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 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 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()