Different value from this version and keras

[hjwdzh]

I try to print the maximum value of the feature produced by first conv layer, and it is different from keras output (more than 10x difference). I am using the same image (with the right scale), and pretrained with imagenet. Is that normal?

[malicd]

@hjwdzh Interesting that you ask that. I am inspecting the values in the feature maps (pool1-pool5) and observing really huge values (sometimes even >20k). I don't think this is desired behavior but the funny thing is that it still manages to classify the data correctly.