<p>Existing parametric clustering models aim to fit the distribution of data. Despite the simplicity, it binds model parameters with the cluster number, thus explaining its vulnerability when the true cluster number is not available. Remarkable nonparametric extensions have been proposed, whose essential advantage lies in their target of fitting the distribution of centroids and their ability to infer the cluster number automatically. Restricting the centroid distribution to be discrete, the introduced Dirichlet process prior allows the centroid distribution with a limited number of supports to achieve the best fit. However, it becomes disadvantaged due to the resulting complex and inefficient inference techniques. Instead of learning a discrete distribution over centroids laboriously, we resort to a fit-then-prune strategy, named auto-ClusTering (ACT), equipped with continuous centroid distribution estimation. To be specific, a continuous distribution, approximated by a group of particles, is introduced to model the centroid distribution, no longer compromised by the unknown true cluster number. Then, we deliver an equivalent but succinct centroid distribution by safely pruning redundant particles, inferring the cluster number close to the true one. Benefiting from its simple gradient-based inference technique, ACT is easily extended to jointly train with representation learning for clustering complex text and image data. Extensive experiments demonstrate that our ACT can effectively determine the cluster number close to the ground truth with superior clustering performance.</p>