Knowledge-guided pre-training and fine-tuning: Video representation learning for action recognition.