[:,2:3])*(-np.exp(cvlf0_cyc[:,3:])))
f0_rmse_cyc = np.sqrt(np.mean((cvf0_cyc-f0)**2))
logging.info('F0_rmse_cyc: %lf Hz' % (f0_rmse_cyc))
cvf0_cyc_mean = np.mean(cvf0_cyc)
f0_mean = np.mean(f0)
f0_corr_cyc = np.sum((cvf0_cyc-cvf0_cyc_mean)*(f0-f0_mean))/\
(np.sqrt(np.sum((cvf0_cyc-cvf0_cyc_mean)**2))*np.sqrt(np.sum((f0-f0_mean)**2)))
logging.info('F0_corr_cyc: %lf' % (f0_corr_cyc))
codeap_rmse_cyc = np.sqrt(np.mean((cvcodeap_cyc-codeap)**2, axis=0))
for i in range(codeap_rmse_cyc.shape[-1]):
logging.info('codeap-%d_rmse_cyc: %lf dB' % (i+1, codeap_rmse_cyc[i]))
spcidx = read_hdf5(feat_file, "/spcidx_range")[0]
_, _, _, mcdpow_arr = dtw.dtw_org_to_trg(np.array(cvmcep_src[np.array(spcidx),:], \
dtype=np.float64), np.array(mcep[np.array(spcidx),:], dtype=np.float64))
_, _, _, mcd_arr = dtw.dtw_org_to_trg(np.array(cvmcep_src[np.array(spcidx),1:], \
dtype=np.float64), np.array(mcep[np.array(spcidx),1:], dtype=np.float64))
mcdpow_mean = np.mean(mcdpow_arr)
mcdpow_std = np.std(mcdpow_arr)
mcd_mean = np.mean(mcd_arr)
mcd_std = np.std(mcd_arr)
logging.info("mcdpow_rec: %.6f dB +- %.6f" % (mcdpow_mean, mcdpow_std))
logging.info("mcd_rec: %.6f dB +- %.6f" % (mcd_mean, mcd_std))
_, _, _, mcdpow_arr = dtw.dtw_org_to_trg(np.array(cvmcep_cyc[np.array(spcidx),:], \
dtype=np.float64), np.array(mcep[np.array(spcidx),:], dtype=np.float64))
_, _, _, mcd_arr = dtw.dtw_org_to_trg(np.array(cvmcep_cyc[np.array(spcidx),1:], \
dtype=np.float64), np.array(mcep[np.array(spcidx),1:], dtype=np.float64))
mcdpow_mean_cyc = np.mean(mcdpow_arr)
mcdpow_std_cyc = np.std(mcdpow_arr)
mcd_mean_cyc = np.mean(mcd_arr)
mcd_std_cyc = np.std(mcd_arr)
logging.info("mcdpow_cyc: %.6f dB +- %.6f" % (mcdpow_mean_cyc, mcdpow_std_cyc))
logging.info("mcd_cyc: %.6f dB +- %.6f" % (mcd_mean_cyc, mcd_std_cyc))
logging.info('org f0')
logging.info(f0[10:15])
logging.info('rec f0')
logging.info(cvf0_src[10:15])
logging.info('cyc f0')
logging.info(cvf0_cyc[10:15])
logging.info('org cap')
logging.info(codeap[10:15])
logging.info('rec cap')
logging.info(cvcodeap_src[10:15])
logging.info('cyc cap')
logging.info(cvcodeap_cyc[10:15])
dataset = feat_file.split('/')[1].split('_')[0]
if 'tr' in dataset:
logging.info('trn')
f0rmse_cvlist.append(f0_rmse)
f0corr_cvlist.append(f0_corr)
caprmse_cvlist.append(codeap_rmse)
mcdpow_cvlist.append(mcdpow_mean)
mcdpow_cvlist.append(mcdpow_mean)
mcdpowstd_cvlist.append(mcdpow_std)
mcd_cvlist.append(mcd_mean)
mcdstd_cvlist.append(mcd_std)
cvlist.append(np.var(cvmcep_src[:,1:], axis=0))
logging.info(len(cvlist))
f0rmse_cvlist_cyc.append(f0_rmse_cyc)
f0corr_cvlist_cyc.append(f0_corr_cyc)
caprmse_cvlist_cyc.append(codeap_rmse_cyc)
mcdpow_cvlist_cyc.append(mcdpow_mean_cyc)
mcdpow_cvlist_cyc.append(mcdpow_mean_cyc)
mcdpowstd_cvlist_cyc.append(mcdpow_std_cyc)
mcd_cvlist_cyc.append(mcd_mean_cyc)
mcdstd_cvlist_cyc.append(mcd_std_cyc)
cvlist_cyc.append(np.var(cvmcep_cyc[:,1:], axis=0))
elif 'dv' in dataset:
logging.info('dev')
f0rmse_cvlist_dv.append(f0_rmse)
f0corr_cvlist_dv.append(f0_corr)
caprmse_cvlist_dv.append(codeap_rmse)
mcdpow_cvlist_dv.append(mcdpow_mean)
mcdpowstd_cvlist_dv.append(mcdpow_std)
mcd_cvlist_dv.append(mcd_mean)
mcdstd_cvlist_dv.append(mcd_std)
cvlist_dv.append(np.var(cvmcep_src[:,1:], axis=0))
logging.info(len(cvlist_dv))
f0rmse_cvlist_cyc_dv.append(f0_rmse_cyc)
f0corr_cvlist_cyc_dv.append(f0_corr_cyc)
caprmse_cvlist_cyc_dv.append(codeap_rmse_cyc)
mcdpow_cvlist_cyc_dv.append(mcdpow_mean_cyc)
mcdpow_cvlist_cyc_dv.append(mcdpow_mean_cyc)
mcdpowstd_cvlist_cyc_dv.append(mcdpow_std_cyc)
mcd_cvlist_cyc_dv.append(mcd_mean_cyc)
mcdstd_cvlist_cyc_dv.append(mcd_std_cyc)
cvlist_cyc_dv.append(np.var(cvmcep_cyc[:,1:], axis=0))
logging.info('write rec to h5')
outh5dir = os.path.join(os.path.dirname(os.path.dirname(feat_file)), args.spk+"-"+args.spk)
if not os.path.exists(outh5dir):
os.makedirs(outh5dir)
feat_file = os.path.join(outh5dir, os.path.basename(feat_file))
logging.info(feat_file + ' ' + args.string_path)
logging.info(feat_rec.shape)
write_hdf5(feat_file, args.string_path, feat_rec)
logging.info('write cyc to h5')
outh5dir = os.path.join(os.path.dirname(os.path.dirname(feat_file)), args.spk+"-"+args.spk+"-"+args.spk)
if not os.path.exists(outh5dir):
os.makedirs(outh5dir)
feat_file = os.path.join(outh5dir, os.path.basename(feat_file))
logging.info(feat_file + ' ' + args.string_path)
logging.info(feat_cyc.shape)
write_hdf5(feat_file, args.string_path, feat_cyc)
count += 1
def gpu_decode(feat_list, gpu, cvlist=None, mcdpow_cvlist=None, mcdpowstd_cvlist=None, mcd_cvlist=None, \
mcdstd_cvlist=None, cvlist_dv=None, mcdpow_cvlist_dv=None, mcdpowstd_cvlist_dv=None, \
mcd_cvlist_dv=None, mcdstd_cvlist_dv=None, \
cvlist_cyc=None, mcdpow_cvlist_cyc=None, mcdpowstd_cvlist_cyc=None, mcd_cvlist_cyc=None, \
mcdstd_cvlist_cyc=None, cvlist_cyc_dv=None, mcdpow_cvlist_cyc_dv=None, mcdpowstd_cvlist_cyc_dv=None, \
mcd_cvlist_cyc_dv=None, mcdstd_cvlist_cyc_dv=None):
with torch.cuda.device(gpu):
# define model and load parameters
with torch.no_grad():
model_encoder = GRU_VAE_ENCODER(
in_dim=config.mcep_dim + config.excit_dim,
n_spk=n_spk,
lat_dim=config.lat_dim,
hidden_layers=config.hidden_layers_enc,
hidden_units=config.hidden_units_enc,
kernel_size=config.kernel_size_enc,
dilation_size=config.dilation_size_enc,
causal_conv=config.causal_conv_enc,
bi=config.bi_enc,
cont=False,
pad_first=True,
right_size=config.right_size,
ar=config.ar_enc)
logging.info(model_encoder)
model_decoder = GRU_SPEC_DECODER(
feat_dim=config.lat_dim,
out_dim=config.mcep_dim,
n_spk=n_spk,
hidden_layers=config.hidden_layers_dec,
hidden_units=config.hidden_units_dec,
kernel_size=config.kernel_size_dec,
dilation_size=config.dilation_size_dec,
causal_conv=config.causal_conv_dec,
bi=config.bi_dec,
pad_first=True,
ar=config.ar_dec)
logging.info(model_decoder)
model_vq = torch.nn.Embedding(config.ctr_size, config.lat_dim)
logging.info(model_vq)
model_encoder.load_state_dict(
torch.load(args.model)["model_encoder"])
model_decoder.load_state_dict(
torch.load(args.model)["model_decoder"])
model_vq.load_state_dict(torch.load(args.model)["model_vq"])
model_encoder.cuda()
model_decoder.cuda()
model_vq.cuda()
model_encoder.eval()
model_decoder.eval()
model_vq.eval()
for param in model_encoder.parameters():
param.requires_grad = False
for param in model_decoder.parameters():
param.requires_grad = False
for param in model_vq.parameters():
param.requires_grad = False
if config.ar_enc:
yz_in = torch.zeros((1, 1, n_spk + config.lat_dim)).cuda()
if config.ar_dec:
mean_stats = torch.FloatTensor(
read_hdf5(
config.stats,
"/mean_" + config.string_path.replace("/", "")))
scale_stats = torch.FloatTensor(
read_hdf5(
config.stats,
"/scale_" + config.string_path.replace("/", "")))
x_in = ((torch.zeros((1, 1, config.mcep_dim)) -
mean_stats[config.excit_dim:]) /
scale_stats[config.excit_dim:]).cuda()
count = 0
pad_left = (model_encoder.pad_left + model_decoder.pad_left) * 2
pad_right = (model_encoder.pad_right + model_decoder.pad_right) * 2
outpad_lefts = [None] * 3
outpad_rights = [None] * 3
outpad_lefts[0] = pad_left - model_encoder.pad_left
outpad_rights[0] = pad_right - model_encoder.pad_right
outpad_lefts[1] = outpad_lefts[0] - model_decoder.pad_left
outpad_rights[1] = outpad_rights[0] - model_decoder.pad_right
outpad_lefts[2] = outpad_lefts[1] - model_encoder.pad_left
outpad_rights[2] = outpad_rights[1] - model_encoder.pad_right
for feat_file in feat_list:
# convert mcep
logging.info("recmcep " + feat_file)
feat_org = read_hdf5(feat_file, "/feat_mceplf0cap")
logging.info(feat_org.shape)
mcep = np.array(feat_org[:, -model_decoder.out_dim:])
with torch.no_grad():
feat = torch.FloatTensor(feat_org).cuda().unsqueeze(0)
feat_excit = feat[:, :, :config.excit_dim]
if config.ar_enc:
spk_logits, lat_src, _, _ = model_encoder(F.pad(feat.transpose(1,2), (pad_left,pad_right), "replicate").transpose(1,2), \
yz_in=yz_in)
else:
spk_logits, lat_src, _ = model_encoder(
F.pad(feat.transpose(1, 2), (pad_left, pad_right),
"replicate").transpose(1, 2))
idx_vq = nn_search_batch(lat_src, model_vq.weight)
lat_src = model_vq(idx_vq)
if outpad_rights[0] > 0:
unique, counts = np.unique(
idx_vq[:, outpad_lefts[0]:-outpad_rights[0]].cpu(
).data.numpy(),
return_counts=True)
else:
unique, counts = np.unique(
idx_vq[:, outpad_lefts[0]:].cpu().data.numpy(),
return_counts=True)
logging.info("input vq")
logging.info(dict(zip(unique, counts)))
logging.info('input spkpost')
if outpad_rights[0] > 0:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[0]:
-outpad_rights[0]],
dim=-1), 1))
else:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[0]:],
dim=-1), 1))
src_code = (torch.ones(
(1, lat_src.shape[1])) * spk_idx).cuda().long()
if config.ar_dec:
cvmcep_src, _, _ = model_decoder(src_code,
lat_src,
x_in=x_in)
else:
cvmcep_src, _ = model_decoder(src_code, lat_src)
if config.ar_enc:
spk_logits, lat_rec, _, _ = model_encoder(torch.cat((F.pad(feat_excit.transpose(1,2), \
(outpad_lefts[1],outpad_rights[1]), "replicate").transpose(1,2), cvmcep_src), 2),
yz_in=yz_in)
else:
spk_logits, lat_rec, _ = model_encoder(torch.cat((F.pad(feat_excit.transpose(1,2), \
(outpad_lefts[1],outpad_rights[1]), "replicate").transpose(1,2), cvmcep_src), 2))
idx_vq = nn_search_batch(lat_rec, model_vq.weight)
lat_rec = model_vq(idx_vq)
if outpad_rights[2] > 0:
unique, counts = np.unique(
idx_vq[:, outpad_lefts[2]:-outpad_rights[2]].cpu(
).data.numpy(),
