Abstract: Dynamic pruning leverage the feature information of each input sample to dynamically adjust the network structure, generating multiple subnetworks suitable for different sample complexity.