choice of mean values in mean_subtract() function

[akshay-vaidya]

How was the mean values of BGR decided?
In the mean_subtract() function, how are the values 123.68, 116.779 and 103.939 decided?
def mean_subtract(img):
img = T.set_subtensor(img[:,0,:,:],img[:,0,:,:] - 123.68)
img = T.set_subtensor(img[:,1,:,:],img[:,1,:,:] - 116.779)
img = T.set_subtensor(img[:,2,:,:],img[:,2,:,:] - 103.939)

return img / 255.0

And how do I customize it?

[duggalrahul]

Those are the channel-wise mean of images in the ImageNet dataset on which the model was pre-trained. This is done to keep the channel-wise mean (per batch) close to zero, which itself is known to speed up training [1].

I believe subtracting this mean, become less of a factor when the pixel intensities are scaled from [0,255] to [0,1]. So either you could compute channel-wise mean of images on your training set or otherwise not subtract any mean and divide by 255.

You could setup a simple experiment to confirm (or discard) this hypothesis. Would love to hear back on your results!

[1] Clevert, Djork-Arné, Thomas Unterthiner, and Sepp Hochreiter. "Fast and accurate deep network learning by exponential linear units (elus)." arXiv preprint arXiv:1511.07289 (2015).

[akshay-vaidya]

@duggalrahul Thanks much for your response. I will experiment on the hypothesis you have mentioned and will update you on the results. But do you think it is still valid to use subtract channel-wise mean of images of the ImageNet dataset when I am actually not using the pretrained weights and I am training the network from scratch?