Slim

Batch-shaping for learning conditional channel gated networks,ICLR20

Spatially Adaptive Inference with Stochastic Feature Sampling and Interpolation,ECCV20,oral

traditional: prune weights and neurons

this paper: prune feature map.

We further remove the interpolation module from our method and fill the features of unsampled points with 0. Results show that removing interpolation does not affect performance on the ImageNet validation set. This is inconsistent with object detection and semantic segmentation. We believe that this is because the classification network is focused on extracting global feature representations. Therefore, as long as the features of certain key points are calculated and preserved, the global features will not be affected and the performance will not be hurt. In other words, in the image classification task, it is not important to reconstruct the features of unsampled points by interpolation.