return_counts=True)
else:
unique, counts = np.unique(
idx_vq[:, outpad_lefts[2]:].cpu().data.numpy(),
return_counts=True)
logging.info("rec vq")
logging.info(dict(zip(unique, counts)))
logging.info('rec spkpost')
if outpad_rights[2] > 0:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[2]:
-outpad_rights[2]],
dim=-1), 1))
else:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[2]:],
dim=-1), 1))
src_code = (torch.ones(
(1, lat_rec.shape[1])) * spk_idx).cuda().long()
if config.ar_dec:
cvmcep_cyc, _, _ = model_decoder(src_code,
lat_rec,
x_in=x_in)
else:
cvmcep_cyc, _ = model_decoder(src_code, lat_rec)
if outpad_rights[1] > 0:
cvmcep_src = cvmcep_src[:, outpad_lefts[1]:
-outpad_rights[1]]
else:
cvmcep_src = cvmcep_src[:, outpad_lefts[1]:]
feat_rec = torch.cat((feat_excit, cvmcep_src),
2)[0].cpu().data.numpy()
feat_cyc = torch.cat((feat_excit, cvmcep_cyc),
2)[0].cpu().data.numpy()
cvmcep_src = np.array(cvmcep_src[0].cpu().data.numpy(),
dtype=np.float64)
cvmcep_cyc = np.array(cvmcep_cyc[0].cpu().data.numpy(),
dtype=np.float64)
logging.info(cvmcep_src.shape)
logging.info(cvmcep_cyc.shape)
spcidx = read_hdf5(feat_file, "/spcidx_range")[0]
_, _, _, mcdpow_arr = dtw.dtw_org_to_trg(np.array(cvmcep_src[np.array(spcidx),:], \
dtype=np.float64), np.array(mcep[np.array(spcidx),:], dtype=np.float64))
_, _, _, mcd_arr = dtw.dtw_org_to_trg(np.array(cvmcep_src[np.array(spcidx),1:], \
dtype=np.float64), np.array(mcep[np.array(spcidx),1:], dtype=np.float64))
mcdpow_mean = np.mean(mcdpow_arr)
mcdpow_std = np.std(mcdpow_arr)
mcd_mean = np.mean(mcd_arr)
mcd_std = np.std(mcd_arr)
logging.info("mcdpow_rec: %.6f dB +- %.6f" %
(mcdpow_mean, mcdpow_std))
logging.info("mcd_rec: %.6f dB +- %.6f" % (mcd_mean, mcd_std))
_, _, _, mcdpow_arr = dtw.dtw_org_to_trg(np.array(cvmcep_cyc[np.array(spcidx),:], \
dtype=np.float64), np.array(mcep[np.array(spcidx),:], dtype=np.float64))
_, _, _, mcd_arr = dtw.dtw_org_to_trg(np.array(cvmcep_cyc[np.array(spcidx),1:], \
dtype=np.float64), np.array(mcep[np.array(spcidx),1:], dtype=np.float64))
mcdpow_mean_cyc = np.mean(mcdpow_arr)
mcdpow_std_cyc = np.std(mcdpow_arr)
mcd_mean_cyc = np.mean(mcd_arr)
mcd_std_cyc = np.std(mcd_arr)
logging.info("mcdpow_cyc: %.6f dB +- %.6f" %
(mcdpow_mean_cyc, mcdpow_std_cyc))
logging.info("mcd_cyc: %.6f dB +- %.6f" %
(mcd_mean_cyc, mcd_std_cyc))
dataset = feat_file.split('/')[1].split('_')[0]
if 'tr' in dataset:
logging.info('trn')
mcdpow_cvlist.append(mcdpow_mean)
mcdpow_cvlist.append(mcdpow_mean)
mcdpowstd_cvlist.append(mcdpow_std)
mcd_cvlist.append(mcd_mean)
mcdstd_cvlist.append(mcd_std)
cvlist.append(np.var(cvmcep_src[:, 1:], axis=0))
logging.info(len(cvlist))
mcdpow_cvlist_cyc.append(mcdpow_mean_cyc)
mcdpow_cvlist_cyc.append(mcdpow_mean_cyc)
mcdpowstd_cvlist_cyc.append(mcdpow_std_cyc)
mcd_cvlist_cyc.append(mcd_mean_cyc)
mcdstd_cvlist_cyc.append(mcd_std_cyc)
cvlist_cyc.append(np.var(cvmcep_cyc[:, 1:], axis=0))
elif 'dv' in dataset:
logging.info('dev')
mcdpow_cvlist_dv.append(mcdpow_mean)
mcdpowstd_cvlist_dv.append(mcdpow_std)
mcd_cvlist_dv.append(mcd_mean)
mcdstd_cvlist_dv.append(mcd_std)
cvlist_dv.append(np.var(cvmcep_src[:, 1:], axis=0))
logging.info(len(cvlist_dv))
mcdpow_cvlist_cyc_dv.append(mcdpow_mean_cyc)
mcdpow_cvlist_cyc_dv.append(mcdpow_mean_cyc)
mcdpowstd_cvlist_cyc_dv.append(mcdpow_std_cyc)
mcd_cvlist_cyc_dv.append(mcd_mean_cyc)
mcdstd_cvlist_cyc_dv.append(mcd_std_cyc)
cvlist_cyc_dv.append(np.var(cvmcep_cyc[:, 1:], axis=0))
logging.info('write rec to h5')
outh5dir = os.path.join(
os.path.dirname(os.path.dirname(feat_file)),
args.spk + "-" + args.spk)
if not os.path.exists(outh5dir):
os.makedirs(outh5dir)
feat_file = os.path.join(outh5dir, os.path.basename(feat_file))
logging.info(feat_file + ' ' + args.string_path)
logging.info(feat_rec.shape)
write_hdf5(feat_file, args.string_path, feat_rec)
logging.info('write cyc to h5')
outh5dir = os.path.join(
os.path.dirname(os.path.dirname(feat_file)),
args.spk + "-" + args.spk + "-" + args.spk)
if not os.path.exists(outh5dir):
os.makedirs(outh5dir)
feat_file = os.path.join(outh5dir, os.path.basename(feat_file))
logging.info(feat_file + ' ' + args.string_path)
logging.info(feat_cyc.shape)
write_hdf5(feat_file, args.string_path, feat_cyc)
count += 1
def decode_RNN(feat_list, gpu, cvlist=None,
mcd_cvlist_src=None, mcdstd_cvlist_src=None, mcdpow_cvlist_src=None, mcdpowstd_cvlist_src=None,\
mcd_cvlist_cyc=None, mcdstd_cvlist_cyc=None, mcdpow_cvlist_cyc=None, mcdpowstd_cvlist_cyc=None,\
mcd_cvlist=None, mcdstd_cvlist=None, mcdpow_cvlist=None, mcdpowstd_cvlist=None, \
lat_dist_rmse_list=None, lat_dist_cosim_list=None):
with torch.cuda.device(gpu):
# define model and load parameters
with torch.no_grad():
model_encoder = GRU_VAE_ENCODER(
in_dim=config.mcep_dim+config.excit_dim,
n_spk=n_spk,
lat_dim=config.lat_dim,
hidden_layers=config.hidden_layers_enc,
hidden_units=config.hidden_units_enc,
kernel_size=config.kernel_size_enc,
dilation_size=config.dilation_size_enc,
causal_conv=config.causal_conv_enc,
bi=False,
ar=False,
pad_first=True,
right_size=config.right_size_enc)
logging.info(model_encoder)
model_decoder = GRU_SPEC_DECODER(
feat_dim=config.lat_dim,
out_dim=config.mcep_dim,
n_spk=n_spk,
hidden_layers=config.hidden_layers_dec,
hidden_units=config.hidden_units_dec,
kernel_size=config.kernel_size_dec,
dilation_size=config.dilation_size_dec,
causal_conv=config.causal_conv_dec,
bi=False,
ar=False,
pad_first=True,
right_size=config.right_size_dec)
logging.info(model_decoder)
model_post = GRU_POST_NET(
spec_dim=config.mcep_dim,
excit_dim=2,
n_spk=n_spk,
hidden_layers=config.hidden_layers_post,
hidden_units=config.hidden_units_post,
kernel_size=config.kernel_size_post,
dilation_size=config.dilation_size_post,
causal_conv=config.causal_conv_post,
pad_first=True,
right_size=config.right_size_post)
#excit_dim=config.excit_dim,
#excit_dim=None,
logging.info(model_post)
model_encoder.load_state_dict(torch.load(args.model)["model_encoder"])
model_decoder.load_state_dict(torch.load(args.model)["model_decoder"])
model_post.load_state_dict(torch.load(args.model)["model_post"])
model_encoder.remove_weight_norm()
model_decoder.remove_weight_norm()
model_post.remove_weight_norm()
model_encoder.cuda()
model_decoder.cuda()
model_post.cuda()
model_encoder.eval()
model_decoder.eval()
model_post.eval()
for param in model_encoder.parameters():
param.requires_grad = False
for param in model_decoder.parameters():
param.requires_grad = False
for param in model_post.parameters():
param.requires_grad = False
count = 0
pad_left = (model_encoder.pad_left + model_decoder.pad_left + model_post.pad_left)*2
pad_right = (model_encoder.pad_right + model_decoder.pad_right + model_post.pad_right)*2
outpad_lefts = [None]*5
outpad_rights = [None]*5
outpad_lefts[0] = pad_left-model_encoder.pad_left
outpad_rights[0] = pad_right-model_encoder.pad_right
outpad_lefts[1] = outpad_lefts[0]-model_decoder.pad_left
outpad_rights[1] = outpad_rights[0]-model_decoder.pad_right
outpad_lefts[2] = outpad_lefts[1]-model_post.pad_left
outpad_rights[2] = outpad_rights[1]-model_post.pad_right
outpad_lefts[3] = outpad_lefts[2]-model_encoder.pad_left
outpad_rights[3] = outpad_rights[2]-model_encoder.pad_right
outpad_lefts[4] = outpad_lefts[3]-model_decoder.pad_left
outpad_rights[4] = outpad_rights[3]-model_decoder.pad_right
logging.info(f'{pad_left} {pad_right}')
logging.info(outpad_lefts)
logging.info(outpad_rights)
for feat_file in feat_list:
# convert mcep
spk_src = os.path.basename(os.path.dirname(feat_file))
src_idx = spk_list.index(spk_src)
logging.info('%s --> %s' % (spk_src, args.spk_trg))
file_trg = os.path.join(os.path.dirname(os.path.dirname(feat_file)), args.spk_trg, os.path.basename(feat_file))
trg_exist = False
if os.path.exists(file_trg):
logging.info('exist: %s' % (file_trg))
feat_trg = read_hdf5(file_trg, config.string_path)
mcep_trg = feat_trg[:,-config.mcep_dim:]
logging.info(mcep_trg.shape)
trg_exist = True
feat_org = read_hdf5(feat_file, config.string_path)
mcep = np.array(feat_org[:,-config.mcep_dim:])
codeap = np.array(np.rint(feat_org[:,2:3])*(-np.exp(feat_org[:,3:config.excit_dim])))
sp = np.array(ps.mc2sp(mcep, args.mcep_alpha, args.fftl))
ap = pw.decode_aperiodicity(codeap, args.fs, args.fftl)
feat_cvf0_lin = np.expand_dims(convert_f0(np.exp(feat_org[:,1]), src_f0_mean, src_f0_std, trg_f0_mean, trg_f0_std), axis=-1)
feat_cv = np.c_[feat_org[:,:1], np.log(feat_cvf0_lin), feat_org[:,2:config.excit_dim]]
logging.info("generate")
with torch.no_grad():
feat = F.pad(torch.FloatTensor(feat_org).cuda().unsqueeze(0).transpose(1,2), (pad_left,pad_right), "replicate").transpose(1,2)
feat_excit = torch.FloatTensor(feat_org[:,:config.excit_dim]).cuda().unsqueeze(0)
feat_excit_cv = torch.FloatTensor(feat_cv).cuda().unsqueeze(0)
spk_logits, _, lat_src, _ = model_encoder(feat, sampling=False)
logging.info('input spkpost')
if outpad_rights[0] > 0:
logging.info(torch.mean(F.softmax(spk_logits[:,outpad_lefts[0]:-outpad_rights[0]], dim=-1), 1))
else:
logging.info(torch.mean(F.softmax(spk_logits[:,outpad_lefts[0]:], dim=-1), 1))
if trg_exist:
spk_trg_logits, _, lat_trg, _ = model_encoder(F.pad(torch.FloatTensor(feat_trg).cuda().unsqueeze(0).transpose(1,2), \
(model_encoder.pad_left,model_encoder.pad_right), "replicate").transpose(1,2), sampling=False)
logging.info('target spkpost')
logging.info(torch.mean(F.softmax(spk_trg_logits, dim=-1), 1))
