Thanks for your work. But it's a bit strange here in the coreset sampling part

[mengxianghan123]

PatchCore_anomaly_detection/sampling_methods/kcenter_greedy.py

Lines 104 to 120 in 59b67d9

	for _ in range(N):
	if self.already_selected is None:
	# Initialize centers with a randomly selected datapoint
	ind = np.random.choice(np.arange(self.n_obs))
	else:
	ind = np.argmax(self.min_distances)
	# New examples should not be in already selected since those points
	# should have min_distance of zero to a cluster center.
	assert ind not in already_selected
	self.update_distances([ind], only_new=True, reset_dist=False)
	new_batch.append(ind)
	print('Maximum distance from cluster centers is %0.2f'
	% max(self.min_distances))
	self.already_selected = already_selected

I think the wanted logic here should be initialing a centre point by random choice (the first if-branch) and all the entire centre points are chosen by argmax the distance (the second if-branch).
But for the "self.already_selected" is initialized as an empty list in line 49(not None), so we'll never get into the first if-branch. As for the centre point initializing process, the expression "np.argmax(self.min_distances)" will return 0 for "np.argmax(None)" will return 0.
So, as a result, the program selects index-0-feature as algorithm initialization everytime instead of random selecting one as a common practice.

[JefferyChiang]

I agree with your point. I think the first if-branch should be modify into "if not self.already_selected: " for more accurate to the kcenter_greedy algorithm. Have you modify the code and test the difference ?

[mengxianghan123]

Yeah I modifyed the code like the following:

for _ in range(N):
            if not self.already_selected:
                # Initialize centers with a randomly selected datapoint
                ind = np.random.choice(np.arange(self.n_obs))
            else:
                ind = np.argmax(self.min_distances)
           
            assert ind not in self.already_selected

            self.update_distances([ind], only_new=True, reset_dist=False)
            new_batch.append(ind)
            print("Maximum distance from cluster centers is %0.2f" % max(self.min_distances))

            self.already_selected.append(ind)

I didn't test the difference with thorough experiments.
I conjectured that there will not be significant difference between the two implementation but our modification is more up to standard and more consistent with the intention of the algorithm

[HoseinHashemi]

This point makes total sense. I also modified the code but haven't thoroughly tested.