This is an official implementation for "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows" on Object Detection and Instance Segmentation.
In order to expand the parameters of the swing transformer backbone network and improve its monitoring effect, I added Swin Transformer MoE (referred to as Swin-T MoE hereafter) as a new backbone network.
Modification
I added Swin Transformer MoE (referred to as Swin-T MoE hereafter) as a new backbone network :
.\mmdet\models\backbones\swin_transformer_moe.py
As the output of Swin-T MoE is different from Swin-T, I modified the extract_feat function in .\mmdet\models\detectors\two_stage.py.
BC-breaking (Optional)
I think no.
Use cases (Optional)
I provided the relevant configuration files for reference:
contains the parameters for the Swin-T MoE backbone network:
Motivation
In order to expand the parameters of the swing transformer backbone network and improve its monitoring effect, I added Swin Transformer MoE (referred to as Swin-T MoE hereafter) as a new backbone network.
Modification
I added Swin Transformer MoE (referred to as Swin-T MoE hereafter) as a new backbone network :
.\mmdet\models\backbones\swin_transformer_moe.pyAs the output of Swin-T MoE is different from Swin-T, I modified the
extract_featfunction in.\mmdet\models\detectors\two_stage.py.BC-breaking (Optional)
I think no.
Use cases (Optional)
I provided the relevant configuration files for reference: contains the parameters for the Swin-T MoE backbone network:
contains the modified configuration for the backbone network:
Thanks for your read.