[Model Deployment] kpu and ncc_infer give different outputs

With kmodelV4

Doing a 2-classes classification(Mobilenet)

Running on a same jpg,
kpu gives an output (0.9994016, 0.0005984303),
while ncc infer gives (0.3547561466693878, 0.6452438831329346) (decoding from the .bin file)

MAIXPY codes are following:

import KPU as kpu
import image,lcd
import uos,os

print(os.listdir())
task = kpu.load(0x300000)
print(task)

a = kpu.set_outputs(task,0,1,1,2)

l = os.listdir()

for i in l :
if i.endswith('.jpg'):
    print(i)
    img = image.Image(i)
    
    a=img.pix_to_ai()
    a = kpu.forward(task, img)
    print(a[0],a[1])

a = kpu.deinit(task)