ParentFull threadfzliu·This is typically done by taking the activations from a neural network. Activations from layer depths are generally preferred since they are more representative.View on HN