hi folks,
i think i may have found a subtle bug in the meanshift algo as shown in lesson 12.
I’m referring to this code (not the GPU optimized or matrix mult versions):
def one_update(X):
for i, x in enumerate(X):
dist = torch.sqrt( ((x-X)**2).sum(1) )
weight = gaussian(dist, 2.5)
#weight = tri(dist, 8)
X[i] = (weight[:,None]*X).sum(0)/weight.sum()
Note that as we iterate thru each point in X we update it immediately in place. We then use this updated point in all future calculations including the current iteration of the loop. So for example when we update X[1] we will be using the already updated X[0], while the rest of the array has not been updated yet. The first point in X will be calculated based on points that haven’t been updated yet, while the last point in the array will be calculated based using points that have already been updated. This effectively introduces a bias.
I tested to see if this ‘bug’ results in different centroids and it does. This version of the code removes this issue and produces slightly different centroids.
def one_update2(X):
res = torch.zeros_like(X)
for i, x in enumerate(X):
dist = torch.sqrt( ((x-X)**2).sum(1) )
weight = gaussian(dist, 2.5)
#weight = tri(dist, 8)
res[i] = (weight[:,None]*X).sum(0)/weight.sum()
return res
Am I understanding this algo and code correctly?
thanks, jason
