def enrollment_dic_kwsTrials(self,
listfilename,
uttpath,
utt2label,
save_path,
print_interval=10,
num_eval=10,
eval_frames=None,
save_dic=False):
self.eval()
lines = []
enroll_utt = []
feats = {}
tstart = time.time()
## Read all lines
with open(listfilename) as listfile:
lines = listfile.readlines()
for line in lines:
data = line.split()
enroll_utt.append(data[0])
enroll_utt.append(data[1])
enroll_utt.append(data[2])
set_enroll_utt = list(set(enroll_utt))
set_enroll_utt.sort()
##extract enrollment data embeddings
for idx, uttid in enumerate(set_enroll_utt):
with torch.no_grad():
inp = torch.FloatTensor(loadWAV(uttpath + uttid, 0, True,
10)).cuda()
embd = self.__S__.forward(inp).cpu()
feats[uttid] = embd
telapsed = time.time() - tstart
if idx % print_interval == 0:
sys.stdout.write(
"\rReading %d of %d: %.2f Hz, embedding size %d" %
(idx, len(set_enroll_utt), idx / telapsed, embd.size()[1]))
feats_np = {}
for utt in feats:
feats_np[utt] = feats[utt].numpy()
if save_dic == True:
savenpy_path = save_path
numpy.save(savenpy_path, feats_np)
end = time.time() - tstart
print("\n total time %.2f" % (end))
return feats_np
def get_embeddings(model, test_list, test_path, max_frames=1000):
"""2D scatter plot
Parameters
----------
model : torch.nn.Module
The loaded model.
test_list : str
The list of wav file names.
test_path : str
The wav file location.
max_frames : int
The max number of frames to process per utterance
Returns
-------
embeddings : numpy.ndarray
Stacked speaker embeddings per utterance. Each row is an embedding.
labels : list
List of str. Corresponding speaker IDs.
"""
# TODO(alexbooth): what does num_eval do?
num_eval = 10
wav_paths = parse_test_list(test_list)
embeddings = None
labels = []
for idx, wav_path in enumerate(wav_paths):
input_ = loadWAV(os.path.join(test_path, wav_path),
max_frames,
evalmode=True,
num_eval=num_eval).to(device)
output = model.forward(input_).detach().cpu()
if embeddings == None:
embeddings = output.view(-1).unsqueeze(0)
else:
output = output.view(-1).unsqueeze(0)
embeddings = torch.cat([embeddings, output])
labels.append(wav_path.split('/')[0])
return embeddings, labels
def loadAudio(file):
audio = loadWAV(file, args.eval_frames, evalmode=True)
return torch.FloatTensor(audio)
def evaluateFromList(self,
listfilename,
print_interval=100,
test_path='',
num_eval=10,
eval_frames=None):
self.eval()
lines = []
files = []
feats = {}
tstart = time.time()
## Read all lines
with open(listfilename) as listfile:
while True:
line = listfile.readline()
if (not line):
break
data = line.split()
## Append random label if missing
if len(data) == 2: data = [random.randint(0, 1)] + data
files.append(data[1])
files.append(data[2])
lines.append(line)
setfiles = list(set(files))
setfiles.sort()
## Save all features to file
for idx, file in enumerate(setfiles):
inp1 = loadWAV(os.path.join(test_path, file),
eval_frames,
evalmode=True,
num_eval=num_eval).to(device)
ref_feat = self.__S__.forward(inp1).detach().cpu()
filename = '%06d.wav' % idx
feats[file] = ref_feat
telapsed = time.time() - tstart
if idx % print_interval == 0:
sys.stdout.write(
"\rReading %d of %d: %.2f Hz, embedding size %d" %
(idx, len(setfiles), idx / telapsed, ref_feat.size()[1]))
print('')
all_scores = []
all_labels = []
all_trials = []
tstart = time.time()
## Read files and compute all scores
for idx, line in enumerate(lines):
data = line.split()
## Append random label if missing
if len(data) == 2: data = [random.randint(0, 1)] + data
ref_feat = feats[data[1]].to(device)
com_feat = feats[data[2]].to(device)
if self.__L__.test_normalize:
ref_feat = F.normalize(ref_feat, p=2, dim=1)
com_feat = F.normalize(com_feat, p=2, dim=1)
dist = F.pairwise_distance(ref_feat.unsqueeze(-1),
com_feat.unsqueeze(-1).transpose(
0, 2)).detach().cpu().numpy()
score = -1 * numpy.mean(dist)
all_scores.append(score)
all_labels.append(int(data[0]))
all_trials.append(data[1] + " " + data[2])