cvmcep_src, _ = model_decoder((torch.ones((1, lat_src.shape[1]))*src_idx).cuda().long(), lat_src)
cvmcep_src_post, _ = model_post(cvmcep_src, y=(torch.ones((1, cvmcep_src.shape[1]))*src_idx).cuda().long(),
e=F.pad(feat_excit[:,:,:2].transpose(1,2), (outpad_lefts[1],outpad_rights[1]), "replicate").transpose(1,2))
#e=F.pad(feat_excit.transpose(1,2), (outpad_lefts[1],outpad_rights[1]), "replicate").transpose(1,2))
if model_post.pad_right > 0:
spk_logits, _, lat_rec, _ = model_encoder(torch.cat((F.pad(feat_excit.transpose(1,2), \
(outpad_lefts[2],outpad_rights[2]), "replicate").transpose(1,2), cvmcep_src[:,model_post.pad_left:-model_post.pad_right]), 2),
sampling=False)
else:
spk_logits, _, lat_rec, _ = model_encoder(torch.cat((F.pad(feat_excit.transpose(1,2), \
(outpad_lefts[2],outpad_rights[2]), "replicate").transpose(1,2), cvmcep_src[:,model_post.pad_left:]), 2),
sampling=False)
logging.info('rec spkpost')
if outpad_rights[3] > 0:
logging.info(torch.mean(F.softmax(spk_logits[:,outpad_lefts[3]:-outpad_rights[3]], dim=-1), 1))
else:
logging.info(torch.mean(F.softmax(spk_logits[:,outpad_lefts[3]:], dim=-1), 1))
cvmcep, _ = model_decoder((torch.ones((1, lat_src.shape[1]))*trg_idx).cuda().long(), lat_src)
cvmcep_post, _ = model_post(cvmcep, y=(torch.ones((1, cvmcep.shape[1]))*trg_idx).cuda().long(),
e=F.pad(feat_excit_cv[:,:,:2].transpose(1,2), (outpad_lefts[1],outpad_rights[1]), "replicate").transpose(1,2))
#e=F.pad(feat_excit_cv.transpose(1,2), (outpad_lefts[1],outpad_rights[1]), "replicate").transpose(1,2))
if model_post.pad_right > 0:
spk_logits, _, lat_cv, _ = model_encoder(torch.cat((F.pad(feat_excit_cv.transpose(1,2), \
(outpad_lefts[2],outpad_rights[2]), "replicate").transpose(1,2), cvmcep[:,model_post.pad_left:-model_post.pad_right]), 2),
sampling=False)
else:
spk_logits, _, lat_cv, _ = model_encoder(torch.cat((F.pad(feat_excit_cv.transpose(1,2), \
(outpad_lefts[2],outpad_rights[2]), "replicate").transpose(1,2), cvmcep[:,model_post.pad_left:]), 2),
sampling=False)
logging.info('cv spkpost')
if outpad_rights[3] > 0:
logging.info(torch.mean(F.softmax(spk_logits[:,outpad_lefts[3]:-outpad_rights[3]], dim=-1), 1))
else:
logging.info(torch.mean(F.softmax(spk_logits[:,outpad_lefts[3]:], dim=-1), 1))
cvmcep_cyc, _ = model_decoder((torch.ones((1, lat_cv.shape[1]))*src_idx).cuda().long(), lat_cv)
cvmcep_cyc_post, _ = model_post(cvmcep_cyc, y=(torch.ones((1, cvmcep_cyc.shape[1]))*src_idx).cuda().long(),
e=F.pad(feat_excit[:,:,:2].transpose(1,2), (outpad_lefts[4],outpad_rights[4]), "replicate").transpose(1,2))
#e=F.pad(feat_excit.transpose(1,2), (outpad_lefts[4],outpad_rights[4]), "replicate").transpose(1,2))
if outpad_rights[2] > 0:
cvmcep_src = np.array(cvmcep_src_post[0,outpad_lefts[2]:-outpad_rights[2]].cpu().data.numpy(), dtype=np.float64)
cvmcep = np.array(cvmcep_post[0,outpad_lefts[2]:-outpad_rights[2]].cpu().data.numpy(), dtype=np.float64)
else:
cvmcep_src = np.array(cvmcep_src_post[0,outpad_lefts[2]:].cpu().data.numpy(), dtype=np.float64)
cvmcep = np.array(cvmcep_post[0,outpad_lefts[2]:].cpu().data.numpy(), dtype=np.float64)
cvmcep_cyc = np.array(cvmcep_cyc_post[0].cpu().data.numpy(), dtype=np.float64)
if trg_exist:
if outpad_rights[0] > 0:
lat_src = lat_src[:,outpad_lefts[0]:-outpad_rights[0]]
else:
lat_src = lat_src[:,outpad_lefts[0]:]
logging.info(cvmcep_src.shape)
logging.info(cvmcep.shape)
logging.info(cvmcep_cyc.shape)
if trg_exist:
logging.info(lat_src.shape)
logging.info(lat_trg.shape)
cvlist.append(np.var(cvmcep[:,1:], axis=0))
logging.info("cvf0lin")
f0_range = read_hdf5(feat_file, "/f0_range")
cvf0_range_lin = convert_f0(f0_range, src_f0_mean, src_f0_std, trg_f0_mean, trg_f0_std)
uv_range_lin, cont_f0_range_lin = convert_continuos_f0(np.array(cvf0_range_lin))
unique, counts = np.unique(uv_range_lin, return_counts=True)
logging.info(dict(zip(unique, counts)))
cont_f0_lpf_range_lin = \
low_pass_filter(cont_f0_range_lin, int(1.0 / (args.shiftms * 0.001)), cutoff=20)
uv_range_lin = np.expand_dims(uv_range_lin, axis=-1)
cont_f0_lpf_range_lin = np.expand_dims(cont_f0_lpf_range_lin, axis=-1)
# plain converted feat for neural vocoder
feat_cv = np.c_[uv_range_lin, np.log(cont_f0_lpf_range_lin), feat_cv[:,2:config.excit_dim], cvmcep]
logging.info(feat_cv.shape)
logging.info("mcd acc")
spcidx = np.array(read_hdf5(feat_file, "/spcidx_range")[0])
_, mcdpow_arr = dtw.calc_mcd(np.array(mcep[spcidx], dtype=np.float64), np.array(cvmcep_src[spcidx], dtype=np.float64))
_, mcd_arr = dtw.calc_mcd(np.array(mcep[spcidx,1:], dtype=np.float64), np.array(cvmcep_src[spcidx,1:], dtype=np.float64))
mcdpow_mean = np.mean(mcdpow_arr)
mcdpow_std = np.std(mcdpow_arr)
mcd_mean = np.mean(mcd_arr)
mcd_std = np.std(mcd_arr)
logging.info("mcdpow_src_cv: %.6f dB +- %.6f" % (mcdpow_mean, mcdpow_std))
logging.info("mcd_src_cv: %.6f dB +- %.6f" % (mcd_mean, mcd_std))
mcdpow_cvlist_src.append(mcdpow_mean)
mcdpowstd_cvlist_src.append(mcdpow_std)
mcd_cvlist_src.append(mcd_mean)
mcdstd_cvlist_src.append(mcd_std)
if trg_exist:
spcidx_trg = np.array(read_hdf5(file_trg, "/spcidx_range")[0])
_, _, _, mcdpow_arr = dtw.dtw_org_to_trg(np.array(cvmcep[spcidx], \
dtype=np.float64), np.array(mcep_trg[spcidx_trg], dtype=np.float64))
_, _, _, mcd_arr = dtw.dtw_org_to_trg(np.array(cvmcep[spcidx,1:], \
dtype=np.float64), np.array(mcep_trg[spcidx_trg,1:], dtype=np.float64))
mcdpow_mean = np.mean(mcdpow_arr)
mcdpow_std = np.std(mcdpow_arr)
mcd_mean = np.mean(mcd_arr)
mcd_std = np.std(mcd_arr)
logging.info("mcdpow_trg: %.6f dB +- %.6f" % (mcdpow_mean, mcdpow_std))
logging.info("mcd_trg: %.6f dB +- %.6f" % (mcd_mean, mcd_std))
mcdpow_cvlist.append(mcdpow_mean)
mcdpowstd_cvlist.append(mcdpow_std)
mcd_cvlist.append(mcd_mean)
mcdstd_cvlist.append(mcd_std)
spcidx_src = torch.LongTensor(spcidx).cuda()
spcidx_trg = torch.LongTensor(spcidx_trg).cuda()
trj_lat_src = np.array(torch.index_select(lat_src[0],0,spcidx_src).cpu().data.numpy(), dtype=np.float64)
trj_lat_trg = np.array(torch.index_select(lat_trg[0],0,spcidx_trg).cpu().data.numpy(), dtype=np.float64)
aligned_lat_srctrg, _, _, _ = dtw.dtw_org_to_trg(trj_lat_src, trj_lat_trg)
lat_dist_srctrg = np.mean(np.sqrt(np.mean((aligned_lat_srctrg-trj_lat_trg)**2, axis=0)))
_, _, lat_cdist_srctrg, _ = dtw.dtw_org_to_trg(trj_lat_trg, trj_lat_src, mcd=0)
aligned_lat_trgsrc, _, _, _ = dtw.dtw_org_to_trg(trj_lat_trg, trj_lat_src)
lat_dist_trgsrc = np.mean(np.sqrt(np.mean((aligned_lat_trgsrc-trj_lat_src)**2, axis=0)))
_, _, lat_cdist_trgsrc, _ = dtw.dtw_org_to_trg(trj_lat_src, trj_lat_trg, mcd=0)
logging.info("%lf %lf %lf %lf" % (lat_dist_srctrg, lat_cdist_srctrg, lat_dist_trgsrc, lat_cdist_trgsrc))
lat_dist_rmse = (lat_dist_srctrg+lat_dist_trgsrc)/2
lat_dist_cosim = (lat_cdist_srctrg+lat_cdist_trgsrc)/2
lat_dist_rmse_list.append(lat_dist_rmse)
lat_dist_cosim_list.append(lat_dist_cosim)
logging.info("lat_dist: %.6f %.6f" % (lat_dist_rmse, lat_dist_cosim))
_, mcdpow_arr = dtw.calc_mcd(np.array(mcep[spcidx], dtype=np.float64), np.array(cvmcep_cyc[spcidx], dtype=np.float64))
_, mcd_arr = dtw.calc_mcd(np.array(mcep[spcidx,1:], dtype=np.float64), np.array(cvmcep_cyc[spcidx,1:], dtype=np.float64))
mcdpow_mean = np.mean(mcdpow_arr)
mcdpow_std = np.std(mcdpow_arr)
mcd_mean = np.mean(mcd_arr)
mcd_std = np.std(mcd_arr)
logging.info("mcdpow_cyc_cv: %.6f dB +- %.6f" % (mcdpow_mean, mcdpow_std))
logging.info("mcd_cyc_cv: %.6f dB +- %.6f" % (mcd_mean, mcd_std))
mcdpow_cvlist_cyc.append(mcdpow_mean)
mcdpowstd_cvlist_cyc.append(mcdpow_std)
mcd_cvlist_cyc.append(mcd_mean)
mcdstd_cvlist_cyc.append(mcd_std)
logging.info("synth anasyn")
wav = np.clip(pw.synthesize(f0_range, sp, ap, args.fs, frame_period=args.shiftms), -1, 1)
wavpath = os.path.join(args.outdir,os.path.basename(feat_file).replace(".h5","_anasyn.wav"))
sf.write(wavpath, wav, args.fs, 'PCM_16')
logging.info(wavpath)
logging.info("synth voco rec")
cvsp_src = ps.mc2sp(cvmcep_src, args.mcep_alpha, args.fftl)
logging.info(cvsp_src.shape)
wav = np.clip(pw.synthesize(f0_range, cvsp_src, ap, args.fs, frame_period=args.shiftms), -1, 1)
wavpath = os.path.join(args.outdir, os.path.basename(feat_file).replace(".h5", "_rec.wav"))
sf.write(wavpath, wav, args.fs, 'PCM_16')
logging.info(wavpath)
logging.info("synth voco cv")
cvsp = ps.mc2sp(cvmcep, args.mcep_alpha, args.fftl)
logging.info(cvsp.shape)
wav = np.clip(pw.synthesize(cvf0_range_lin, cvsp, ap, args.fs, frame_period=args.shiftms), -1, 1)
wavpath = os.path.join(args.outdir, os.path.basename(feat_file).replace(".h5", "_cv.wav"))
sf.write(wavpath, wav, args.fs, 'PCM_16')
logging.info(wavpath)
logging.info("synth voco cv GV")
datamean = np.mean(cvmcep[:,1:], axis=0)
cvmcep_gv = np.c_[cvmcep[:,0], args.gv_coeff*(np.sqrt(gv_mean_trg/cvgv_mean) * \
(cvmcep[:,1:]-datamean) + datamean) + (1-args.gv_coeff)*cvmcep[:,1:]]
cvmcep_gv = mod_pow(cvmcep_gv, cvmcep, alpha=args.mcep_alpha, irlen=IRLEN)
cvsp_gv = ps.mc2sp(cvmcep_gv, args.mcep_alpha, args.fftl)
logging.info(cvsp_gv.shape)
wav = np.clip(pw.synthesize(cvf0_range_lin, cvsp_gv, ap, args.fs, frame_period=args.shiftms), -1, 1)
wavpath = os.path.join(args.outdir, os.path.basename(feat_file).replace(".h5", "_cvGV.wav"))
sf.write(wavpath, wav, args.fs, 'PCM_16')
logging.info(wavpath)
#logging.info("synth diffGV")
#shiftl = int(args.fs/1000*args.shiftms)
#mc_cv_diff = cvmcep_gv-mcep
#b = np.apply_along_axis(ps.mc2b, 1, mc_cv_diff, args.mcep_alpha)
#logging.info(b.shape)
#assert np.isfinite(b).all
#mlsa_fil = ps.synthesis.Synthesizer(MLSADF(mcep_dim, alpha=args.mcep_alpha), shiftl)
