def recog_eval():
correct_cnt = 0
total_cnt = 0
print('Start Recognization Eval:')
for wavpath, people_valid in valid_list:
filename = os.path.join(root_dir, wavpath)
data_aud = loadWAV(filename, 160)
out_content, out_id = S.__S__.forward_aud(data_aud.cuda())
out_id = out_id.cpu().detach().numpy()[0]
magnitude = np.linalg.norm(out_id, axis=0)
out_id = out_id / magnitude
out_id = np.average(out_id, axis=1)
max_sim = None
max_arg = None
for people in register_dict.keys():
sim = cosine_similarity(out_id, register_dict[people])
if max_sim is None or max_sim < sim:
max_sim = sim
max_arg = people
total_cnt += 1
if max_arg == people_valid:
correct_cnt += 1
print('\rIter:%04d\tAcc:%.3f%%' %
(total_cnt, correct_cnt * 100 / total_cnt),
end='')
def veri_eval(threshold):
if os.path.exists('npy/match_list.npy') and os.path.exists(
'npy/unmatch_list.npy'):
match_list = np.load('npy/match_list.npy')
unmatch_list = np.load('npy/unmatch_list.npy')
else:
match_list = []
unmatch_list = []
for wavpath, people_valid in tqdm(valid_list):
filename = os.path.join(root_dir, wavpath)
data_aud = loadWAV(filename, 160)
out_content, out_id = S.__S__.forward_aud(data_aud.cuda())
out_id = out_id.cpu().detach().numpy()[0]
magnitude = np.linalg.norm(out_id, axis=0)
out_id = out_id / magnitude
out_id = np.average(out_id, axis=1)
for people in register_dict.keys():
sim = cosine_similarity(out_id, register_dict[people])
if people == people_valid:
match_list.append(sim)
else:
unmatch_list.append(sim)
np.save('npy/match_list.npy', match_list)
np.save('npy/unmatch_list.npy', unmatch_list)
tn, fp, fn, tp = 0, 0, 0, 0
for i in match_list:
if i < threshold:
fn += 1
else:
tp += 1
for i in unmatch_list:
if i < threshold:
tn += 1
else:
fp += 1
acc = tn * 100 / (tn + fp)
tpr = tp * 100 / (tp + fn)
fpr = fp * 100 / (tn + fp)
fnr = fn * 100 / (tp + fn)
# print('ACC: %.3f%% TPR: %.3f%% FPR: %.3f%% FNR: %.3f%%'%(acc, tpr, fpr, fnr))
return fpr, fnr
def evaluateFromListSave(self,
listfilename,
print_interval=10,
test_path='',
num_eval=10):
self.eval()
lines = []
files = []
filedict = {}
feats = {}
tstart = time.time()
## Read all lines
with open(listfilename) as listfile:
while True:
line = listfile.readline()
if (not line):
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__ * 4,
evalmode=True,
num_eval=num_eval).cuda()
out_a, out_A = self.__S__.forward_aud(inp1.cuda())
out_AA = torch.mean(out_A, 2)
feats[file] = out_AA.detach().cpu()
telapsed = time.time() - tstart
if idx % print_interval == 0:
sys.stdout.write("\rReading %d: %.2f Hz, embed size %d" %
(idx, idx / telapsed, out_AA.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()
ref_feat = feats[data[1]].cuda()
com_feat = feats[data[2]].cuda()
dist = F.cosine_similarity(
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 = 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()
print('\n')
return (all_scores, all_labels)
for data_pair in valid_list:
wavpath, index = data_pair
if index not in valid_dict.keys():
valid_dict[index] = [wavpath]
else:
valid_dict[index].append(wavpath)
if os.path.exists('npy/register_dict.npy'):
register_dict = np.load('npy/register_dict.npy', allow_pickle=True).item()
else:
register_dict = dict()
for people in tqdm(train_dict.keys()):
id_features = []
for wavpath in train_dict[people]:
filename = os.path.join(root_dir, wavpath)
data_aud = loadWAV(filename, 160)
out_content, out_id = S.__S__.forward_aud(data_aud.cuda())
out_id = out_id.cpu().detach().numpy()[0]
magnitude = np.linalg.norm(out_id, axis=0)
out_id = out_id / magnitude
out_id = np.average(out_id, axis=1)
id_features.append(out_id)
register_dict[people] = np.average(id_features, axis=0)
np.save('npy/register_dict.npy', register_dict)
def recog_eval():
correct_cnt = 0
total_cnt = 0
print('Start Recognization Eval:')
for wavpath, people_valid in valid_list: