8p, Conclusion, the cost of collecting ImageNet has been largely ignored, but the ‘pre-training’ step in the ‘pre-training + fine-tuning’ paradigm is in fact not free when we scale out this paradigm