#x, fs_ = sf.read(os.path.join(os.path.dirname(feat_file).replace("hdf5", "wav_filtered"), os.path.basename(feat_file).replace(".h5", ".wav")))
#assert(fs_ == args.fs)
#wav = mlsa_fil.synthesis(x, b)
#wav = np.clip(wav, -1, 1)
#wavpath = os.path.join(args.outdir, os.path.basename(feat_file).replace(".h5", "_DiffGV.wav"))
#sf.write(wavpath, wav, args.fs, 'PCM_16')
#logging.info(wavpath)
#logging.info("synth diffGVF0")
#time_axis = read_hdf5(feat_file, "/time_axis")
#sp_diff = pw.cheaptrick(wav, f0_range, time_axis, args.fs, fft_size=args.fftl)
#logging.info(sp_diff.shape)
#ap_diff = pw.d4c(wav, f0_range, time_axis, args.fs, fft_size=args.fftl)
#logging.info(ap_diff.shape)
#wav = pw.synthesize(cvf0_range_lin, sp_diff, ap_diff, args.fs, frame_period=args.shiftms)
#wav = np.clip(wav, -1, 1)
#wavpath = os.path.join(args.outdir,os.path.basename(feat_file).replace(".h5", "_DiffGVF0.wav"))
#sf.write(wavpath, wav, args.fs, 'PCM_16')
#logging.info(wavpath)
#logging.info("analysis diffGVF0")
#sp_diff_anasyn = pw.cheaptrick(wav, cvf0_range_lin, time_axis, args.fs, fft_size=args.fftl)
#logging.info(sp_diff_anasyn.shape)
#mc_cv_diff_anasyn = ps.sp2mc(sp_diff_anasyn, mcep_dim, args.mcep_alpha)
#ap_diff_anasyn = pw.d4c(wav, cvf0_range_lin, time_axis, args.fs, fft_size=args.fftl)
#code_ap_diff_anasyn = pw.code_aperiodicity(ap_diff_anasyn, args.fs)
## convert to continouos codeap with uv
#for i in range(code_ap_diff_anasyn.shape[-1]):
# logging.info('codeap: %d' % (i+1))
# uv_codeap_i, cont_codeap_i = convert_continuos_codeap(np.array(code_ap_diff_anasyn[:,i]))
# cont_codeap_i = np.log(-np.clip(cont_codeap_i, a_min=np.amin(cont_codeap_i), a_max=MAX_CODEAP))
# if i > 0:
# cont_codeap = np.c_[cont_codeap, np.expand_dims(cont_codeap_i, axis=-1)]
# else:
# uv_codeap = np.expand_dims(uv_codeap_i, axis=-1)
# cont_codeap = np.expand_dims(cont_codeap_i, axis=-1)
# uv_codeap_i = np.expand_dims(uv_codeap_i, axis=-1)
# unique, counts = np.unique(uv_codeap_i, return_counts=True)
# logging.info(dict(zip(unique, counts)))
## postprocessed converted feat for neural vocoder
#feat_diffgv_anasyn = np.c_[feat_cv[:,:2], uv_codeap, cont_codeap, mc_cv_diff_anasyn]
#logging.info("write lat")
#outTxtDir = os.path.join(args.outdir, os.path.basename(os.path.dirname(feat_file)))
#if not os.path.exists(outTxtDir):
# os.mkdir(outTxtDir)
#outTxt = os.path.join(outTxtDir, os.path.basename(feat_file).replace(".wav", ".txt"))
#logging.info(outTxt)
#g = open(outTxt, "wt")
#idx_frm = 0
#nfrm = trj_lat_src.shape[0]
#dim = trj_lat_src.shape[1]
#if not args.time_flag:
##if True:
# while idx_frm < nfrm:
# idx_elmt = 1
# for elmt in trj_lat_src[idx_frm]:
# if idx_elmt < dim:
# g.write("%lf " % (elmt))
# else:
# g.write("%lf\n" % (elmt))
# idx_elmt += 1
# idx_frm += 1
#else:
# while idx_frm < nfrm:
# idx_elmt = 1
# for elmt in trj_lat_src[idx_frm]:
# if idx_elmt < dim:
# if idx_elmt > 1:
# g.write("%lf " % (elmt))
# else:
# g.write("%lf %lf " % (time_axis[idx_frm], elmt))
# else:
# g.write("%lf\n" % (elmt))
# idx_elmt += 1
# idx_frm += 1
#g.close()
logging.info('write to h5')
outh5dir = os.path.join(os.path.dirname(os.path.dirname(feat_file)), spk_src+"-"+args.spk_trg)
if not os.path.exists(outh5dir):
os.makedirs(outh5dir)
feat_file = os.path.join(outh5dir, os.path.basename(feat_file))
# cv
write_path = args.string_path
logging.info(feat_file + ' ' + write_path)
logging.info(feat_cv.shape)
write_hdf5(feat_file, write_path, feat_cv)
## diffGVF0
#write_path = args.string_path+"_diffgvf0"
#logging.info(feat_file + ' ' + write_path)
#logging.info(feat_diffgv_anasyn.shape)
#write_hdf5(feat_file, write_path, feat_diffgv_anasyn)
count += 1
def gpu_decode(feat_list, gpu, cvlist=None, mcdpow_cvlist=None, mcdpowstd_cvlist=None, mcd_cvlist=None, \
mcdstd_cvlist=None, cvlist_dv=None, mcdpow_cvlist_dv=None, mcdpowstd_cvlist_dv=None, \
mcd_cvlist_dv=None, mcdstd_cvlist_dv=None, \
cvlist_cyc=None, mcdpow_cvlist_cyc=None, mcdpowstd_cvlist_cyc=None, mcd_cvlist_cyc=None, \
mcdstd_cvlist_cyc=None, cvlist_cyc_dv=None, mcdpow_cvlist_cyc_dv=None, mcdpowstd_cvlist_cyc_dv=None, \
mcd_cvlist_cyc_dv=None, mcdstd_cvlist_cyc_dv=None):
with torch.cuda.device(gpu):
# define model and load parameters
with torch.no_grad():
model_encoder = GRU_VAE_ENCODER(
in_dim=config.mcep_dim + config.excit_dim,
n_spk=n_spk,
lat_dim=config.lat_dim,
hidden_layers=config.hidden_layers_enc,
hidden_units=config.hidden_units_enc,
kernel_size=config.kernel_size_enc,
dilation_size=config.dilation_size_enc,
causal_conv=config.causal_conv_enc,
bi=False,
ar=False,
pad_first=True,
right_size=config.right_size_enc)
logging.info(model_encoder)
model_decoder = GRU_SPEC_DECODER(
feat_dim=config.lat_dim,
out_dim=config.mcep_dim,
n_spk=n_spk,
hidden_layers=config.hidden_layers_dec,
hidden_units=config.hidden_units_dec,
kernel_size=config.kernel_size_dec,
dilation_size=config.dilation_size_dec,
causal_conv=config.causal_conv_dec,
bi=False,
ar=False,
pad_first=True,
right_size=config.right_size_dec)
logging.info(model_decoder)
model_post = GRU_POST_NET(
spec_dim=config.mcep_dim,
excit_dim=2,
n_spk=n_spk,
hidden_layers=config.hidden_layers_post,
hidden_units=config.hidden_units_post,
kernel_size=config.kernel_size_post,
dilation_size=config.dilation_size_post,
causal_conv=config.causal_conv_post,
pad_first=True,
right_size=config.right_size_post)
#excit_dim=config.excit_dim,
#excit_dim=None,
logging.info(model_post)
model_encoder.load_state_dict(
torch.load(args.model)["model_encoder"])
model_decoder.load_state_dict(
torch.load(args.model)["model_decoder"])
model_post.load_state_dict(
torch.load(args.model)["model_post"])
model_encoder.remove_weight_norm()
model_decoder.remove_weight_norm()
model_post.remove_weight_norm()
model_encoder.cuda()
model_decoder.cuda()
model_post.cuda()
model_encoder.eval()
model_decoder.eval()
model_post.eval()
for param in model_encoder.parameters():
param.requires_grad = False
for param in model_decoder.parameters():
param.requires_grad = False
for param in model_post.parameters():
param.requires_grad = False
count = 0
pad_left = (model_encoder.pad_left + model_decoder.pad_left +
model_post.pad_left) * 2
pad_right = (model_encoder.pad_right + model_decoder.pad_right +
model_post.pad_right) * 2
outpad_lefts = [None] * 5
outpad_rights = [None] * 5
outpad_lefts[0] = pad_left - model_encoder.pad_left
outpad_rights[0] = pad_right - model_encoder.pad_right
outpad_lefts[1] = outpad_lefts[0] - model_decoder.pad_left
outpad_rights[1] = outpad_rights[0] - model_decoder.pad_right
outpad_lefts[2] = outpad_lefts[1] - model_post.pad_left
outpad_rights[2] = outpad_rights[1] - model_post.pad_right
outpad_lefts[3] = outpad_lefts[2] - model_encoder.pad_left
outpad_rights[3] = outpad_rights[2] - model_encoder.pad_right
outpad_lefts[4] = outpad_lefts[3] - model_decoder.pad_left
outpad_rights[4] = outpad_rights[3] - model_decoder.pad_right
logging.info(f'{pad_left} {pad_right}')
logging.info(outpad_lefts)
logging.info(outpad_rights)
for feat_file in feat_list:
# convert mcep
logging.info("recmcep " + feat_file)
feat_org = read_hdf5(feat_file, "/feat_mceplf0cap")
logging.info(feat_org.shape)
mcep = np.array(feat_org[:, -config.mcep_dim:])
with torch.no_grad():
feat = F.pad(
torch.FloatTensor(feat_org).cuda().unsqueeze(
0).transpose(1, 2), (pad_left, pad_right),
"replicate").transpose(1, 2)
feat_excit = torch.FloatTensor(
feat_org[:, :config.excit_dim]).cuda().unsqueeze(0)
spk_logits, _, lat_src, _ = model_encoder(feat,
sampling=False)
logging.info('input spkpost')
if outpad_rights[0] > 0:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[0]:
-outpad_rights[0]],
dim=-1), 1))
else:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[0]:],
dim=-1), 1))
cvmcep_src, _ = model_decoder((torch.ones(
(1, lat_src.shape[1])) * spk_idx).cuda().long(),
lat_src)
cvmcep_src_post, _ = model_post(
cvmcep_src,
y=(torch.ones(
(1, cvmcep_src.shape[1])) * spk_idx).cuda().long(),
e=F.pad(feat_excit[:, :, :2].transpose(1, 2),
(outpad_lefts[1], outpad_rights[1]),
"replicate").transpose(1, 2))
#e=F.pad(feat_excit.transpose(1,2), (outpad_lefts[1],outpad_rights[1]), "replicate").transpose(1,2))
if model_post.pad_right > 0:
spk_logits, _, lat_rec, _ = model_encoder(torch.cat((F.pad(feat_excit.transpose(1,2), \
(outpad_lefts[2],outpad_rights[2]), "replicate").transpose(1,2), cvmcep_src[:,model_post.pad_left:-model_post.pad_right]), 2),
sampling=False)
else:
spk_logits, _, lat_rec, _ = model_encoder(torch.cat((F.pad(feat_excit.transpose(1,2), \
(outpad_lefts[2],outpad_rights[2]), "replicate").transpose(1,2), cvmcep_src[:,model_post.pad_left:]), 2),
sampling=False)
logging.info('rec spkpost')
if outpad_rights[3] > 0:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[3]:
-outpad_rights[3]],
dim=-1), 1))
else:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[3]:],
dim=-1), 1))
cvmcep_cyc, _ = model_decoder((torch.ones(
(1, lat_rec.shape[1])) * spk_idx).cuda().long(),
lat_rec)
cvmcep_cyc_post, _ = model_post(
cvmcep_cyc,
y=(torch.ones(
(1, cvmcep_cyc.shape[1])) * spk_idx).cuda().long(),
e=F.pad(feat_excit[:, :, :2].transpose(1, 2),
(outpad_lefts[4], outpad_rights[4]),
"replicate").transpose(1, 2))
#e=F.pad(feat_excit.transpose(1,2), (outpad_lefts[4],outpad_rights[4]), "replicate").transpose(1,2))
if outpad_rights[2] > 0:
feat_rec = torch.cat(
(feat_excit,
cvmcep_src_post[:,
outpad_lefts[2]:-outpad_rights[2]]
