Interpretable Explanations of Black Boxes by Meaningful Perturbation with PyTorch

⭐ Star us on GitHub — it helps!!

PyTorch implementation for Interpretable Explanations of Black Boxes by Meaningful Perturbation

Install

You will need a machine with a GPU and CUDA installed.
Then, you prepare runtime environment:

pip install -r requirements.txt

Use

This codes are baesed on ImageNet dataset

python main.py --model_path=vgg19 --img_path=examples/catdog.png

Arguments:

• model_path - Choose a pretrained model in torchvision.models or saved model (.pt)
  • Examples of available list: ['alexnet', 'vgg19', 'resnet50', 'densenet169', 'mobilenet_v2' ,'wide_resnet50_2', ...]
• img_path - Image Path
• perturb - Choose a perturbation method (blur, noise)
• tv_coeff - Coefficient of TV
• tv_beta - TV beta value
• l1_coeff - L1 regularization
• factor - Factor to upsampling
• lr - Learning rate
• iter - Iteration

How to use your customized model

If you want to use customized model that has a type 'OrderedDict', you shoud type a code that loads model object.

Search 'load model' function in utils.py and type a code such as:

from yourNetwork import yourNetwork())
model=yourNetwork()

Understanding this Paper!

✅ Check my blog!! Here