19、自动机器学习助力DenseNet压缩与贝叶斯脉冲决策电路探索

自动机器学习助力DenseNet压缩与贝叶斯脉冲决策电路探索

1. 自动DenseNet稀疏化算法

自动DenseNet稀疏化算法（ADS）旨在自动修剪DenseNet中冗余的跳跃连接，以提高网络性能并实现高效压缩。该算法的主要流程如下：

Algorithm 1. Automatic DenseNet Sparsification
Input: A pre-trained DenseNet D, Maximum iteration step N, Batch size b and Compression ratio c
Output: The best-performing sparsified DenseNet under compression ratio
1: Compute the number of the agent’s actions, T, according to the compression ratio c
2: for iteration step=1 to N do
3:    Given a set of matrices, M = {M01, M02, ... , M0b}, which represents b uncompressed DenseNets D
4:    for t=1 to T do
5:        The encoder network encodes the initial state of M as S = {St1, St2, ... , Stb}
6:        The policy network takes the state S as input, and