def _load_nnet(self, cpt_dir):
nnet_conf = load_json(cpt_dir, "mdl.json")
nnet = ConvTasNet(**nnet_conf)
cpt_fname = os.path.join(cpt_dir, "best.pt.tar")
cpt = th.load(cpt_fname, map_location="cpu")
nnet.load_state_dict(cpt["model_state_dict"])
logger.info("Load checkpoint from {}, epoch {:d}".format(
cpt_fname, cpt["epoch"]))
return nnet
def _load_nnet(self, cpt_dir, online=False, init_dump=False):
nnet_conf = load_json(cpt_dir, "mdl.json")
nnet = TCNSENet(**nnet_conf, online=online, init_dump=init_dump)
cpt_fname = os.path.join(cpt_dir, "best.pt.tar")
cpt = th.load(cpt_fname, map_location="cpu")
nnet.load_state_dict(cpt["model_state_dict"])
logger.info("Load checkpoint from {}, epoch {:d}".format(
cpt_fname, cpt["epoch"]))
return nnet
def run(args):
start = time.time()
logger = get_logger(
os.path.join(args.checkpoint, 'separate.log'), file=True)
dataset = Dataset(mix_scp=args.mix_scp, ref_scp=args.ref_scp, aux_scp=args.aux_scp)
# Load model
nnet_conf = load_json(args.checkpoint, "mdl.json")
nnet = ConvTasNet(**nnet_conf)
cpt_fname = os.path.join(args.checkpoint, "best.pt.tar")
cpt = th.load(cpt_fname, map_location="cpu")
nnet.load_state_dict(cpt["model_state_dict"])
logger.info("Load checkpoint from {}, epoch {:d}".format(
cpt_fname, cpt["epoch"]))
device = th.device(
"cuda:{}".format(args.gpuid)) if args.gpuid >= 0 else th.device("cpu")
nnet = nnet.to(device) if args.gpuid >= 0 else nnet
nnet.eval()
with th.no_grad():
total_cnt = 0
for i, data in enumerate(dataset):
mix = th.tensor(data['mix'], dtype=th.float32, device=device)
aux = th.tensor(data['aux'], dtype=th.float32, device=device)
aux_len = th.tensor(data['aux_len'], dtype=th.float32, device=device)
key = data['key']
if args.gpuid >= 0:
mix = mix.cuda()
aux = aux.cuda()
aux_len = aux_len.cuda()
# Forward
ests = nnet(mix, aux, aux_len)
ests = ests.cpu().numpy()
norm = np.linalg.norm(mix.cpu().numpy(), np.inf)
ests = ests[:mix.shape[-1]]
# for each utts
logger.info("Separate Utt{:d}".format(total_cnt + 1))
# norm
ests = ests * norm / np.max(np.abs(ests))
write_wav(os.path.join(args.dump_dir, key),
ests,
fs=args.fs)
total_cnt += 1
break
end = time.time()
logger.info('Utt={:d} | Time Elapsed: {:.1f}s'.format(total_cnt, end-start))
def run(args):
computer = NnetComputer(args.checkpoint, args.gpu)
num_done = 0
feats_conf = load_json(args.checkpoint, "feats.json")
spectra = Processor(args.spectra, **feats_conf)
spatial = ScriptReader(args.spatial) if args.spatial else None
dump_dir = Path(args.dump_dir)
dump_dir.mkdir(exist_ok=True, parents=True)
for key, feats in spectra:
logger.info("Compute on utterance {}...".format(key))
if spatial:
spa = spatial[key]
feats = np.hstack([feats, spa])
spk_masks = computer.compute(feats)
for i, m in enumerate(spk_masks):
(dump_dir / f"spk{i + 1:d}").mkdir(exist_ok=True)
np.save(dump_dir / f"spk{i + 1:d}" / key, m)
num_done += 1
logger.info("Compute over {:d} utterances".format(num_done))
def run(args):
computer = NnetComputer(args.checkpoint, args.gpu)
num_done = 0
feats_conf = load_json(args.checkpoint, "feats.json")
spectra = Processor(args.spectra, **feats_conf)
spatial = ScriptReader(args.spatial) if args.spatial else None
for key, feats in spectra:
logger.info("Compute on utterance {}...".format(key))
if spatial:
spa = spatial[key]
feats = np.hstack([feats, spa])
spk_masks = computer.compute(feats)
for i, m in enumerate(spk_masks):
fdir = os.path.join(args.dump_dir, "spk{:d}".format(i + 1))
make_dir(fdir)
np.save(os.path.join(fdir, key), m)
num_done += 1
logger.info("Compute over {:d} utterances".format(num_done))
def evaluate(args):
start = time.time()
total_SISNR = 0
total_SDR = 0
total_cnt = 0
# build the logger object
logger = get_logger(os.path.join(args.checkpoint, 'eval.log'), file=True)
# Load model
nnet_conf = load_json(args.checkpoint, "mdl.json")
nnet = ConvTasNet(**nnet_conf)
cpt_fname = os.path.join(args.checkpoint, "best.pt.tar")
cpt = th.load(cpt_fname, map_location="cpu")
nnet.load_state_dict(cpt["model_state_dict"])
logger.info("Load checkpoint from {}, epoch {:d}".format(
cpt_fname, cpt["epoch"]))
device = th.device("cuda:{}".format(
args.gpuid)) if args.gpuid >= 0 else th.device("cpu")
nnet = nnet.to(device) if args.gpuid >= 0 else nnet
nnet.eval()
# Load data
dataset = Dataset(mix_scp=args.mix_scp,
ref_scp=args.ref_scp,
aux_scp=args.aux_scp)
with th.no_grad():
for i, data in enumerate(dataset):
mix1 = th.tensor(data['mix1'], dtype=th.float32, device=device)
mix2 = th.tensor(data['mix2'], dtype=th.float32, device=device)
aux = th.tensor(data['aux'], dtype=th.float32, device=device)
aux_len = th.tensor(data['aux_len'],
dtype=th.float32,
device=device)
if args.gpuid >= 0:
mix1 = mix1.cuda()
mix2 = mix2.cuda()
aux = aux.cuda()
aux_len = aux_len.cuda()
# Forward
ref = data['ref']
ests, _ = nnet(mix1, mix2, aux, aux_len)
ests = ests.cpu().numpy()
if ests.size != ref.size:
end = min(ests.size, ref.size)
ests = ests[:end]
ref = ref[:end]
# for each utts
# Compute SDRi
if args.cal_sdr:
SDR, sir, sar, popt = bss_eval_sources(ref, ests)
# avg_SDRi = cal_SDRi(src_ref, src_est, mix)
total_SDR += SDR[0]
# Compute SI-SNR
SISNR = cal_SISNR(ests, ref)
if args.cal_sdr:
logger.info("Utt={:d} | SDR={:.2f} | SI-SNR={:.2f}".format(
total_cnt + 1, SDR[0], SISNR))
else:
logger.info("Utt={:d} | SI-SNR={:.2f}".format(
total_cnt + 1, SISNR))
total_SISNR += SISNR
total_cnt += 1
end = time.time()
logger.info('Time Elapsed: {:.1f}s'.format(end - start))
if args.cal_sdr:
logger.info("Average SDR: {0:.2f}".format(total_SDR / total_cnt))
logger.info("Average SI-SNR: {:.2f}".format(total_SISNR / total_cnt))