How do you balance the trade-offs between model accuracy, inference latency, and computational cost when choosing a pre-trained model for a pipeline?