), 2)[0].cpu().data.numpy()
else:
feat_rec = torch.cat(
(feat_excit, cvmcep_src_post[:, outpad_lefts[2]:]),
2)[0].cpu().data.numpy()
feat_cyc = torch.cat((feat_excit, cvmcep_cyc_post),
2)[0].cpu().data.numpy()
if outpad_rights[2] > 0:
cvmcep_src = np.array(
cvmcep_src_post[:,
outpad_lefts[2]:-outpad_rights[2]]
[0].cpu().data.numpy(),
dtype=np.float64)
else:
cvmcep_src = np.array(cvmcep_src_post[:,
outpad_lefts[2]:]
[0].cpu().data.numpy(),
dtype=np.float64)
cvmcep_cyc = np.array(
cvmcep_cyc_post[0].cpu().data.numpy(),
dtype=np.float64)
logging.info(cvmcep_src.shape)
logging.info(cvmcep_cyc.shape)
spcidx = read_hdf5(feat_file, "/spcidx_range")[0]
_, _, _, mcdpow_arr = dtw.dtw_org_to_trg(np.array(cvmcep_src[np.array(spcidx),:], \
dtype=np.float64), np.array(mcep[np.array(spcidx),:], dtype=np.float64))
_, _, _, mcd_arr = dtw.dtw_org_to_trg(np.array(cvmcep_src[np.array(spcidx),1:], \
dtype=np.float64), np.array(mcep[np.array(spcidx),1:], dtype=np.float64))
mcdpow_mean = np.mean(mcdpow_arr)
mcdpow_std = np.std(mcdpow_arr)
mcd_mean = np.mean(mcd_arr)
mcd_std = np.std(mcd_arr)
logging.info("mcdpow_rec: %.6f dB +- %.6f" %
(mcdpow_mean, mcdpow_std))
logging.info("mcd_rec: %.6f dB +- %.6f" % (mcd_mean, mcd_std))
_, _, _, mcdpow_arr = dtw.dtw_org_to_trg(np.array(cvmcep_cyc[np.array(spcidx),:], \
dtype=np.float64), np.array(mcep[np.array(spcidx),:], dtype=np.float64))
_, _, _, mcd_arr = dtw.dtw_org_to_trg(np.array(cvmcep_cyc[np.array(spcidx),1:], \
dtype=np.float64), np.array(mcep[np.array(spcidx),1:], dtype=np.float64))
mcdpow_mean_cyc = np.mean(mcdpow_arr)
mcdpow_std_cyc = np.std(mcdpow_arr)
mcd_mean_cyc = np.mean(mcd_arr)
mcd_std_cyc = np.std(mcd_arr)
logging.info("mcdpow_cyc: %.6f dB +- %.6f" %
(mcdpow_mean_cyc, mcdpow_std_cyc))
logging.info("mcd_cyc: %.6f dB +- %.6f" %
(mcd_mean_cyc, mcd_std_cyc))
dataset = feat_file.split('/')[1].split('_')[0]
if 'tr' in dataset:
logging.info('trn')
mcdpow_cvlist.append(mcdpow_mean)
mcdpowstd_cvlist.append(mcdpow_std)
mcd_cvlist.append(mcd_mean)
mcdstd_cvlist.append(mcd_std)
cvlist.append(np.var(cvmcep_src[:, 1:], axis=0))
logging.info(len(cvlist))
mcdpow_cvlist_cyc.append(mcdpow_mean_cyc)
mcdpowstd_cvlist_cyc.append(mcdpow_std_cyc)
mcd_cvlist_cyc.append(mcd_mean_cyc)
mcdstd_cvlist_cyc.append(mcd_std_cyc)
cvlist_cyc.append(np.var(cvmcep_cyc[:, 1:], axis=0))
elif 'dv' in dataset:
logging.info('dev')
mcdpow_cvlist_dv.append(mcdpow_mean)
mcdpowstd_cvlist_dv.append(mcdpow_std)
mcd_cvlist_dv.append(mcd_mean)
mcdstd_cvlist_dv.append(mcd_std)
cvlist_dv.append(np.var(cvmcep_src[:, 1:], axis=0))
logging.info(len(cvlist_dv))
mcdpow_cvlist_cyc_dv.append(mcdpow_mean_cyc)
mcdpowstd_cvlist_cyc_dv.append(mcdpow_std_cyc)
mcd_cvlist_cyc_dv.append(mcd_mean_cyc)
mcdstd_cvlist_cyc_dv.append(mcd_std_cyc)
cvlist_cyc_dv.append(np.var(cvmcep_cyc[:, 1:], axis=0))
logging.info('write rec to h5')
outh5dir = os.path.join(
os.path.dirname(os.path.dirname(feat_file)),
args.spk + "-" + args.spk)
if not os.path.exists(outh5dir):
os.makedirs(outh5dir)
feat_file = os.path.join(outh5dir, os.path.basename(feat_file))
logging.info(feat_file + ' ' + args.string_path)
logging.info(feat_rec.shape)
write_hdf5(feat_file, args.string_path, feat_rec)
logging.info('write cyc to h5')
outh5dir = os.path.join(
os.path.dirname(os.path.dirname(feat_file)),
args.spk + "-" + args.spk + "-" + args.spk)
if not os.path.exists(outh5dir):
os.makedirs(outh5dir)
feat_file = os.path.join(outh5dir, os.path.basename(feat_file))
logging.info(feat_file + ' ' + args.string_path)
logging.info(feat_cyc.shape)
write_hdf5(feat_file, args.string_path, feat_cyc)
count += 1
def gpu_decode(feat_list,
gpu,
cvlist=None,
lsd_cvlist=None,
lsdstd_cvlist=None,
cvlist_dv=None,
lsd_cvlist_dv=None,
lsdstd_cvlist_dv=None,
f0rmse_cvlist=None,
f0corr_cvlist=None,
caprmse_cvlist=None,
f0rmse_cvlist_dv=None,
f0corr_cvlist_dv=None,
caprmse_cvlist_dv=None,
cvlist_cyc=None,
lsd_cvlist_cyc=None,
lsdstd_cvlist_cyc=None,
cvlist_cyc_dv=None,
lsd_cvlist_cyc_dv=None,
lsdstd_cvlist_cyc_dv=None,
f0rmse_cvlist_cyc=None,
f0corr_cvlist_cyc=None,
caprmse_cvlist_cyc=None,
f0rmse_cvlist_cyc_dv=None,
f0corr_cvlist_cyc_dv=None,
caprmse_cvlist_cyc_dv=None):
with torch.cuda.device(gpu):
# define model and load parameters
with torch.no_grad():
model_encoder_melsp = GRU_VAE_ENCODER(
in_dim=config.mel_dim,
n_spk=n_spk,
lat_dim=config.lat_dim,
hidden_layers=config.hidden_layers_enc,
hidden_units=config.hidden_units_enc,
kernel_size=config.kernel_size_enc,
dilation_size=config.dilation_size_enc,
causal_conv=config.causal_conv_enc,
bi=False,
ar=False,
pad_first=True,
right_size=config.right_size_enc)
logging.info(model_encoder_melsp)
model_decoder_melsp = GRU_SPEC_DECODER(
feat_dim=config.lat_dim + config.lat_dim_e,
excit_dim=config.excit_dim,
out_dim=config.mel_dim,
n_spk=n_spk,
hidden_layers=config.hidden_layers_dec,
hidden_units=config.hidden_units_dec,
kernel_size=config.kernel_size_dec,
dilation_size=config.dilation_size_dec,
causal_conv=config.causal_conv_dec,
bi=False,
ar=False,
pad_first=True,
right_size=config.right_size_dec)
logging.info(model_decoder_melsp)
model_encoder_excit = GRU_VAE_ENCODER(
in_dim=config.mel_dim,
n_spk=n_spk,
lat_dim=config.lat_dim_e,
hidden_layers=config.hidden_layers_enc,
hidden_units=config.hidden_units_enc,
kernel_size=config.kernel_size_enc,
dilation_size=config.dilation_size_enc,
causal_conv=config.causal_conv_enc,
bi=False,
ar=False,
pad_first=True,
right_size=config.right_size_enc)
logging.info(model_encoder_excit)
model_decoder_excit = GRU_EXCIT_DECODER(
feat_dim=config.lat_dim_e,
cap_dim=config.cap_dim,
n_spk=n_spk,
hidden_layers=config.hidden_layers_lf0,
hidden_units=config.hidden_units_lf0,
kernel_size=config.kernel_size_lf0,
dilation_size=config.dilation_size_lf0,
causal_conv=config.causal_conv_lf0,
bi=False,
ar=False,
pad_first=True,
right_size=config.right_size_lf0)
logging.info(model_decoder_excit)
if (config.spkidtr_dim > 0):
model_spkidtr = SPKID_TRANSFORM_LAYER(
n_spk=n_spk, spkidtr_dim=config.spkidtr_dim)
logging.info(model_spkidtr)
model_encoder_melsp.load_state_dict(
torch.load(args.model)["model_encoder_melsp"])
model_decoder_melsp.load_state_dict(
torch.load(args.model)["model_decoder_melsp"])
model_encoder_excit.load_state_dict(
torch.load(args.model)["model_encoder_excit"])
model_decoder_excit.load_state_dict(
torch.load(args.model)["model_decoder_excit"])
if (config.spkidtr_dim > 0):
model_spkidtr.load_state_dict(
torch.load(args.model)["model_spkidtr"])
model_encoder_melsp.cuda()
model_decoder_melsp.cuda()
model_encoder_excit.cuda()
model_decoder_excit.cuda()
if (config.spkidtr_dim > 0):
model_spkidtr.cuda()
model_encoder_melsp.eval()
model_decoder_melsp.eval()
model_encoder_excit.eval()
model_decoder_excit.eval()
if (config.spkidtr_dim > 0):
model_spkidtr.eval()
for param in model_encoder_melsp.parameters():
param.requires_grad = False
for param in model_decoder_melsp.parameters():
param.requires_grad = False
for param in model_encoder_excit.parameters():
param.requires_grad = False
for param in model_decoder_excit.parameters():
param.requires_grad = False
if (config.spkidtr_dim > 0):
for param in model_spkidtr.parameters():
param.requires_grad = False
count = 0
pad_left = (model_encoder_melsp.pad_left +
model_decoder_excit.pad_left +
model_decoder_melsp.pad_left) * 2
pad_right = (model_encoder_melsp.pad_right +
model_decoder_excit.pad_right +
model_decoder_melsp.pad_right) * 2
outpad_lefts = [None] * 5
outpad_rights = [None] * 5
outpad_lefts[0] = pad_left - model_encoder_melsp.pad_left
outpad_rights[0] = pad_right - model_encoder_melsp.pad_right
outpad_lefts[1] = outpad_lefts[0] - model_decoder_excit.pad_left
outpad_rights[1] = outpad_rights[0] - model_decoder_excit.pad_right
outpad_lefts[2] = outpad_lefts[1] - model_decoder_melsp.pad_left
outpad_rights[2] = outpad_rights[1] - model_decoder_melsp.pad_right
outpad_lefts[3] = outpad_lefts[2] - model_encoder_melsp.pad_left
outpad_rights[3] = outpad_rights[2] - model_encoder_melsp.pad_right
outpad_lefts[4] = outpad_lefts[3] - model_decoder_excit.pad_left
outpad_rights[4] = outpad_rights[3] - model_decoder_excit.pad_right
for feat_file in feat_list:
# reconst. melsp
logging.info("recmelsp " + feat_file)
feat_org = read_hdf5(feat_file, "/log_1pmelmagsp")
logging.info(feat_org.shape)
with torch.no_grad():
feat = F.pad(
torch.FloatTensor(feat_org).cuda().unsqueeze(
0).transpose(1, 2), (pad_left, pad_right),
"replicate").transpose(1, 2)
spk_logits, _, lat_src, _ = model_encoder_melsp(
feat, sampling=False)
spk_logits_e, _, lat_src_e, _ = model_encoder_excit(
feat, sampling=False)
logging.info('input spkpost')
if outpad_rights[0] > 0:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[0]:
-outpad_rights[0]],
dim=-1), 1))
else:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[0]:],
dim=-1), 1))
logging.info('input spkpost_e')
if outpad_rights[0] > 0:
logging.info(
torch.mean(
F.softmax(spk_logits_e[:, outpad_lefts[0]:
-outpad_rights[0]],
dim=-1), 1))
else:
logging.info(
torch.mean(
F.softmax(spk_logits_e[:, outpad_lefts[0]:],
dim=-1), 1))
if config.spkidtr_dim > 0:
src_code = model_spkidtr((torch.ones(
(1, lat_src_e.shape[1])) * spk_idx).cuda().long())
else:
src_code = (torch.ones(
(1, lat_src_e.shape[1])) * spk_idx).cuda().long()
cvlf0_src, _ = model_decoder_excit(src_code, lat_src_e)
if model_decoder_excit.pad_right > 0:
lat_cat = torch.cat((
lat_src_e[:, model_decoder_excit.
