Darknet: Why are the prediction results different from pjreddie/darknet?

[keto33]

I use the same files in both, but the results are different:

/darknet classifier predict cfg/imagenet1k.data cfg/darknet19.cfg darknet19.weights data/dog.jpg

pjreddie/darknet

data/dog.jpg: Predicted in 0.003919 seconds.
42.35%: malamute
23.06%: Eskimo dog
12.67%: Siberian husky
 2.75%: bicycle-built-for-two
 1.21%: mountain bike

AlexeyAB/darknet

data/dog.jpg: Predicted in 4.415000 milli-seconds.
malamute: 0.521573
Eskimo dog: 0.272391
Siberian husky: 0.116594
dogsled: 0.021984
bicycle-built-for-two: 0.008623

This difference is much more significant for other images.

My question is: when using the same pre-trained models and data, why are the predictions different?

[WongKinYiu]

The pre-pocessing of pjreddie/darknet and AlexeyAB/darknet are different.

https://github.com/pjreddie/darknet/blob/master/examples/classifier.c#L589

https://github.com/AlexeyAB/darknet/blob/master/src/classifier.c#L847-L848