torch7: Unexpected 'counts' in k-Means Clustering - k-means

Question

I am trying to apply k-means clustering on a set of images (images are loaded as float torch.Tensors) using the following segment of code:
print('[Clustering all samples...]')
local points = torch.Tensor(trsize, 3, 221, 221)
for i = 1,trsize do
points[i] = trainData.data[i]:clone() -- dont want to modify the original tensors
end
points:resize(trsize, 3*221*221) -- to convert it to a 2-D tensor
local centroids, counts = unsup.kmeans(points, total_classes, 40, total_classes, nil, true)
print(counts)
When I observe the values in the counts tensor, I observe that it contains unexpected values, in the form of some entries being more than trsize, whereas the documentation says that counts stores the counts per centroid. I expected that it means counts[i] equals the number of samples out of trsize belonging to cluster with centroid centroids[i]. Am I wrong in assuming so?
If that indeed is the case, shouldn't sample-to-centroid be a hard-assignment (i.e. shouldn't counts[i] sum to trsize, which clearly is not the case with my clustering)? Am I missing something here?
Thanks in advance.

Answer 1

In the current version of the code, counts are accumulated after each iteration
for i = 1,niter do
-- k-means computations...
-- total counts
totalcounts:add(counts)
end
So in the end counts:sum() is a multiple of niter.
As a workaround you can use the callback to obtain the final counts (non-accumulated):
local maxiter = 40
local centroids, counts = unsup.kmeans(
points,
total_classes,
maxiter,
total_classes,
function(i, _, totalcounts) if i < maxiter then totalcounts:zero() end end,
true
)
As an alternative you can use vlfeat.torch and explicitly quantize your input points after kmeans to obtain these counts:
local assignments = kmeans:quantize(points)
local counts = torch.zeros(total_classes):int()
for i=1,total_classes do
counts[i] = assignments:eq(i):sum()
end