I use my dataset, when run the diversity_sampler, it encounter "assert len(cluster_i) >= 1" problem

[syyxtl]

[ChenhongyiYang]

Hi, sorry for the late reply. This problem emerges when an image is assigned to more than one centroid, which is often caused by problems in distance computing. Could you please check the computed image distances?

[BroadswordZhang]

I also have the problem when I use my dataset, please tell me how can I do if I know the image distance

[dust-removal]

Hi, I also want to train my own dataset, can you tell me what changes you made to train your own dataset

[sharat29ag]

@ChenhongyiYang I also faced this error while training on COCO.

[shiyuanyu123]

@ChenhongyiYang I also faced this error while training on my dataset.

[Parkkkkk]

Try with that code

@staticmethod
def kmeans(dis_matrix, K, n_iter=100):
    N = dis_matrix.shape[0]
    centroids = DiversitySampler.k_centroid_greedy(dis_matrix, K)
    data_indices = np.arange(N)

    assign_dis_records = []
    for _ in range(n_iter):
        centroid_dis = dis_matrix[:, centroids]
        cluster_assign = np.argmin(centroid_dis, axis=1)
        assign_dis = centroid_dis.min(axis=1).sum()
        assign_dis_records.append(assign_dis)

        new_centroids = []
        for i in range(K):
            cluster_i = data_indices[cluster_assign == i]
            if len(cluster_i) == 0:
                new_centroid_i = np.random.choice(data_indices)
            else:
                dis_mat_i = dis_matrix[cluster_i][:, cluster_i]
                new_centroid_i = cluster_i[np.argmin(dis_mat_i.sum(axis=1))]
            new_centroids.append(new_centroid_i)
        centroids = np.array(new_centroids)
    return centroids.tolist()