[FEATURE] Add ParC-Net (EdgeFormer), a pure-conv version of MobileViT, which has better performance than MobileViT v1 and v2.

[hkzhang-git]

Is your feature request related to a problem? Please describe. We propose a new basic operation, position aware circular convolution (ParC). Differing from previous convolution operations, the proposed ParC has global receptive field. Experimental results show that ParC uniformly improves performance of various typical models.

Based on ParC, we build ParC-Net (was named edgeformer), where Mobilevit is used as baseline and self attention operations are replaced with our proposed ParC. For classification on imagenet-1k, ParC-Net saves 11% parameters and 13% computation cost but gaining 0.2% higher accuracy and 23% faster inference speed (on ARM based RK3288) and 3.77 × faster inference speed (on deep learing chip) compared with MobileViT.

Describe the solution you'd like

This work has been accpted by ECCV 2022, we hope the proposed ParC can be used by other researchers. Please refer to refer to https://arxiv.org/abs/2203.03952 to find more details.

Code is available at: https://github.com/hkzhang91/ParC-Net

Additional context Add any other context or screenshots about the feature request here.