Top-Down Bottom-Up Attention on top of Swin Transformer

Implementation of top-down bottom-up attentional modules on top of Swin Transformer object detection network.

Installation:

• Follow official doc from here.
• Download the model from here

How to run

Test

Single image test

python -m swin_td_bu_att demo.jpg configs/td_bu_attention/topdown_bottomup_attentional_swin.py swin_tiny_patch4_window7_224.pth --device cuda --out-file result.jpg

Evaluate all test set:

python swin_td_bu_att/eval.py

Train

python swin_td_bu_att/train.py

Current implementation progress:

• Implement modules
• Implement functioning test code
• Fix training parameters in config file
• Implement missing training functionality
• Decrease the training loss