FourierHeatmap (latest release: v0.2.0)

This is an unofficial pytorch implementation of Fourier Heat Map which is proposed in the paper, A Fourier Perspective on Model Robustness in Computer Vision [Yin+, NeurIPS2019].

Fourier Heat Map allows to investigate the sensitivity of CNNs to high and low frequency corruptions via a perturbation analysis in the Fourier domain.

News

• We release v0.2.0. API is renewed and some useful libraries (e.g. hydra) are added.

• Previous version is still available as v0.1.0.

• Docker is supported. Now, you can evaluate Fourier Heat Map on the Docker container.

Requirements

This library requires following as a pre-requisite.

• python 3.9+
• poetry

Note that I run the code with Ubuntu 20, Pytorch 1.8.1, CUDA 11.0.

Installation

This repo uses poetry as a package manager. The following code will install all necessary libraries under .venv/.

$ git clone git@github.com:gatheluck/FourierHeatmap.git
$ cd FourierHeatmap
$ pip install poetry  # If you haven't installed poetry yet.
$ poetry install

Setup

Dataset

This codes expect datasets exist under data/. For example, if you want to evaluate Fourier Heat Map for ImageNet, please set up like follows:

FourierHeatmap
├── data
│	└── imagenet
│		├── train/
│		└── val/

Usage

Visualizing Fourier basis

The script fhmap/fourier/basis.py generates Fourier base functions. For example:

$ poetry run python fhmap/fourier/basis.py

will generate 31x31 2D Fourier basis and save as an image under outputs/basis.png. The generated image should be like follows.

Evaluating Fourier Heat Map

The script fhmap/apps/eval_fhmap.py eveluate Fourier Heat Map for a model. For example:

$ poetry run python fhmap/apps/eval_fhmap.py dataset=cifar10 arch=resnet56 weightpath=[PYTORCH_MODEL_WEIGHT_PATH] eps=4.0

will generate 31x31 Fourier Heat Map for ResNet56 on CIFAR-10 dataset and save as an image under outputs/eval_fhmap/. The generated image should be like follows.

Note that the L2 norm size (=eps) of Fourier basis use in original paper is following:

dataset	eps
CIFAR-10	4.0
ImageNet	15.7

Evaluating custom dataset and model

Evaluating your custom dataset

If you want to evaluate Fourier Heat Map on your custom dataset, please refer follwing instraction.

• Implement YourCustomDatasetStats class:

  • This class holds basic dataset information.
  • YourCustomDatasetStats class should inherit from original DatasetStats class in factory/dataset module and also shoud be placed in factory/dataset module.
  • For details, please refer to the Cifar10Stats class in factory/dataset module.

• Implement YourCustomDataModule class:

  • This class is responsible for preprocess, transform (includes adding Fourier Noise to image) and create test dataset.
  • YourCustomDataModule class should inherit from BaseDataModule class in factory/dataset module and also shoud be placed in factory/dataset module.
  • For details, please refer to the Cifar10DataModule class in factory/dataset module.

• Implement YourCustomDatasetConfig class:

  • This class is needed for applying hydra's dynamic object instantiation to dataset class.
  • YourCustomDatasetConfig class should inherit from DatasetConfig class in schema/dataset module and also shoud be placed in schema/dataset module. Please add YourCustomDatasetConfig to schema/__init__.
  • For details, please refer to the Cifar10Config class in schema/dataset module.

• Add option for your custom dataset:

  • Lastly, please add the config of your custom dataset to ConfigStore class by adding a follwing line to apps/eval_fhmap.

  cs.store(group="dataset", name="yourcustomdataset", node=schema.YourCustomDatasetConfig)
  

Now, you will be able to call your custom dataset like following.

$ poetry run python fhmap/apps/eval_fhmap.py dataset=yourcustomdataset arch=resnet50 weightpath=[PYTORCH_MODEL_WEIGHT_PATH] eps=4.0

Evaluating your custom architecture (model)

If you want to evaluate Fourier Heat Map on your custom architecture (model), please refer follwing instraction.

