def embedding_computation_loop(split, set_loader, stat_file):
"""Extracts embeddings for a given dataset loader
"""
# Extract embeddings (skip if already done)
if not os.path.isfile(stat_file):
logger.debug("Extracting deep embeddings and diarizing")
embeddings = np.empty(shape=[0, params["emb_dim"]], dtype=np.float64)
modelset = []
segset = []
# Different data may have different statistics
params["mean_var_norm_emb"].count = 0
for batch in set_loader: # t:
ids = batch.id
wavs, lens = batch.sig
mod = [x for x in ids]
seg = [x for x in ids]
modelset = modelset + mod
segset = segset + seg
# Embedding computation
emb = (
compute_embeddings(wavs, lens)
.contiguous()
.squeeze(1)
.cpu()
.numpy()
)
embeddings = np.concatenate((embeddings, emb), axis=0)
modelset = np.array(modelset, dtype="|O")
segset = np.array(segset, dtype="|O")
# Intialize variables for start, stop and stat0
s = np.array([None] * embeddings.shape[0])
b = np.array([[1.0]] * embeddings.shape[0])
stat_obj = StatObject_SB(
modelset=modelset,
segset=segset,
start=s,
stop=s,
stat0=b,
stat1=embeddings,
)
logger.debug("Saving Embeddings...")
stat_obj.save_stat_object(stat_file)
else:
logger.debug("Skipping embedding extraction (as already present)")
logger.debug("Loading previously saved embeddings")
with open(stat_file, "rb") as in_file:
stat_obj = pickle.load(in_file)
return stat_obj
def emb_computation_loop(split, set_loader, stat_file):
"""Computes the embeddings and saves the in a stat file"""
# Extract embeddings (skip if already done)
if not os.path.isfile(stat_file):
embeddings = numpy.empty(shape=[0, params["emb_dim"]],
dtype=numpy.float64)
modelset = []
segset = []
with tqdm(set_loader, dynamic_ncols=True) as t:
for batch in t:
ids = batch.id
wavs, lens = batch.sig
mod = [x for x in ids]
seg = [x for x in ids]
modelset = modelset + mod
segset = segset + seg
# Enrolment and test embeddings
embs = compute_embeddings(wavs, lens)
xv = embs.squeeze().cpu().numpy()
embeddings = numpy.concatenate((embeddings, xv), axis=0)
modelset = numpy.array(modelset, dtype="|O")
segset = numpy.array(segset, dtype="|O")
# Intialize variables for start, stop and stat0
s = numpy.array([None] * embeddings.shape[0])
b = numpy.array([[1.0]] * embeddings.shape[0])
# Stat object (used to collect embeddings)
stat_obj = StatObject_SB(
modelset=modelset,
segset=segset,
start=s,
stop=s,
stat0=b,
stat1=embeddings,
)
logger.info(f"Saving stat obj for {split}")
stat_obj.save_stat_object(stat_file)
else:
logger.info(f"Skipping embedding Extraction for {split}")
logger.info(f"Loading previously saved stat_object for {split}")
with open(stat_file, "rb") as input:
stat_obj = pickle.load(input)
return stat_obj
emb = compute_embeddings(wav, lens)
xv = emb.squeeze(1).cpu().numpy()
embeddings = numpy.concatenate((embeddings, xv), axis=0)
# Speaker IDs and utterance IDs
modelset = numpy.array(modelset, dtype="|O")
segset = numpy.array(segset, dtype="|O")
# Intialize variables for start, stop and stat0
s = numpy.array([None] * embeddings.shape[0])
b = numpy.array([[1.0]] * embeddings.shape[0])
embeddings_stat = StatObject_SB(
modelset=modelset,
segset=segset,
start=s,
stop=s,
stat0=b,
stat1=embeddings,
)
del embeddings
# Save TRAINING embeddings in StatObject_SB object
embeddings_stat.save_stat_object(xv_file)
else:
# Load the saved stat object for train embedding
logger.info("Skipping embedding Extraction for training set")
logger.info(
"Loading previously saved stat_object for train embeddings..")
with open(xv_file, "rb") as input: