Should I subtract imagenet precomputation for inception_v3 model on keras inception_v3.py?
def preprocess_input(x):
x /= 255.
x -= 0.5
x *= 2.
return x
& emsp; I am using the pre-configured keras inception_v3 imagenet model ( inception_v3.py ) to complete the dataset myself.
& emsp; When I want to subtract the average of the image (123.68, 116.779, 103.939) and the RGB inverse axis in BGR , as we often do, I found that the author provided the _preprocess_input () _ function at end.I is confused about it.
& EPRS; Should I use the provided preprocess_input () function or subtract the mean and the return axis as usual?
& EPRS; Many thanks.
source to share
In fact, in the original Inception , the authors specify as the data preprocessor the function you provided (the one that has all channels zero centered and resized to an interval [-1, 1]
). As the InceptionV3 paper does not provide for the transformation of new data I think you can assume that you should use the following function:
def preprocess_input(x):
x /= 255.
x -= 0.5
x *= 2.
return x
source to share