• Implement YourCustomArch class:

  • Please implement class or function which return your custom architecture. The custom architecture have to subclass of torch.nn.module.
  • For details, please refer to the factory/archs/resnet module.

• Implement YourCustomArchConfig class:

  • This class is needed for applying hydra's dynamic object instantiation to architecture class.
  • YourCustomArchConfig class should inherit from ArchConfig class in schema/arch module and also shoud be placed in schema/arch module. Please add YourCustomArchConfig to schema/__init__.
  • For details, please refer to the Resnet56Config class in schema/arch module.
  • If you want to use architectures which is provided by other libs like pytorch or timm, please refere to the Resnet50Config class in schema/arch module.

• Add option for your custom architecture:

  • Lastly, please add the config of your custom architecture to ConfigStore class by adding a follwing line to apps/eval_fhmap.

  cs.store(group="arch", name="yourcustomarch", node=schema.YourCustomArchConfig)
  

Now, you will be able to call your custom arch like following.

$ poetry run python fhmap/apps/eval_fhmap.py dataset=cifar10 arch=yourcustomarch weightpath=[PYTORCH_MODEL_WEIGHT_PATH] eps=4.0

Evaluating Fourier Heat Map through Docker

In order to use FourierHeatmap throgh docker, please install Docker with NVIDIA Container Toolkit beforehand. For detail, please refere official installation guide.

If nvidia-smi is able to run through docker like following, it is successfully installed.

$ sudo docker run --rm --gpus all nvidia/cuda:11.0-base nvidia-smi

Tue Apr 27 06:46:09 2021       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.102.04   Driver Version: 450.102.04   CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  GeForce GTX 1080    Off  | 00000000:01:00.0  On |                  N/A |
| N/A   56C    P0    42W /  N/A |   1809MiB /  8114MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
+-----------------------------------------------------------------------------+

We use environmental variables to specify the arguments. The variables that can be specified and their meanings are as follows:

name	optional	default	description
HOST_DATADIR	False		Path to the directory where the dataset is located in the host.
HOST_OUTPUTSDIR	False		Path to the directory where the output will be located in the host.
HOST_WEIGHTDIR	False		Path to the directory where the pretrained wight is located in the host.
WEIGHTFILE	False		File name of the pretrained wight.
ARCH	True	resnet56	Name of the architecture.
BATCH_SIZE	True	512	Size of batch.
DATASET	True	cifar10	Name of dataset.
EPS	True	4.0	L2 norm size of Fourier basis.
IGNORE_EDGE_SIZE	True	0	Size of the edge to ignore.
NUM_SAMPLES	True	-1	Number of samples used from dataset. If -1, use all samples.
NVIDIA_VISIBLE_DEVICES	True	0	Device number (or list of number) visible from CUDA.

For example:

$ export HOST_DATADIR=[DATASET_DIRECTORY_PATH]
$ export HOST_OUTPUTSDIR=[OUTPUTS_DIRECTORY_PATH]
$ export HOST_WEIGHTDIR=[WEIGHT_DIRECTORY_PATH]
$ export WEIGHTFILE=[PYTORCH_MODEL_FILE]
$ cd provision/docker
$ sudo -E docker-compose up  # -E option is needed to inherit environment variables.

will generate 31x31 Fourier Heat Map for ResNet56 on CIFAR-10 dataset and save as an image under OUTPUTS_DIRECTORY_PATH/eval_fhmap/.

References

• Dong Yin, Raphael Gontijo Lopes, Jonathon Shlens, Ekin D. Cubuk, Justin Gilmer. "A Fourier Perspective on Model Robustness in Computer Vision.", in NeurIPS, 2019.
• Justin Gilmer and Dan Hendrycks. "Adversarial Example Researchers Need to Expand What is Meant by ‘Robustness’.", in Distill, 2019.