I have a feature map that needs to be divied with another 2d constant, the feature map has shape [H, W, ...] , the 2d constant has shape [H, W] , however when I use np.divide it says these two tensor can not be broadcast to a common shape:
import numpy as np
a = np.random.rand(4, 3 ,3)
b = np.random.rand(4, 3)
c = np.divide(a, b, where=b > 0)
as the numpy broadcast rules say b will be broadcast to shape [1, 4, 3] instead of [4, 3, 1] , thus I can't do this operation with boradcast.
I would like to divide my feature map with that constant tensor where the constant is not 0 to avoid zero-div error, what should I do ?
Personally, I would probably write:
import numpy as np
np.random.seed(3) # Set seed for reproducibility
a = np.random.rand(4, 3 ,3)
b = np.random.normal(size=(4, 3)) # Use `normal` to produce more `where` misses
b = b[..., np.newaxis]
c = np.where(b > 0, a / b, a)
However, to stay closer to your code, one could write equivalently:
import numpy as np
np.random.seed(3) # Set seed for reproducibility
a = np.random.rand(4, 3 ,3)
b = np.random.normal(size=(4, 3)) # Use `normal` to produce more `where` misses
c = a.copy()
np.divide(a, b[..., np.newaxis], where=b[..., np.newaxis] > 0, out=c)
Notable changes to your code:
b so that its shape is compatible with a (i.e. expand the 4×3 array to a 4×3×1 array).out=c with np.divide(), because its where argument only produces meaningful results in connection with a given out array (see documentation of np.divide()).a as the output array, meaning that in places where b <= 0, the given a value will be the result instead (which is what you want, if I read your question correctly).In either case, the result will be contained in c.