pad_left:-model_decoder_excit.pad_right],
lat_src[:, model_decoder_excit.
pad_left:-model_decoder_excit.pad_right]),
2)
else:
lat_cat = torch.cat(
(lat_src_e[:, model_decoder_excit.pad_left:],
lat_src[:, model_decoder_excit.pad_left:]), 2)
if config.spkidtr_dim > 0:
src_code = model_spkidtr((torch.ones(
(1, lat_cat.shape[1])) * spk_idx).cuda().long())
else:
src_code = (torch.ones(
(1, lat_cat.shape[1])) * spk_idx).cuda().long()
cvmelsp_src, _ = model_decoder_melsp(
lat_cat,
y=src_code,
e=cvlf0_src[:, :, :config.excit_dim])
spk_logits, _, lat_rec, _ = model_encoder_melsp(
cvmelsp_src, sampling=False)
spk_logits_e, _, lat_rec_e, _ = model_encoder_excit(
cvmelsp_src, sampling=False)
logging.info('rec spkpost')
if outpad_rights[3] > 0:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[3]:
-outpad_rights[3]],
dim=-1), 1))
else:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[3]:],
dim=-1), 1))
logging.info('rec spkpost_e')
if outpad_rights[3] > 0:
logging.info(
torch.mean(
F.softmax(spk_logits_e[:, outpad_lefts[3]:
-outpad_rights[3]],
dim=-1), 1))
else:
logging.info(
torch.mean(
F.softmax(spk_logits_e[:, outpad_lefts[3]:],
dim=-1), 1))
if config.spkidtr_dim > 0:
src_code = model_spkidtr((torch.ones(
(1, lat_rec_e.shape[1])) * spk_idx).cuda().long())
else:
src_code = (torch.ones(
(1, lat_rec_e.shape[1])) * spk_idx).cuda().long()
cvlf0_cyc, _ = model_decoder_excit(src_code, lat_rec_e)
if model_decoder_excit.pad_right > 0:
lat_cat = torch.cat((
lat_rec_e[:, model_decoder_excit.
pad_left:-model_decoder_excit.pad_right],
lat_rec[:, model_decoder_excit.
pad_left:-model_decoder_excit.pad_right]),
2)
else:
lat_cat = torch.cat(
(lat_rec_e[:, model_decoder_excit.pad_left:],
lat_rec[:, model_decoder_excit.pad_left:]), 2)
if config.spkidtr_dim > 0:
src_code = model_spkidtr((torch.ones(
(1, lat_cat.shape[1])) * spk_idx).cuda().long())
else:
src_code = (torch.ones(
(1, lat_cat.shape[1])) * spk_idx).cuda().long()
cvmelsp_cyc, _ = model_decoder_melsp(
lat_cat,
y=src_code,
e=cvlf0_cyc[:, :, :config.excit_dim])
if outpad_rights[1] > 0:
cvlf0_src = cvlf0_src[:, outpad_lefts[1]:
-outpad_rights[1]]
else:
cvlf0_src = cvlf0_src[:, outpad_lefts[1]:]
if outpad_rights[2] > 0:
cvmelsp_src = cvmelsp_src[:, outpad_lefts[2]:
-outpad_rights[2]]
else:
cvmelsp_src = cvmelsp_src[:, outpad_lefts[2]:]
if outpad_rights[4] > 0:
cvlf0_cyc = cvlf0_cyc[:, outpad_lefts[4]:
-outpad_rights[4]]
else:
cvlf0_cyc = cvlf0_cyc[:, outpad_lefts[4]:]
feat_rec = cvmelsp_src[0].cpu().data.numpy()
feat_cyc = cvmelsp_cyc[0].cpu().data.numpy()
cvmelsp_src = np.array(cvmelsp_src[0].cpu().data.numpy(),
dtype=np.float64)
cvlf0_src = np.array(cvlf0_src[0].cpu().data.numpy(),
dtype=np.float64)
cvmelsp_cyc = np.array(cvmelsp_cyc[0].cpu().data.numpy(),
dtype=np.float64)
cvlf0_cyc = np.array(cvlf0_cyc[0].cpu().data.numpy(),
dtype=np.float64)
logging.info(cvlf0_src.shape)
logging.info(cvmelsp_src.shape)
logging.info(cvlf0_cyc.shape)
logging.info(cvmelsp_cyc.shape)
melsp = np.array(feat_org)
feat_world = read_hdf5(feat_file, "/feat_mceplf0cap")
f0 = np.array(
np.rint(feat_world[:, 0]) * np.exp(feat_world[:, 1]))
codeap = np.array(
np.rint(feat_world[:, 2:3]) *
(-np.exp(feat_world[:, 3:config.full_excit_dim])))
cvf0_src = np.array(
np.rint(cvlf0_src[:, 0]) * np.exp(cvlf0_src[:, 1]))
cvcodeap_src = np.array(
np.rint(cvlf0_src[:, 2:3]) * (-np.exp(cvlf0_src[:, 3:])))
f0_rmse = np.sqrt(np.mean((cvf0_src - f0)**2))
logging.info('F0_rmse_rec: %lf Hz' % (f0_rmse))
cvf0_src_mean = np.mean(cvf0_src)
f0_mean = np.mean(f0)
f0_corr = np.sum((cvf0_src-cvf0_src_mean)*(f0-f0_mean))/\
(np.sqrt(np.sum((cvf0_src-cvf0_src_mean)**2))*np.sqrt(np.sum((f0-f0_mean)**2)))
logging.info('F0_corr_rec: %lf' % (f0_corr))
codeap_rmse = np.sqrt(
np.mean((cvcodeap_src - codeap)**2, axis=0))
for i in range(codeap_rmse.shape[-1]):
logging.info('codeap-%d_rmse_rec: %lf dB' %
(i + 1, codeap_rmse[i]))
cvf0_cyc = np.array(
np.rint(cvlf0_cyc[:, 0]) * np.exp(cvlf0_cyc[:, 1]))
cvcodeap_cyc = np.array(
np.rint(cvlf0_cyc[:, 2:3]) * (-np.exp(cvlf0_cyc[:, 3:])))
f0_rmse_cyc = np.sqrt(np.mean((cvf0_cyc - f0)**2))
logging.info('F0_rmse_cyc: %lf Hz' % (f0_rmse_cyc))
cvf0_cyc_mean = np.mean(cvf0_cyc)
f0_mean = np.mean(f0)
f0_corr_cyc = np.sum((cvf0_cyc-cvf0_cyc_mean)*(f0-f0_mean))/\
(np.sqrt(np.sum((cvf0_cyc-cvf0_cyc_mean)**2))*np.sqrt(np.sum((f0-f0_mean)**2)))
logging.info('F0_corr_cyc: %lf' % (f0_corr_cyc))
codeap_rmse_cyc = np.sqrt(
np.mean((cvcodeap_cyc - codeap)**2, axis=0))
for i in range(codeap_rmse_cyc.shape[-1]):
logging.info('codeap-%d_rmse_cyc: %lf dB' %
(i + 1, codeap_rmse_cyc[i]))
spcidx = np.array(read_hdf5(feat_file, "/spcidx_range")[0])
melsp_rest = (np.exp(melsp) - 1) / 10000
melsp_src_rest = (np.exp(cvmelsp_src) - 1) / 10000
melsp_cyc_rest = (np.exp(cvmelsp_cyc) - 1) / 10000
lsd_arr = np.sqrt(np.mean((20*(np.log10(np.clip(melsp_src_rest[spcidx], a_min=1e-16, a_max=None))\
-np.log10(np.clip(melsp_rest[spcidx], a_min=1e-16, a_max=None))))**2, axis=-1))
lsd_mean = np.mean(lsd_arr)
lsd_std = np.std(lsd_arr)
logging.info("lsd_rec: %.6f dB +- %.6f" % (lsd_mean, lsd_std))
lsd_arr = np.sqrt(np.mean((20*(np.log10(np.clip(melsp_cyc_rest[spcidx], a_min=1e-16, a_max=None))\
-np.log10(np.clip(melsp_rest[spcidx], a_min=1e-16, a_max=None))))**2, axis=-1))
lsd_mean_cyc = np.mean(lsd_arr)
lsd_std_cyc = np.std(lsd_arr)
logging.info("lsd_cyc: %.6f dB +- %.6f" %
(lsd_mean_cyc, lsd_std_cyc))
logging.info('org f0')
logging.info(f0[10:15])
logging.info('rec f0')
logging.info(cvf0_src[10:15])
logging.info('cyc f0')
logging.info(cvf0_cyc[10:15])
logging.info('org cap')
logging.info(codeap[10:15])
logging.info('rec cap')
logging.info(cvcodeap_src[10:15])
logging.info('cyc cap')
logging.info(cvcodeap_cyc[10:15])
dataset = feat_file.split('/')[1].split('_')[0]
if 'tr' in dataset:
logging.info('trn')
f0rmse_cvlist.append(f0_rmse)
f0corr_cvlist.append(f0_corr)
caprmse_cvlist.append(codeap_rmse)
lsd_cvlist.append(lsd_mean)
lsdstd_cvlist.append(lsd_std)
cvlist.append(np.var(melsp_src_rest, axis=0))
logging.info(len(cvlist))
f0rmse_cvlist_cyc.append(f0_rmse_cyc)
f0corr_cvlist_cyc.append(f0_corr_cyc)
caprmse_cvlist_cyc.append(codeap_rmse_cyc)
lsd_cvlist_cyc.append(lsd_mean_cyc)
lsdstd_cvlist_cyc.append(lsd_std_cyc)
cvlist_cyc.append(np.var(melsp_cyc_rest, axis=0))
elif 'dv' in dataset:
logging.info('dev')
f0rmse_cvlist_dv.append(f0_rmse)
f0corr_cvlist_dv.append(f0_corr)
caprmse_cvlist_dv.append(codeap_rmse)
lsd_cvlist_dv.append(lsd_mean)
lsdstd_cvlist_dv.append(lsd_std)
cvlist_dv.append(np.var(melsp_src_rest, axis=0))
logging.info(len(cvlist_dv))
f0rmse_cvlist_cyc_dv.append(f0_rmse_cyc)
f0corr_cvlist_cyc_dv.append(f0_corr_cyc)
caprmse_cvlist_cyc_dv.append(codeap_rmse_cyc)
lsd_cvlist_cyc_dv.append(lsd_mean_cyc)
lsdstd_cvlist_cyc_dv.append(lsd_std_cyc)
cvlist_cyc_dv.append(np.var(melsp_cyc_rest, axis=0))
logging.info('write rec to h5')
outh5dir = os.path.join(
os.path.dirname(os.path.dirname(feat_file)),
args.spk + "-" + args.spk)
if not os.path.exists(outh5dir):
os.makedirs(outh5dir)
feat_file = os.path.join(outh5dir, os.path.basename(feat_file))
logging.info(feat_file + ' ' + args.string_path)
logging.info(feat_rec.shape)
write_hdf5(feat_file, args.string_path, feat_rec)
logging.info('write cyc to h5')
outh5dir = os.path.join(
os.path.dirname(os.path.dirname(feat_file)),
args.spk + "-" + args.spk + "-" + args.spk)
if not os.path.exists(outh5dir):
os.makedirs(outh5dir)
feat_file = os.path.join(outh5dir, os.path.basename(feat_file))
logging.info(feat_file + ' ' + args.string_path)
logging.info(feat_cyc.shape)
write_hdf5(feat_file, args.string_path, feat_cyc)
count += 1
def main():
parser = argparse.ArgumentParser()
# decode setting
parser.add_argument("--model",
required=True,
type=str,
help="GRU_RNN model file")
parser.add_argument("--config",
required=True,
type=str,
help="GRU_RNN configure file")
parser.add_argument("--outdir",
required=True,
type=str,
help="directory to save generated samples")
# other setting
#parser.add_argument("--GPU_device", default=None,
# type=int, help="selection of GPU device")
#parser.add_argument("--GPU_device_str", default=None,
# type=str, help="selection of GPU device")
parser.add_argument("--verbose",
default=VERBOSE,
type=int,
help="log level")
args = parser.parse_args()
#if args.GPU_device is not None or args.GPU_device_str is not None:
# os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
# if args.GPU_device_str is None:
# os.environ["CUDA_VISIBLE_DEVICES"] = str(args.GPU_device)
# else:
# os.environ["CUDA_VISIBLE_DEVICES"] = args.GPU_device_str
os.environ["CUDA_VISIBLE_DEVICES"] = ""
# check directory existence
if not os.path.exists(args.outdir):
os.makedirs(args.outdir)
# set log level
if args.verbose > 0:
logging.basicConfig(
level=logging.INFO,
format=
'%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s',
datefmt='%m/%d/%Y %I:%M:%S',
filemode='w',
filename=args.outdir + "/decode.log")
logging.getLogger().addHandler(logging.StreamHandler())
elif args.verbose > 1:
logging.basicConfig(
level=logging.DEBUG,
format=
'%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s',
datefmt='%m/%d/%Y %I:%M:%S',
filemode='w',
filename=args.outdir + "/decode.log")
logging.getLogger().addHandler(logging.StreamHandler())
else:
logging.basicConfig(
level=logging.WARN,
format=
'%(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s',
datefmt='%m/%d/%Y %I:%M:%S',
filemode='w',
filename=args.outdir + "/decode.log")
logging.getLogger().addHandler(logging.StreamHandler())
logging.warn("logging is disabled.")
