Skip to content

Latest commit

 

History

History

tools

Folders and files

NameName
Last commit message
Last commit date

parent directory

..

images

images

 
 

modelvshuman

modelvshuman

 
 

README.md

README.md

 
 

cnn_attention.py

cnn_attention.py

 
 

decision_bound.py

decision_bound.py

 
 

kernal_visualization.py

kernal_visualization.py

 
 

shape_texture_bias.py

shape_texture_bias.py

 
 

vit_attenton.py

vit_attenton.py

 
 

README.md

Usage of Analytical Tools

Attention Visualization of CNNs

Usage: 
    python cnn_attention.py [OPTIONS...]

OPTIONS:
    --model [ARCH in timm]
    --ckpt_path [URL or PATH of the model weights]
    --input_image [PATH or URL of the input image]
    --method [METHOD of CNN visualization]
  • CNN Attention for Standard ResNet50:
python cnn_attention.py --model resnet50 --ckpt_path https://download.pytorch.org/models/resnet50-19c8e357.pth --input_image http://alisec-competition.oss-cn-shanghai.aliyuncs.com/xiaofeng/easy_robust/assets/test.png

  • CNN Attention for Robust ResNet50:
python cnn_attention.py --model resnet50 --ckpt_path http://alisec-competition.oss-cn-shanghai.aliyuncs.com/xiaofeng/imagenet_pretrained_models/advtrain_models/advtrain_resnet50_ep4.pth --input_image http://alisec-competition.oss-cn-shanghai.aliyuncs.com/xiaofeng/easy_robust/assets/test.png

Convolution Kernel Visualization

Usage: 
    python kernal_visualization.py [OPTIONS...]

OPTIONS:
    --model [ARCH in timm]
    --ckpt_path [URL or PATH of the model weights]
  • Visualize the Filters of a Standard ResNet50::
python kernal_visualization.py --model resnet50 --ckpt_path https://download.pytorch.org/models/resnet50-19c8e357.pth

  • Visualize the Filters of a Robust ResNet50::
python kernal_visualization.py --model resnet50 --ckpt_path http://alisec-competition.oss-cn-shanghai.aliyuncs.com/xiaofeng/imagenet_pretrained_models/advtrain_models/advtrain_resnet50_ep4.pth

Visualization of Decision Boundary

Usage: 
    python decision_bound.py [OPTIONS...]

OPTIONS:
    --model [ARCH in timm]
    --ckpt_path [URL or PATH of the model weights]
    --input_image [URL or PATH of the input image]
  • Visualize the Decision Boundary of a Standard ResNet50:
python decision_bound.py --model resnet50 --ckpt_path https://download.pytorch.org/models/resnet50-19c8e357.pth --input_image http://alisec-competition.oss-cn-shanghai.aliyuncs.com/xiaofeng/easy_robust/assets/test.png --max-eps 120 --eps-step 3

  • Visualize the Decision Boundary of a Robust ResNet50:
python decision_bound.py --model resnet50 --ckpt_path http://alisec-competition.oss-cn-shanghai.aliyuncs.com/xiaofeng/imagenet_pretrained_models/advtrain_models/advtrain_resnet50_ep4.pth --input_image http://alisec-competition.oss-cn-shanghai.aliyuncs.com/xiaofeng/easy_robust/assets/test.png --max-eps 120 --eps-step 3

Attention Visualization of ViTs

Usage: 
    python vit_attenton.py [OPTIONS...]

OPTIONS:
    --model [ARCH in timm]
    --ckpt_path [URL or PATH of the model weights]
    --input_image [PATH or URL of the input image]
  • ViT Attention for Standard ViT-B/16:
python vit_attenton.py --model vit_base_patch16_224 --ckpt_path http://alisec-competition.oss-cn-shanghai.aliyuncs.com/xiaofeng/imagenet_pretrained_models/clean_models/timm_model/vit_base_patch16_224.pth --input_image http://alisec-competition.oss-cn-shanghai.aliyuncs.com/xiaofeng/easy_robust/assets/test.png

  • ViT Attention for Robust ViT-B/16:
python vit_attenton.py --model vit_base_patch16_224 --ckpt_path http://alisec-competition.oss-cn-shanghai.aliyuncs.com/xiaofeng/imagenet_pretrained_models/advtrain_models/advtrain_vit_base_patch16_224_ep4.pth --input_image http://alisec-competition.oss-cn-shanghai.aliyuncs.com/xiaofeng/easy_robust/assets/test.png

Shape vs. Texture Biases Analysis

Examples:

python shape_texture_bias.py

then check tools/modelvshuman/figures/example-figures/.