Key points are not available for this paper at this time.
Transformer-based models have recently adopted increasingly complex structure (e. g. , deeper or wider stacked network) to promote the representation learning capabilities of vision recognition. However, progressively deeper or wider stacked network cause the expensive computation cost, which hinders their effective deployment in resource-constrained edge clouds or end devices. In this paper, we propose DTSNet, a dynamic transformer slimming model, which scales vision transformers (ViTs) down across layers from both of the model depth and input width. This is the first time to explore the joint reduction of input tokens and model parameters for ViTs under maintaining performance. Specifically, DTSNet adopts a diversity-enhanced weight sharing module to reduce network parameters, where the weight knowledge of multiple adjacent blocks is effectively integrated into one block. Furthermore, DTSNet designs a unified and massively scalable token pruning mechanism that dynamically discarding less important tokens with a model-driven manner, by introducing a series of discriminant parameters, which is a simple change to the common architecture of vision transformers. Extensive experiments are conducted to verify that DTSNet is able to yield high efficacy in compressing parameter space and accelerating model inference. DTSNet-T/-S/-B on ImageNet achieves 3. 0M/11. 1M/42. 9M parameters and 0. 8/2. 9/13. 7 GFLOPs, where number of parameters are reduced by 48%51% and inference speed are improved by 1. 3 1. 5. Experiments results on semantic segmentation and object detection dataset further demonstrate the potential of DTSNet on complex dense prediction tasks. Code will be available upon publication.
Xiao et al. (Tue,) studied this question.