shape-bias: A Python repository from anguyen8

The shape and simplicity biases of adversarially robust ImageNet-trained CNN

This repository contains source code necessary to reproduce some of the main results in our paper:

Chen*, Agarwal*, Nguyen* (2020). The shape and simplicity biases of adversarially robust ImageNet-trained CNNs. paper | code

If you use this software, please consider citing:

@article{chen2020shape,
  title={The shape and simplicity biases of adversarially robust ImageNet-trained CNNs},
  author={Chen, Peijie and Agarwal, Chirag and Nguyen, Anh},
  journal={arXiv preprint arXiv:2006.09373},
  year={2020}
}

1. Robustness package

We use Madry's robustness packages to train the robust models for our experiment. This repo already include their packages (in robustness folder), so you don't need to download or install the robustness packages.

2. Pre-trained models

The pretrained models are avaliable here.

3. Requirements

We recommend using Anaconda to create new enviroment:

conda env create -f enviroment.yml -n {NAME OF NEW ENV}

4. Network Dissection

Our Network Dissection experiment was developed on top of Bolei Zhou's network dissection.

In order to run Network Dissection, you will need follow network dissection repo to download the Broden dataset using:

    ./script/dlbroden.sh
    ./script/dlzoo_example.sh

Setup the model, target layers and other settings in settings/settings.py.
Then run:

    python netdissect.py

5. An example of loading a robust model and setting up feature hooks

This sample is avaliable at load_sample.py:

from settings import settings
from loader.model_loader import loadrobust
# for loading robust version
from robustness.model_utils import make_and_restore_model
from robustness.datasets import DATASETS
import torch
from PIL import Image
from torchvision import transforms

# target layers can be changed in settings file. For this sample, it will hook layer ["1", "4", "7", "9", "11"] of AlexNet-R
features_blobs = []
def hook_feature(module, input, output): 
    features_blobs.append(output.data.cpu().numpy())

def sample_loading_and_hooking():
    global features_blobs
    # load robust model
    if settings.MODEL[-1] in ['r', 'R']:
        '''loading robust model'''
        dataset = DATASETS['imagenet']('robustness/dataset')
        model, checkpoint = make_and_restore_model(arch=settings.MODEL[:-2],
                                                dataset=dataset, parallel=settings.MODEL_PARALLEL,
                                                resume_path=settings.MODEL_FILE)
        # add hooker on target layers
        model = loadrobust(hook_feature, model, checkpoint,  settings.FEATURE_NAMES)
    
    img = Image.open('figures/real/ILSVRC2012_val_00002691.JPEG')

    TEST_TRANSFORMS_IMAGENET = transforms.Compose([
            transforms.Resize(256),
            transforms.CenterCrop(224),
            transforms.ToTensor(),
        ])

    input_tensor = TEST_TRANSFORMS_IMAGENET(img).unsqueeze(0).cuda()
    # make predictions
    logit, _ = model(input_tensor)
    pred_class = logit.argmax()
    # put hooked features in to a dict.
    hooked_features = dict(zip(["1", "4", "7", "9", "11"], features_blobs)) # dict for features of layer1, 4, 7, 9, 11.
    return logit, pred_class, hooked_features

if __name__ == '__main__':
    #* A naive sample for load and hook features from robust model
    logit, pred_class, hooked_features = sample_loading_and_hooking()

6. ImageNet-CL

We provide the index of ImageNet-CL(.npy), where the robust model and standard model both make correct predictions. There are three .npy files, one for each pair of network. i.e. (AlexNet, AlexNet-R).

Sample:

import numpy as np
alexnet_intersect_img_idx = np.load('data/intersection_alexnet_alexnet-r.npy')

The .npy file contain a list of image names of ImageNet val set.

>>> print(intersect_img_idx)
['ILSVRC2012_val_00000003.JPEG' 'ILSVRC2012_val_00000007.JPEG'
 'ILSVRC2012_val_00000012.JPEG' ... 'ILSVRC2012_val_00049990.JPEG'
 'ILSVRC2012_val_00049991.JPEG' 'ILSVRC2012_val_00049999.JPEG']

7. Example of evaluating adversarially robust models on ImageNet

To reproduce Table. 1 prediction accuracy of robust model in our paper, use the following script:

python eval_on_imagenet.py --model_name {MODEL ARCHITECTURE} --model_path {CHECK POINT PATH} --data_path {PATH TO ILSVRC2012} --madry_model

for example,

python eval_on_imagenet.py --model_name alexnet --model_path zoo/alexnet_r.pt --data_path path/to/ILSVRC2012 --madry_model

anguyen8/shape-bias