def kmeans(data, k, max_iter=100):
    n, m = data.size()
    centers = data[torch.randperm(n)[:k]]
    for _ in range(max_iter):
        distance = torch.sum((data.unsqueeze(1) - centers.unsqueeze(0)).pow(2), 2)
        label = torch.argmin(distance, 1)
        new_centers = torch.zeros_like(centers)
        for i in range(k):
            new_centers[i] = torch.mean(data[label == i], 0)
        if torch.all(new_centers == centers):
            break
        centers = new_centers
    return centers, label