PaperReading

Network Slimming: Learning Efficient Convolutional Networks through Network Slimming

May 2020

tl;dr: Channel pruning by learning with L1 sparse constraint on batch norm.

Overall impression

This paper proposes a simple idea of gamma (channel scaling factor) decay. It adds a L1 sparse constraint on BN scale parameter gamma. During inference time, when gamma is smaller than a global threshold, set the entire channel to zero.

The work is concurrent with batchnorm pruning which has a very similar idea.

Key ideas

Technical details

Notes