How to get non negative, nonzero min value from tensor?

1

I want to get a non-negative, nonzero min value from the tensor. Can someone help me to do this in an easy and efficient way

3

Will this work?

>>> x = (torch.rand((1,1,4,4)) * 2) - 1  # creating a tensor with a mixture of positive and negative value
>>> x
tensor([[[[ 0.8101, -0.5538,  0.2049, -0.8969],
          [-0.5394, -0.8587, -0.5657,  0.0481],
          [-0.9651,  0.8118, -0.3242, -0.5439],
          [ 0.0358, -0.1856, -0.1928,  0.7507]]]])
>>> min_x = torch.min(x[x>0]) # taking non-negative non-zero min value from the tensor
>>> min_x
tensor(0.0358)
4

I discovered a trick for this which avoids construction of a mask.

# how to determine the smallest positive value in the tensor x

# by shifting and taking inverse, we accomplish the following
# - zero maps to "negative infinity"
# - small positive numbers go to large positive numbers
# - negative numbers remain negative
small_shift = 0.001
invs = 1.0/(x - small_shift)

# undo the mapping
min_value = 1.0/invs.max().values() + small_shift