Hierarchical Convolution and Graph Net for Utilizing Structural Information of Image
Pytorch 1.7.0, timm 0.3.2, torchprofile 0.0.4, apex
- GraphConvNet
Model | Params (M) | FLOPs (B) | Top-1 | BaiduDisk URL |
---|---|---|---|---|
GraphConvNet-Ti | 7.7 | 1.3 | 77.1 | BaiduDisk URL |
GraphConvNet-S | 24.5 | 4.9 | 82.0 | BaiduDisk URL |
- Pyramid GraphConvNet
Model | Params (M) | FLOPs (B) | Top-1 | BaiduDisk URL |
---|---|---|---|---|
Pyramid GraphConvNet-Ti | 11.4 | 1.8 | 80.5 | BaiduDisk URL |
Pyramid GraphConvNet-S | 29.2 | 4.9 | 82.4 | BaiduDisk URL |
see run.sh
The visualization code only available to GraphConvNet and ViG
- Create a fold named 'ckpt' in './viz_nodes' and download the checkpoints of GraphConvNet-Ti or GraphConvNet-S and put them in './viz_nodes/ckpt'
- Open
viz_demo.ipnb
,and set arguments(arch,) - Run cells
- The first row: gradcam heatmaps of GraphConvNet-Ti in 4th,8th,12th layers.
- The second row: the patch(node) that has the max gradcam value(most discriminative) and its corresponding neighbors in different layers.
- The third row: add edges,the pentagram is the most discriminative node.(you can draw edges using tools such as PowerPoint, OmniGraffle..)
This repo partially uses code from vig