文章目录

• why
• -
• Abstract
• Introduction
• • Contributions
• Related Works
• • Filter-level pruning
  • Block-level pruning
  • Data Limited Knowledge Distillation
• Method
• • Overview
  • The motivation to drop blocks
  • The recoverability of the pruned model
  • Recover the accuracy of the pruned model
• Experimental Results
• • Different amounts of training data
  • Different acceleration ratios
  • The data-latency-accuracy tradeoff
• Results on MobileNetV2
• • Train with synthesized/out-of-domain images
  • Different criteria for dropping blocks
• Conclusions and Future Works