nickb- / layer_masking

Code to reproduce our ICCV paper "Towards Improved Input Masking for Convolutional Neural Networks"

First install the prerequisites in requirements.txt

Then, if you want to run layer masking on timm resnet (the data augmented one robust to greyout), replace timm/models/resnet.py with timm_resnet.py and rename it to resnet.py. Also, place a copy of sal_layers.py in timm/models .

Then run the jupyter notebooks:

1. Analysis: Segment removal experiments
2. Mask shape bias: Analysis of shape bias of masking tecnique
3. LIME (quantitative/qualitative): Experiments on LIME
4. Measure Linearity: Linearity experiments

Make sure to insert the correct paths for Salient ImageNet and Pixel ImageNet in the notebooks and code