The implementation of Eq. (10)

[panlinchao]

Hello,
Thanks for this promising work! I am reproducing the results from your paper using the provided configuration. I noticed a difference between how Equation (10) is implemented in the code and how it is defined. Considering the similarity condition between different centroids, Equation (10) uses larger ACD clusters while the code uses smaller ACD clusters. Could you clarify if there is a specific reason for this difference or guide me to the correct implementation?
Thanks for your help.

TABASCO/experiment/train_cifar_ssl.py

Lines 361 to 369 in 712992b

	if (abs((current_centriod_distance-max_centriod_distance[0]).item()) < 0.1 * current_centriod_distance.item() and (sample_num_list[class_num] < sample_num_list[max_centriod_indice[0].item()])):
	select_sample_index_list.extend(select_sample_index_list_1[class_num])
	select_sample_prob_list.extend(select_sample_prob_list_1[class_num])
	else:
	select_sample_index_list.extend(select_sample_index_list_2[class_num])
	select_sample_prob_list.extend(select_sample_prob_list_2[class_num])

TABASCO/experiment/train_cifar_ssl.py

Lines 342 to 347 in 712992b

	select_sample_dimensionnal_status.append(0)
	select_sample_index_list_1.append(cluster_index_1_acd)
	select_sample_prob_list_1.append(prob_acd[:,gmm_acd.means_.argmin()][cluster_select_index_1_acd])
	select_sample_index_list_2.append(cluster_index_2_acd)
	select_sample_prob_list_2.append(prob_acd[:,gmm_acd.means_.argmax()][cluster_select_index_2_acd])