def calc_128_vec(self, img):
face_img = utils.pre_process(img)
# print('face image is: ', face_img.shape)
facenet_inputs = []
facenet_outputs = []
facenet_inputs.append(tritonclient.http.InferInput(self.Facenet_inputs[0], face_img.shape, 'FP32'))
facenet_inputs[0].set_data_from_numpy(face_img.astype(np.float32), binary_data=True)
facenet_outputs.append(tritonclient.http.InferRequestedOutput(self.Facenet_outputs[0], binary_data=True))
output = self.triton_client.infer("facenet_tf", inputs=facenet_inputs, outputs=facenet_outputs)
pre = output.as_numpy(self.Facenet_outputs[0])
pre = utils.l2_normalize(np.concatenate(pre))
pre = np.reshape(pre, [128])
# print('pre is: ', pre.shape)
return pre
def aggregate_templates(templates, features, method, T=1.0):
for i,t in enumerate(templates):
if len(t.indices) > 0:
if method == 'mean':
t.feature = utils.l2_normalize(np.mean(features[t.indices], axis=0))
if method == 'PFE_fuse':
t.mu, t.sigma_sq = utils.aggregate_PFE(features[t.indices], normalize=True, concatenate=False, T=T)
t.feature = t.mu
if method == 'PFE_fuse_match':
if not hasattr(t, 'mu'):
t.mu, t.sigma_sq = utils.aggregate_PFE(features[t.indices], normalize=True, concatenate=False, T=T)
t.feature = np.concatenate([t.mu, t.sigma_sq])
else:
t.feature = None
if i % 1000 == 0:
sys.stdout.write('Fusing templates {}/{}...\t\r'.format(i, len(templates)))
print('')