if idx % print_interval == 0:
telapsed = time.time() - tstart
sys.stdout.write("\rComputing %d of %d: %.2f Hz" %
(idx, len(lines), idx / telapsed))
sys.stdout.flush()
print('\n')
return (all_scores, all_labels, all_trials)
def evaluateFromListSave(self, listfilename, print_interval=5000, feat_dir='', test_path='', num_eval=10):
self.eval();
lines = []
files = []
filedict = {}
feats = {}
tstart = time.time()
if feat_dir != '':
print('Saving temporary files to %s'%feat_dir)
if not(os.path.exists(feat_dir)):
os.makedirs(feat_dir)
## Read all lines
with open(listfilename) as listfile:
while True:
line = listfile.readline();
if (not line): # or (len(all_scores)==1000)
break;
data = line.split();
files.append(data[1])
files.append(data[2])
lines.append(line)
setfiles = list(set(files))
setfiles.sort()
## Save all features to file
for idx, file in enumerate(setfiles):
inp1 = loadWAV(os.path.join(test_path,file), self.__max_frames__, evalmode=True, num_eval=num_eval).cuda()
ref_feat = self.__S__.forward(inp1).detach().cpu()
filename = '%06d.wav'%idx
if feat_dir == '':
feats[file] = ref_feat
else:
filedict[file] = filename
torch.save(ref_feat,os.path.join(feat_dir,filename))
telapsed = time.time() - tstart
if idx % print_interval == 0:
sys.stdout.write("\rReading %d: %.2f Hz, embed size %d"%(idx,idx/telapsed,ref_feat.size()[1]));
print('')
all_scores = [];
all_labels = [];
tstart = time.time()
## Read files and compute all scores
for idx, line in enumerate(lines):
data = line.split();
if feat_dir == '':
ref_feat = feats[data[1]].cuda()
com_feat = feats[data[2]].cuda()
else:
ref_feat = torch.load(os.path.join(feat_dir,filedict[data[1]])).cuda()
com_feat = torch.load(os.path.join(feat_dir,filedict[data[2]])).cuda()
if self.__test_normalize__:
ref_feat = F.normalize(ref_feat, p=2, dim=1)
com_feat = F.normalize(com_feat, p=2, dim=1)
dist = F.pairwise_distance(ref_feat.unsqueeze(-1).expand(-1,-1,num_eval), com_feat.unsqueeze(-1).expand(-1,-1,num_eval).transpose(0,2)).detach().cpu().numpy();
score = -1 * numpy.mean(dist);
all_scores.append(score);
all_labels.append(int(data[0]));
if idx % print_interval == 0:
telapsed = time.time() - tstart
sys.stdout.write("\rComputing %d: %.2f Hz"%(idx,idx/telapsed));
sys.stdout.flush();
if feat_dir != '':
print(' Deleting temporary files.')
shutil.rmtree(feat_dir)
print('\n')
return (all_scores, all_labels);
threshold = parameter_dic['eer_threshold']
if len(lines[0].strip().split()) == 5: # on dev set
print('Deal with dev set')
tsatrt = time.time()
for line in tqdm.tqdm(lines):
data = line.strip().split()
final_labels.append(data[4])
if utt2label[data[3]] == 'non-trigger':
output_score.append('negative')
continue
else:
with torch.no_grad():
uttid = data[3] + '.wav'
if uttid not in eval_dic:
inp = torch.FloatTensor(
loadWAV(args.uttpath + uttid, 0, True, 10)).cuda()
eval_embd = s.__S__.forward(inp).cpu().numpy()
eval_dic[uttid] = eval_embd
else:
eval_embd = eval_dic[uttid]
eval_embd = numpy.squeeze(eval_embd)
enroll_embd = (enroll_dic[data[0]] + enroll_dic[data[1]] +
enroll_dic[data[2]]) / 3
enroll_embd = numpy.squeeze(enroll_embd)
result = 1 - spatial.distance.cosine(eval_embd, enroll_embd)
if result < threshold:
output_score.append('negative')
else:
output_score.append('positive')
tend = time.time() - tsatrt
print('total time: %.2f' % (tend))
tsatrt = time.time()
for line in tqdm.tqdm(lines):
# Only use 40000 lines of trial file, because the back contains stitched audio
data = line.strip().split()
final_labels.append(data[4])
if utt2label[data[3]] == 'negative':
output_score.append('negative')
continue
elif (utt2label[data[3]] == 'trigger') & (u2l_template[data[3]]
== 'positive'):
with torch.no_grad():
uttid = data[3] + '.wav'
if uttid not in eval_dic:
inp = torch.FloatTensor(
loadWAV(args.uttpath + uttid, 0, True, 10)).cuda()
eval_embd = s.__S__.forward(inp).cpu().numpy()
eval_dic[uttid] = eval_embd
else:
eval_embd = eval_dic[uttid]
eval_embd = numpy.squeeze(eval_embd)
enroll_embd = (enroll_dic[data[0]] + enroll_dic[data[1]] +
enroll_dic[data[2]]) / 3
enroll_embd = numpy.squeeze(enroll_embd)
result = 1 - spatial.distance.cosine(eval_embd, enroll_embd)
scores.append(result)
if data[4] == 'negative':
labels.append(0)
else:
labels.append(1)
output_score.append('tbd')
def evaluateFromList(self,
wav1,
wav2,
print_interval=100,
test_path='',
num_eval=10,
eval_frames=None):
self.eval()
lines = []
files = []
feats = {}
tstart = time.time()
lines = [wav1, wav2]
setfiles = list(set(lines))
setfiles.sort()
## Save all features to file
for idx, file in enumerate(setfiles):
inp1 = torch.FloatTensor(
loadWAV(os.path.join(test_path, file),
eval_frames,
evalmode=True,
num_eval=num_eval)).cuda()
ref_feat = self.__S__.forward(inp1).detach().cpu()
filename = '%06d.wav' % idx
feats[file] = ref_feat
telapsed = time.time() - tstart
if idx % print_interval == 0:
sys.stdout.write(
"\rReading %d of %d: %.2f Hz, embedding size %d" %
(idx, len(setfiles), idx / telapsed, ref_feat.size()[1]))
print('')
tstart = time.time()
## Read files and compute all scores
idx = 0
ref_feat = feats[wav1].cuda()
com_feat = feats[wav2].cuda()
if self.__L__.test_normalize:
ref_feat = F.normalize(ref_feat, p=2, dim=1)
com_feat = F.normalize(com_feat, p=2, dim=1)
dist = F.pairwise_distance(ref_feat.unsqueeze(-1),
com_feat.unsqueeze(-1).transpose(
0, 2)).detach().cpu().numpy()
score = -1 * numpy.mean(dist)
trial = wav1 + " " + wav2
if idx % print_interval == 0:
telapsed = time.time() - tstart
sys.stdout.write("\rComputing %d of %d: %.2f Hz" %
(idx, len(lines), idx / telapsed))
sys.stdout.flush()
print('\n')
return (score, trial)
def evaluateFromList(self,
listfilename,
print_interval=100,
test_path='',
num_eval=0,
eval_frames=None,
step=0.2,
save_path="./"):
self.eval()
lines = []
files = []
feats = {}
tstart = time.time()
## Read all lines
with open(listfilename) as listfile:
while True:
line = listfile.readline()
if (not line):
break
data = line.strip()
## Append random label if missing
# if len(data) == 2: data = [random.randint(0,1)] + data
files.append(data)
# files.append(data[2])
lines.append(line)
setfiles = list(set(files))
setfiles.sort()
## Save all features to file
for idx, file in enumerate(setfiles):
wavs = loadWAV(os.path.join(test_path, file + '.wav'),
eval_frames,
evalmode=True,
num_eval=num_eval,
step=step)
print('wavs size', len(wavs))
res = []
for c in chunks(wavs, 20):
c = numpy.stack(c, axis=0).astype(numpy.float)
inp1 = torch.FloatTensor(c).to(device)
ref_feat = self.__S__.forward(inp1).detach().cpu()
res.append(ref_feat)
res = torch.cat(res)
# import pickle
# pickle.dump(res, open( f"{file}.p", "wb" ) )
with open(f'{save_path}/{file}.npy', 'wb') as f:
numpy.save(f, res)
filename = '%06d.wav' % idx
feats[file] = ref_feat
telapsed = time.time() - tstart
if idx % print_interval == 0:
sys.stdout.write(
"\rReading %d of %d: %.2f Hz, embedding size %d" %
(idx, len(setfiles), idx / telapsed, ref_feat.size()[1]))
return
all_scores = []
all_labels = []
all_trials = []
tstart = time.time()
## Read files and compute all scores
for idx, line in enumerate(lines):
data = line.split()
## Append random label if missing
if len(data) == 2: data = [random.randint(0, 1)] + data
ref_feat = feats[data[1]].to(device)
com_feat = feats[data[2]].to(device)
if self.__L__.test_normalize:
ref_feat = F.normalize(ref_feat, p=2, dim=1)
com_feat = F.normalize(com_feat, p=2, dim=1)
dist = F.pairwise_distance(ref_feat.unsqueeze(-1),
com_feat.unsqueeze(-1).transpose(
0, 2)).detach().cpu().numpy()
score = -1 * numpy.mean(dist)
all_scores.append(score)
all_labels.append(int(data[0]))
all_trials.append(data[1] + " " + data[2])
if idx % print_interval == 0:
telapsed = time.time() - tstart
sys.stdout.write("\rComputing %d of %d: %.2f Hz" %
(idx, len(lines), idx / telapsed))
sys.stdout.flush()
print('\n')
return (all_scores, all_labels, all_trials)
