Determining Ratio of Prunable Channels in MobileNet by Sparsity for Acoustic Scene Classification
share
MobileNet is widely used for Acoustic Scene Classification (ASC) in embedded systems. Existing works reduce the complexity of ASC algorithms by pruning some components, e.g. pruning channels in the convolutional layer. In practice, the maximum proportion of channels being pruned, which is defined as Ratio of Prunable Channels (R_PC), is often decided empirically. This paper proposes a method that determines the R_PC by simple linear regression models related to the Sparsity of Channels (S_C) in the convolutional layers. In the experiment, R_PC is examined by removing inactive channels until reaching a knee point of performance decrease. Simple methods for calculating the S_C of trained models and resulted R_PC are proposed. The experiment results demonstrate that 1) the decision of R_PC is linearly dependent on S_C and the hyper-parameters have a little impact on the relationship; 2) MobileNet shows a high sensitivity and stability on proposed method.
READ FULL TEXT
