Description:

This repo contains code for the paper Understanding Robustness of Deep Neural Networks under Natural Variations

Environment:

python 3.6+

tensorflow 1.13+

scikit-learn

scikit-image

matplotlib

skimage

Multicore-TSNE

File Structures

The file structure looks like the following:

|--backup(need to be created manually)
    |--cifar10
    |--fmnist
    |--svhn
|--data(need to be created manually)
    |--cifar10
    |--fmnist
    |--svhn
    |--simulation (the online data path needs to be changed to this)
|--load(need to be created manually)
    |--cifar10
    |--fmnist
    |--svhn
    |--simulation (the online pretrained model path needs to be changed to this)
|--scripts
    |--analysis.py
    |--rq1.py
    |--train_classification_model.py
    |--train_detector.py
    |--tSNE.py
    |--utils_new.py

    |--dataloader
        |--loader.py
        |--preprocess_fmnist.py
    |--learning
        |--model_resnet.py
        |--model_wrn.py
        |--model_vgg16.py
        |--detector.py
    |--code-sdc
        ├── Automold.py
        ├── batch_generator.py
        ├── drive.py
        ├── Helpers.py
        ├── model_factory.py
        ├── models
        │   ├── abstract_model_provider.py
        │   ├── chauffeur.py
        │   ├── epoch.py
        │   ├── nvidia_dave2.py
        ├── train_self_driving_car.py
        ├── utils_train_self_driving_car.py
        ├── variational_autoencoder.py
        └── video.py
    |--configs: cmconfiguaration files
    |--tmp_data_with_neighbor (need to be created manually): consists of temporary files for doing analysis
    |--tmp_data_without_neighbor (need to be created manually): consists of temporary files for doing analysis
    |--tsne: consists of temporary files for doing analysis

Datasets and Pretrained Models (only for steering angle prediction now; need to be updated for all others)

cifar10: https://www.cs.toronto.edu/~kriz/cifar.html (CIFAR-10 python version)

svhn: http://ufldl.stanford.edu/housenumbers/ (all three files in Format2)

fmnist: https://www.kaggle.com/zalando-research/fashionmnist#fashion-mnist_test.csv

steering angle prediction: https://academictorrents.com/details/221c3c71ac0b09b1bb31698534d50168dc394cc7

Reproduction Instruction:

Preprocess data and train a classifier

1. Download each dataset online and put the data files into corresponding folder (e.g. one puts the bunch of files got from unzipping the cifar10 data zip file from its official website into the data/cifar10 folder).
2. Run

cd scripts/dataloader
python preprocess_fmnist.py

to process the fmnist dataset into a certain format if one wants to run experiment on it. 3. Train a classification model by running commands like the following (we use cifar10 and resnet here as an example):

  python train_classification_model.py --dataset=cifar10 --architecture=resnet

4. Move the trained model folder from backup/cifar10 to load/cifar10 and may need to remove the ending index if necessary e.g. if one gets the model folder with name natural_resnet_2, one needs to remove the _2 in the end before moving it into load/cifar10.

Extract feature vectors for neighbors and train/test a detector

5. Run

  python train_detector.py --dataset=cifar10 --architecture=resnet --running_mode=preprocess

to extract feature vectors for the trained resnet model on the neighbors of each of the original data point from both traing and testing set of the cifar10 dataset. 6. Run

  python train_detector.py --dataset=cifar10 --architecture=resnet --running_mode=train

to train the detector for the trained classifier on the cifar10 dataset.

7. (Table 6, Table 7, Table 9) (need to change parameters accordingly) Run

  python train_detector.py --dataset=cifar10 --architecture=resnet --running_mode=test

to test the detector for the trained classifier on the cifar10 dataset.

Do extra analysis

rq1

Reproduce results for Table 2, Table 3, Figure 5, and Figure 6 (Need to change parameters inside. See the beginning comments in rq1.py). Run:

python rq1.py

rq2-rq4

Reproduce results for Figure 7 (Need to change parameters inside. See the beginning comments in tSNE.py). Run:

python tSNE.py --dataset=cifar10 --architecture=resnet

Reproduce results for Figure 8, Table 4, Table 5, top1 baseline in Table 6 and Table 8, Table 8 (Need to change parameters inside. See the beginning comments in analysis.py). Run:

python analysis.py