Skip to content

Improvement upon the architecture from "ParC-Net: Position Aware Circular Convolution with Merits from ConvNets and Transformer"

Notifications You must be signed in to change notification settings

MahatKC/ParC-Net-HalfBlocks

 
 

Repository files navigation

Increasing ParC-Net’s Performance with Half-blocks

Project created for the CISC 867 class of Prof. Hazem Abbas, in Fall '23 at Queen's University - Canada, by @dashtiali, @MahaK21 and @MahatKC.


Fork from the original repository for the paper "ParC-Net: Position Aware Circular Convolution with Merits from ConvNets and Transformer". Our changes are mainly to the cvnets/modules/edgeformer_block.py file, specifically the forward function on line 489.

Our goal here was to increase the performance of the original ParC-Net architecture by reducing redundancy. For that matter, we designed the ParC-Half blocks, which do convolution only for the even-indexed rows/columns of the input, which are shown in the image below.

Then, using the proposed ParC-Half blocks, we did 3 experiments to compare our approach with the approach from the paper:

The following results were achieved, which are compared to the pre-trained model of the paper (provided in their repository):

Therefore, we were able to improve upon the accuracy of the original model while decreasing its cost.

About

Improvement upon the architecture from "ParC-Net: Position Aware Circular Convolution with Merits from ConvNets and Transformer"

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 99.9%
  • Shell 0.1%