What to do when there are multiple values in the kernel that equals to the max? For example, for these values:

array([[0., 0.],
       [0., 0.]])

The max is simply 0. What should max indices look like? should it be True's for all occurrence of max:

array([[ True,  True],
       [ True,  True]])

Or the first occurrence of the max:

array([[ True, False],
       [False, False]])

Pytorch uses the first occurrence of the max, while some source uses the other one (he uses pos = np.where(result == view, 1, 0), which essentially records 1 for all occurrence of max).

Edit: for forward prop it doesn't matter, but during backpropagation, for example, if the upstream gradient is [[1,1],[1,1]], then the gradient of current node will be different: [[1,1],[1,1]] (all occurrence) vs [[1,0],[0,0]] (first occurrence)

CodePudding user response：

After even a single back-propagation it is very unlikely that the kernel values will remain equal. So it shouldn't be a real issue, should it?