Abnormal result while using numpy.linalg.norm() on Nano

Same code was runed on PC(Ubuntu 16.04), Xavier(Jetpack 4.2.1) and Nano(Jetpack 4.2.1) seperately.
Same normal result was obtained on PC and Xavier, abnormal result was obtained on Nano.

It is a simple code using dlib to realize face recognition. It feels like numpy.linalg.norm(), a function to calculate norm, has a bug on Nano.
The code and two kind of results are shown below. What causes this problem and how to solve it? Humbly request your advice.

Code:

#-*- coding:utf-8 -*-
import dlib
import os
import numpy

detector = dlib.get_frontal_face_detector()
predictor = dlib.shape_predictor('shape_predictor_68_face_landmarks.dat')
facerec = dlib.face_recognition_model_v1('dlib_face_recognition_resnet_model_v1.dat')
win = dlib.image_window()

def create_known(known_path):
    person_names = []
    face_features = []

    print("Creating Known Face Library...")
    for file_name in os.listdir(known_path):
        if (file_name.find('png')<0) and (file_name.find('jpg')<0):
            continue

        img = dlib.load_rgb_image(os.path.join(known_path,file_name))
        dets = detector(img)
        if (len(dets) == 0):
            continue

        det = dets[0]
        shape = predictor(img, det)
        descriptor = facerec.compute_face_descriptor(img, shape)

        person_name = file_name[:file_name.rfind('.')]
        person_names.append(person_name)
        face_features.append(descriptor)
        print('Appending [' + person_name + '] ...')

        dlib.hit_enter_to_continue()

    return person_names, face_features

if __name__ == "__main__":
    person_names, face_features = create_known('known')
    img = dlib.load_rgb_image('nina_victory.jpg')

    dets = detector(img)
    for det in dets:
        shape = predictor(img, det)
        descriptor = facerec.compute_face_descriptor(img, shape)

        min_dist = 1.0
        person_name = 'unknown'
        for i in range(len(face_features)):
            dist = numpy.linalg.norm(numpy.array(descriptor)-numpy.array(face_features[i]))
            print(person_names[i], dist)
            if dist < min_dist:
                min_dist = dist
                person_name = person_names[i]
        print('Found [' + person_name + '] !')

        dlib.hit_enter_to_continue()

Normal Result on PC and Xavier:

nvidia@nvidia:~/workspace/facereg$ python test_facereg.py 
Creating Known Face Library...
Appending [vitory] ...
Hit enter to continue
Appending [nina] ...
Hit enter to continue
Appending [lidandan] ...
Hit enter to continue
('vitory', 0.6371401155718989)
('nina', 0.4089912978192962)
('lidandan', 0.9013869084740994)
Found [nina] !
Hit enter to continue
('vitory', 0.3036043206987178)
('nina', 0.6542178530243183)
('lidandan', 0.8145808738380327)
Found [vitory] !
Hit enter to continue
nvidia@nvidia:~/workspace/facereg$ python
Python 2.7.15+ (default, Nov 27 2018, 23:36:35) 
[GCC 7.3.0] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy
>>> numpy.__version__
'1.14.0'
>>> import dlib
>>> dlib.__version__
u'19.17.0'
>>>

Abnormal Result on Nano:

nvidia@nvidia:~/workspace/facereg$ python test_facereg.py 
Creating Known Face Library...
Appending [nina] ...
Hit enter to continue
Appending [lidandan] ...
Hit enter to continue
Appending [vitory] ...
Hit enter to continue
('nina', 2.7818110035028154e+24)
('lidandan', 2.781277577877308e+24)
('vitory', 1.1984561081838429e+25)
Found [unknown] !
Hit enter to continue
('nina', nan)
('lidandan', nan)
('vitory', nan)
Found [unknown] !
Hit enter to continue
nvidia@nvidia:~/workspace/facereg$ python
Python 2.7.15+ (default, Nov 27 2018, 23:36:35) 
[GCC 7.3.0] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import numpy
>>> numpy.__version__
'1.14.0'
>>> import dlib
>>> dlib.__version__
u'19.17.0'
>>>

Hi,

It’s known that dlib will return NAN or extremely small value from face_encodings():
[url]https://devtalk.nvidia.com/default/topic/1049660/jetson-nano/issues-with-dlib-library/[/url]

Current workaround is to set the cuDNN convolution algorithm into default and recompile the dlib library.
Here are the detail instructions for your reference:
[url]issues with dlib library - Jetson Nano - NVIDIA Developer Forums

Thanks.