# load config
config = torch.load(args.config)
spk_list = config.spk_list.split('@')
n_spk = len(spk_list)
model_epoch = os.path.basename(args.model).split('.')[0].split('-')[1]
logging.info('epoch: ' + model_epoch)
device = torch.device("cpu")
#with torch.cuda.device(0):
# define model and load parameters
with torch.no_grad():
model_decoder_mcep = GRU_SPEC_DECODER(
feat_dim=config.lat_dim,
out_dim=config.mcep_dim,
n_spk=n_spk,
hidden_layers=config.hidden_layers_dec,
hidden_units=config.hidden_units_dec,
kernel_size=config.kernel_size_dec,
dilation_size=config.dilation_size_dec,
causal_conv=config.causal_conv_dec,
bi=config.bi_dec,
spkidtr_dim=config.spkidtr_dim,
pad_first=True,
ar=config.ar_dec)
logging.info(model_decoder_mcep)
model_decoder_excit = GRU_EXCIT_DECODER(
feat_dim=config.lat_dim,
cap_dim=config.cap_dim,
n_spk=n_spk,
hidden_layers=config.hidden_layers_dec,
hidden_units=config.hidden_units_dec,
kernel_size=config.kernel_size_dec,
dilation_size=config.dilation_size_dec,
causal_conv=config.causal_conv_dec,
bi=config.bi_dec,
spkidtr_dim=config.spkidtr_dim,
pad_first=True,
ar=config.ar_dec)
logging.info(model_decoder_excit)
model_decoder_mcep.load_state_dict(
torch.load(args.model, map_location=device)["model_decoder_mcep"])
model_decoder_excit.load_state_dict(
torch.load(args.model, map_location=device)["model_decoder_excit"])
#model_decoder_mcep.cuda()
#model_decoder_excit.cuda()
model_decoder_mcep.eval()
model_decoder_excit.eval()
for param in model_decoder_mcep.parameters():
param.requires_grad = False
for param in model_decoder_excit.parameters():
param.requires_grad = False
#feat = torch.LongTensor(np.arange(n_spk)).cuda().unsqueeze(0)
feat = torch.LongTensor(np.arange(n_spk)).unsqueeze(0)
logging.info(feat)
logging.info(spk_list)
colormap = np.array(['b', 'r'])
male = ['bdl', 'p237', 'p245', 'p251', 'p252', 'p259', 'p274', 'p304', 'p311', 'p326', 'p345', 'p360', 'p363', \
'SEM1', 'SEM2', 'TFM1', 'TGM1', 'TMM1', 'TEM1', 'TEM2', \
'VCC2SM1', 'VCC2SM2', 'VCC2SM3', 'VCC2TM1', 'VCC2TM2', 'VCC2SM4']
female = ['slt', 'p231', 'p238', 'p248', 'p253', 'p264', 'p265', 'p266', 'p276', 'p305', 'p308', 'p318', 'p335', \
'SEF1', 'SEF2', 'TEF1', 'TEF2', 'TFF1', 'TGF1', 'TMF1', \
'VCC2SF1', 'VCC2SF2', 'VCC2SF3', 'VCC2TF1', 'VCC2TF2', 'VCC2SF4']
gender = []
for i in range(len(spk_list)):
if spk_list[i] in male:
gender.append(0)
elif spk_list[i] in female:
gender.append(1)
else:
logging.info('error %s not in gender list' % (spk_list[i]))
exit()
z = model_decoder_mcep.spkidtr_conv(
F.one_hot(feat,
num_classes=n_spk).float().transpose(1,
2)).transpose(1, 2)
#z_rec = model_decoder_mcep.spkidtr_deconv(z.transpose(1,2)).transpose(1,2)
logging.info(z)
logging.info(args.outdir)
#plt.rcParams["figure.figsize"] = (20,11.25) #1920x1080
plt.rcParams["figure.figsize"] = (11.25, 11.25) #1080x1080
#plt.rcParams["figure.figsize"] = (14.229166667,14.229166667) #1366x1366
#z = z.cpu().data.numpy()
z = z.data.numpy()
logging.info(z.shape)
x = z[0, :, 0]
y = z[0, :, 1]
fig, ax = plt.subplots()
ax.scatter(x, y, s=40, c=colormap[gender])
for i, txt in enumerate(spk_list):
ax.annotate(txt, (x[i], y[i]))
plt.savefig(os.path.join(args.outdir, 'spect.png'))
plt.close()
z_e = model_decoder_excit.spkidtr_conv(
F.one_hot(feat,
num_classes=n_spk).float().transpose(1,
2)).transpose(1, 2)
#z_e_rec = model_decoder_excit.spkidtr_deconv(z_e.transpose(1,2)).transpose(1,2)
logging.info(z_e)
#z_e = z_e.cpu().data.numpy()
z_e = z_e.data.numpy()
x = z_e[0, :, 0]
y = z_e[0, :, 1]
fig, ax = plt.subplots()
ax.scatter(x, y, s=40, c=colormap[gender])
for i, txt in enumerate(spk_list):
ax.annotate(txt, (x[i], y[i]))
plt.savefig(os.path.join(args.outdir, 'excit.png'))
plt.close()
def decode_RNN(feat_list,
gpu,
cvlist=None,
lsd_cvlist_src=None,
lsdstd_cvlist_src=None,
lsd_cvlist_cyc=None,
lsdstd_cvlist_cyc=None,
lsd_cvlist=None,
lsdstd_cvlist=None,
lat_dist_rmse_list=None,
lat_dist_cosim_list=None):
with torch.cuda.device(gpu):
# define model and load parameters
with torch.no_grad():
model_encoder_melsp = GRU_VAE_ENCODER(
in_dim=config.mel_dim,
n_spk=n_spk,
lat_dim=config.lat_dim,
hidden_layers=config.hidden_layers_enc,
hidden_units=config.hidden_units_enc,
kernel_size=config.kernel_size_enc,
dilation_size=config.dilation_size_enc,
causal_conv=config.causal_conv_enc,
pad_first=True,
right_size=config.right_size_enc)
logging.info(model_encoder_melsp)
model_decoder_melsp = GRU_SPEC_DECODER(
feat_dim=config.lat_dim + config.lat_dim_e,
out_dim=config.mel_dim,
n_spk=(config.emb_spk_dim // config.n_weight_emb) *
config.n_weight_emb,
hidden_layers=config.hidden_layers_dec,
hidden_units=config.hidden_units_dec,
kernel_size=config.kernel_size_dec,
dilation_size=config.dilation_size_dec,
causal_conv=config.causal_conv_dec,
pad_first=True,
right_size=config.right_size_dec,
red_dim_upd=config.mel_dim,
pdf_gauss=True)
logging.info(model_decoder_melsp)
model_encoder_excit = GRU_VAE_ENCODER(
in_dim=config.mel_dim,
n_spk=n_spk,
lat_dim=config.lat_dim_e,
hidden_layers=config.hidden_layers_enc,
hidden_units=config.hidden_units_enc,
kernel_size=config.kernel_size_enc,
dilation_size=config.dilation_size_enc,
causal_conv=config.causal_conv_enc,
pad_first=True,
right_size=config.right_size_enc)
logging.info(model_encoder_excit)
model_spkidtr = SPKID_TRANSFORM_LAYER(
n_spk=n_spk,
emb_dim=config.emb_spk_dim,
n_weight_emb=config.n_weight_emb,
conv_emb_flag=True,
spkidtr_dim=config.spkidtr_dim)
logging.info(model_spkidtr)
model_encoder_melsp.load_state_dict(
torch.load(args.model)["model_encoder_melsp"])
model_decoder_melsp.load_state_dict(
torch.load(args.model)["model_decoder_melsp"])
model_encoder_excit.load_state_dict(
torch.load(args.model)["model_encoder_excit"])
model_spkidtr.load_state_dict(
torch.load(args.model)["model_spkidtr"])
model_encoder_melsp.cuda()
model_decoder_melsp.cuda()
model_encoder_excit.cuda()
model_spkidtr.cuda()
model_encoder_melsp.eval()
model_decoder_melsp.eval()
model_encoder_excit.eval()
model_spkidtr.eval()
model_encoder_melsp.remove_weight_norm()
model_decoder_melsp.remove_weight_norm()
model_encoder_excit.remove_weight_norm()
model_spkidtr.remove_weight_norm()
for param in model_encoder_melsp.parameters():
param.requires_grad = False
for param in model_decoder_melsp.parameters():
param.requires_grad = False
for param in model_encoder_excit.parameters():
param.requires_grad = False
for param in model_spkidtr.parameters():
param.requires_grad = False
count = 0
pad_left = (model_encoder_melsp.pad_left +
model_decoder_melsp.pad_left) * 2
pad_right = (model_encoder_melsp.pad_right +
model_decoder_melsp.pad_right) * 2
outpad_lefts = [None] * 3
outpad_rights = [None] * 3
outpad_lefts[0] = pad_left - model_encoder_melsp.pad_left
outpad_rights[0] = pad_right - model_encoder_melsp.pad_right
outpad_lefts[1] = outpad_lefts[0] - model_decoder_melsp.pad_left
outpad_rights[1] = outpad_rights[0] - model_decoder_melsp.pad_right
outpad_lefts[2] = outpad_lefts[1] - model_encoder_melsp.pad_left
outpad_rights[2] = outpad_rights[1] - model_encoder_melsp.pad_right
melfb_t = np.linalg.pinv(
librosa.filters.mel(args.fs, args.fftl, n_mels=config.mel_dim))
temp = 0.675
logging.info(f'temp: {temp}')
for feat_file in feat_list:
# convert melsp
spk_src = os.path.basename(os.path.dirname(feat_file))
logging.info('%s --> %s' % (spk_src, args.spk_trg))
file_trg = os.path.join(
os.path.dirname(os.path.dirname(feat_file)), args.spk_trg,
os.path.basename(feat_file))
trg_exist = False
if os.path.exists(file_trg):
logging.info('exist: %s' % (file_trg))
feat_trg = read_hdf5(file_trg, "/log_1pmelmagsp")
logging.info(feat_trg.shape)
trg_exist = True
feat_org = read_hdf5(feat_file, "/log_1pmelmagsp")
logging.info(feat_org.shape)
logging.info("generate")
with torch.no_grad():
feat = F.pad(
torch.FloatTensor(feat_org).cuda().unsqueeze(
0).transpose(1, 2), (pad_left, pad_right),
"replicate").transpose(1, 2)
spk_logits, _, lat_src, _ = model_encoder_melsp(
feat, sampling=False)
spk_logits_e, _, lat_src_e, _ = model_encoder_excit(
feat, sampling=False)
logging.info('input spkpost')
if outpad_rights[0] > 0:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[0]:
-outpad_rights[0]],
dim=-1), 1))
else:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[0]:],
dim=-1), 1))
logging.info('input spkpost_e')
if outpad_rights[0] > 0:
logging.info(
torch.mean(
F.softmax(spk_logits_e[:, outpad_lefts[0]:
-outpad_rights[0]],
dim=-1), 1))
else:
logging.info(
torch.mean(
F.softmax(spk_logits_e[:, outpad_lefts[0]:],
dim=-1), 1))
if trg_exist:
spk_trg_logits, _, lat_trg, _ = model_encoder_melsp(
F.pad(
torch.FloatTensor(feat_trg).cuda().unsqueeze(
0).transpose(1, 2),
(model_encoder_melsp.pad_left,
model_encoder_melsp.pad_right),
"replicate").transpose(1, 2),
sampling=False)
spk_trg_logits_e, _, lat_trg_e, _ = model_encoder_excit(
F.pad(
torch.FloatTensor(feat_trg).cuda().unsqueeze(
0).transpose(1, 2),
(model_encoder_excit.pad_left,
model_encoder_excit.pad_right),
"replicate").transpose(1, 2),
sampling=False)
logging.info('target spkpost')
logging.info(
torch.mean(F.softmax(spk_trg_logits, dim=-1), 1))
logging.info('target spkpost_e')
logging.info(
torch.mean(F.softmax(spk_trg_logits_e, dim=-1), 1))
_, src_code = model_spkidtr((torch.ones(
(1, lat_src_e.shape[1])) * src_idx).cuda().long())
_, trg_code = model_spkidtr((torch.ones(
(1, lat_src_e.shape[1])) * trg_idx).cuda().long())
lat_cat = torch.cat((lat_src_e, lat_src), 2)
_, cvmelsp_src, _ = model_decoder_melsp(lat_cat,
y=src_code,
temp=temp)
_, cvmelsp, _ = model_decoder_melsp(lat_cat,
y=trg_code,
temp=temp)
trj_lat_cat = lat_cat_src = lat_cat
spk_logits, _, lat_cv, _ = model_encoder_melsp(
cvmelsp, sampling=False)
spk_logits_e, _, lat_cv_e, _ = model_encoder_excit(
cvmelsp, sampling=False)
logging.info('cv spkpost')
if outpad_rights[2] > 0:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[2]:
-outpad_rights[2]],
dim=-1), 1))
else:
logging.info(
torch.mean(
F.softmax(spk_logits[:, outpad_lefts[2]:],
dim=-1), 1))
logging.info('cv spkpost_e')
if outpad_rights[2] > 0:
logging.info(
torch.mean(
F.softmax(spk_logits_e[:, outpad_lefts[2]:
-outpad_rights[2]],
dim=-1), 1))
else:
logging.info(
torch.mean(
F.softmax(spk_logits_e[:, outpad_lefts[2]:],
dim=-1), 1))
_, src_code = model_spkidtr((torch.ones(
(1, lat_cv_e.shape[1])) * src_idx).cuda().long())
lat_cat = torch.cat((lat_cv_e, lat_cv), 2)
_, cvmelsp_cyc, _ = model_decoder_melsp(lat_cat,
y=src_code,
temp=temp)
#if outpad_rights[0] > 0:
# trj_lat_cat = trj_lat_cat[:,outpad_lefts[0]:-outpad_rights[0]]
#else:
# trj_lat_cat = trj_lat_cat[:,outpad_lefts[0]:-outpad_rights[0]]
if outpad_rights[1] > 0:
cvmelsp_src = cvmelsp_src[:, outpad_lefts[1]:
-outpad_rights[1]]
cvmelsp = cvmelsp[:, outpad_lefts[1]:-outpad_rights[1]]
else:
cvmelsp_src = cvmelsp_src[:, outpad_lefts[1]:]
cvmelsp = cvmelsp[:, outpad_lefts[1]:]
feat_cv = cvmelsp[0].cpu().data.numpy()
#feat_lat = trj_lat_cat[0].cpu().data.numpy()
cvmelsp_src = np.array(cvmelsp_src[0].cpu().data.numpy(),
dtype=np.float64)
cvmelsp = np.array(cvmelsp[0].cpu().data.numpy(),
dtype=np.float64)
cvmelsp_cyc = np.array(cvmelsp_cyc[0].cpu().data.numpy(),
dtype=np.float64)
if trg_exist:
if outpad_rights[1] > 0:
lat_src = lat_cat_src[:, outpad_lefts[0]:
-outpad_rights[0]]
else:
lat_src = lat_cat_src[:, outpad_lefts[0]:]
lat_trg = torch.cat((lat_trg_e, lat_trg), 2)
logging.info(cvmelsp_src.shape)
logging.info(cvmelsp.shape)
logging.info(cvmelsp_cyc.shape)
melsp = np.array(feat_org)
if trg_exist:
logging.info(lat_src.shape)
logging.info(lat_trg.shape)
melsp_trg = np.array(feat_trg)
spcidx = np.array(read_hdf5(feat_file, "/spcidx_range")[0])
melsp_rest = (np.exp(melsp) - 1) / 10000
melsp_cv_rest = (np.exp(cvmelsp) - 1) / 10000
melsp_src_rest = (np.exp(cvmelsp_src) - 1) / 10000
melsp_cyc_rest = (np.exp(cvmelsp_cyc) - 1) / 10000
cvlist.append(np.var(melsp_cv_rest, axis=0))
lsd_arr = np.sqrt(np.mean((20*(np.log10(np.clip(melsp_src_rest[spcidx], a_min=1e-16, a_max=None))\
-np.log10(np.clip(melsp_rest[spcidx], a_min=1e-16, a_max=None))))**2, axis=-1))
lsd_mean = np.mean(lsd_arr)
lsd_std = np.std(lsd_arr)
logging.info("lsd_src_cv: %.6f dB +- %.6f" %
(lsd_mean, lsd_std))
lsd_cvlist_src.append(lsd_mean)
lsdstd_cvlist_src.append(lsd_std)
if trg_exist:
melsp_trg_rest = (np.exp(melsp_trg) - 1) / 10000
spcidx_trg = np.array(
read_hdf5(file_trg, "/spcidx_range")[0])
_, twf_melsp, _, _ = dtw.dtw_org_to_trg(np.array(melsp_cv_rest[spcidx], \
dtype=np.float64), np.array(melsp_trg_rest[spcidx_trg], dtype=np.float64), mcd=-1)
twf_melsp = np.array(twf_melsp[:, 0])
lsd_arr = np.sqrt(np.mean((20*(np.log10(np.clip(melsp_cv_rest[twf_melsp], a_min=1e-16, a_max=None))\
-np.log10(np.clip(melsp_rest[twf_melsp], a_min=1e-16, a_max=None))))**2, axis=-1))
lsd_mean = np.mean(lsd_arr)
lsd_std = np.std(lsd_arr)
logging.info("lsd_trg: %.6f dB +- %.6f" %
(lsd_mean, lsd_std))
lsd_cvlist.append(lsd_mean)
lsdstd_cvlist.append(lsd_std)
spcidx_src = torch.LongTensor(spcidx).cuda()
spcidx_trg = torch.LongTensor(spcidx_trg).cuda()
trj_lat_src = np.array(torch.index_select(
lat_src[0], 0, spcidx_src).cpu().data.numpy(),
dtype=np.float64)
trj_lat_trg = np.array(torch.index_select(
lat_trg[0], 0, spcidx_trg).cpu().data.numpy(),
dtype=np.float64)
aligned_lat_srctrg, _, _, _ = dtw.dtw_org_to_trg(
trj_lat_src, trj_lat_trg)
lat_dist_srctrg = np.mean(
np.sqrt(
np.mean((aligned_lat_srctrg - trj_lat_trg)**2,
axis=0)))
_, _, lat_cdist_srctrg, _ = dtw.dtw_org_to_trg(trj_lat_trg,
trj_lat_src,
mcd=0)
aligned_lat_trgsrc, _, _, _ = dtw.dtw_org_to_trg(
trj_lat_trg, trj_lat_src)
lat_dist_trgsrc = np.mean(
np.sqrt(
np.mean((aligned_lat_trgsrc - trj_lat_src)**2,
axis=0)))
_, _, lat_cdist_trgsrc, _ = dtw.dtw_org_to_trg(trj_lat_src,
trj_lat_trg,
mcd=0)
logging.info("%lf %lf %lf %lf" %
(lat_dist_srctrg, lat_cdist_srctrg,
lat_dist_trgsrc, lat_cdist_trgsrc))
lat_dist_rmse = (lat_dist_srctrg + lat_dist_trgsrc) / 2
lat_dist_cosim = (lat_cdist_srctrg + lat_cdist_trgsrc) / 2
lat_dist_rmse_list.append(lat_dist_rmse)
lat_dist_cosim_list.append(lat_dist_cosim)
logging.info("lat_dist: %.6f %.6f" %
(lat_dist_rmse, lat_dist_cosim))
lsd_arr = np.sqrt(np.mean((20*(np.log10(np.clip(melsp_cyc_rest[spcidx], a_min=1e-16, a_max=None))\
-np.log10(np.clip(melsp_rest[spcidx], a_min=1e-16, a_max=None))))**2, axis=-1))
lsd_mean_cyc = np.mean(lsd_arr)
lsd_std_cyc = np.std(lsd_arr)
logging.info("lsd_cyc: %.6f dB +- %.6f" %
(lsd_mean_cyc, lsd_std_cyc))
lsd_cvlist_cyc.append(lsd_mean_cyc)
lsdstd_cvlist_cyc.append(lsd_std_cyc)
logging.info("synth anasyn")
magsp = np.matmul(melfb_t, melsp_rest.T)
logging.info(magsp.shape)
hop_length = int((args.fs / 1000) * args.shiftms)
win_length = int((args.fs / 1000) * args.winms)
wav = np.clip(
librosa.core.griffinlim(magsp,
hop_length=hop_length,
win_length=win_length,
window='hann'), -1,
0.999969482421875)
wavpath = os.path.join(
args.outdir,
os.path.basename(feat_file).replace(".h5", "_anasyn.wav"))
logging.info(wavpath)
sf.write(wavpath, wav, args.fs, 'PCM_16')
#if trg_exist:
# logging.info("synth anasyn_trg")
# wav = np.clip(pw.synthesize(f0_trg, sp_trg, ap_trg, fs, frame_period=args.shiftms), -1, 1)
# wavpath = os.path.join(args.outdir,os.path.basename(feat_file).replace(".h5","_anasyn_trg.wav"))
# sf.write(wavpath, wav, fs, 'PCM_16')
# logging.info(wavpath)
logging.info("synth gf rec")
recmagsp = np.matmul(melfb_t, melsp_src_rest.T)
logging.info(recmagsp.shape)
wav = np.clip(
librosa.core.griffinlim(recmagsp,
hop_length=hop_length,
win_length=win_length,
window='hann'), -1,
0.999969482421875)
wavpath = os.path.join(
args.outdir,
os.path.basename(feat_file).replace(".h5", "_rec.wav"))
logging.info(wavpath)
sf.write(wavpath, wav, args.fs, 'PCM_16')
logging.info("synth gf cv")
cvmagsp = np.matmul(melfb_t, melsp_cv_rest.T)
logging.info(cvmagsp.shape)
wav = np.clip(
librosa.core.griffinlim(cvmagsp,
hop_length=hop_length,
win_length=win_length,
window='hann'), -1,
0.999969482421875)
wavpath = os.path.join(
args.outdir,
os.path.basename(feat_file).replace(".h5", "_cv.wav"))
logging.info(wavpath)
sf.write(wavpath, wav, args.fs, 'PCM_16')
logging.info('write to h5')
outh5dir = os.path.join(
os.path.dirname(os.path.dirname(feat_file)),
spk_src + "-" + args.spk_trg)
if not os.path.exists(outh5dir):
os.makedirs(outh5dir)
feat_file = os.path.join(outh5dir, os.path.basename(feat_file))
# cv
write_path = args.string_path
logging.info(feat_file + ' ' + write_path)
logging.info(feat_cv.shape)
write_hdf5(feat_file, write_path, feat_cv)
#logging.info('write lat to h5')
#logging.info(feat_file + ' ' + args.string_path+'_lat')
#logging.info(feat_lat.shape)
#write_hdf5(feat_file, args.string_path+'_lat', feat_lat)
count += 1