python - getting ValueError: operands could not be broadcast together with shapes (3,224,224) (3,) when trying to subtract from channel wise mean in Caffe -

this follow-up question this. want subtract each image mean. based on issue on github , this other similar question, , this classification example when feed cropped version of images network, need subtract mean pixel using sth :

mu = mean_file.mean(1).mean(1) 

but irony when want this:

.. = (img[:,0:224,0:224] - mu) 

i valueerror: operands not broadcast shapes (3,224,224) (3,) i'm not well-versed @ python , numpy , can't figure out error message trying convey, appreciated.

currently, i'm cropping mean file! not ideal! better nothing!

.. = (img[:,0:224,0:224] - mean_image[:,0:224,0:224]) 

replace

mu = mean_file.mean(1).mean(1) 

with

mu = mean_file.mean(1).mean(1)[:,none,none] 

it seems trying subtract 1d vector (shape of (3,)) 3d array (shape of (3,224,224)). in order numpy needs broadcast 1d vector dimensions of 3d array (much matlab's bsxfun). numpy understand dimensions broadcast, 1 needs add singleton dimensions 1d vector:

mu[:,none,none] 

is of shape (3,1,1) , should enable numpy perform subtraction correctly.