Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/ytongbai/ViTs-vs-CNNs
[NeurIPS 2021]: Are Transformers More Robust Than CNNs? (Pytorch implementation & checkpoints)
https://github.com/ytongbai/ViTs-vs-CNNs
robustness transformer
Last synced: 13 days ago
JSON representation
[NeurIPS 2021]: Are Transformers More Robust Than CNNs? (Pytorch implementation & checkpoints)
- Host: GitHub
- URL: https://github.com/ytongbai/ViTs-vs-CNNs
- Owner: ytongbai
- Created: 2021-10-25T18:48:12.000Z (over 2 years ago)
- Default Branch: main
- Last Pushed: 2021-12-09T18:56:46.000Z (over 2 years ago)
- Last Synced: 2023-10-20T23:57:44.189Z (9 months ago)
- Topics: robustness, transformer
- Language: Python
- Homepage:
- Size: 1.96 MB
- Stars: 161
- Watchers: 14
- Forks: 10
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
Lists
- StarryDivineSky - ytongbai/ViTs-vs-CNNs
README
# Are Transformers More Robust Than CNNs?
Pytorch implementation for NeurIPS 2021 Paper: [Are Transformers More Robust Than CNNs?](https://arxiv.org/pdf/2111.05464.pdf)
Our implementation is based on [DeiT](https://github.com/facebookresearch/deit).
## Introduction
Transformer emerges as a powerful tool for visual recognition. In addition to demonstrating competitive performance on a broad range of visual benchmarks, recent works also argue that Transformers are much more robust than Convolutions Neural Networks (CNNs). Nonetheless, surprisingly, we find these conclusions are drawn from unfair experimental settings, where Transformers and CNNs are compared at different scales and are applied with distinct training frameworks. In this paper, we aim to provide the first fair & in-depth comparisons between Transformers and CNNs, focusing on robustness evaluations.
With our unified training setup, we first challenge the previous belief that Transformers outshine CNNs when measuring adversarial robustness. More surprisingly, we find CNNs can easily be as robust as Transformers on defending against adversarial attacks, if they properly adopt Transformers' training recipes. While regarding generalization on out-of-distribution samples, we show pre-training on (external) large-scale datasets is not a fundamental request for enabling Transformers to achieve better performance than CNNs. Moreover, our ablations suggest such stronger generalization is largely benefited by the Transformer's self-attention-like architectures per se, rather than by other training setups. We hope this work can help the community better understand and benchmark the robustness of Transformers and CNNs.
## Pretrained models
We provide both pretrained vanilla models and adversarially trained models.
### Vanilla Training
#### Main Results
| | Pretrained Model | ImageNet | ImageNet-A | ImageNet-C | Stylized-ImageNet |
| ----------- | :-------------------------------------------------------------------------------------------------: | :------: | :--------: | :--------: | :---------------: |
| Res50-Ori | [download link](https://drive.google.com/file/d/1iNEvIPKYgi1ivfL4v4dKg8j1QN5bgVpn/view?usp=sharing) | 76.9 | 3.2 | 57.9 | 8.3 |
| Res50-Align | [download link](https://drive.google.com/file/d/1SE8u1jctOM5dsbhmrcYW9Mwo3yWfNAaw/view?usp=sharing) | 76.3 | 4.5 | 55.6 | 8.2 |
| Res50-Best | [download link](https://drive.google.com/file/d/12g6Gwn-KXwLBrscjvWgjrK_kXM0twgiS/view?usp=sharing) | 75.7 | 6.3 | 52.3 | 10.8 |
| DeiT-Small | [download link](https://drive.google.com/file/d/1VorHupVJxnBOMS79gYbIhcOaL00h_KTr/view?usp=sharing) | 76.8 | 12.2 | 48.0 | 13.0 |
#### Model Size
ResNets:
- ResNets fully aligned (with DeiT's training recipe) model, denoted as `res*`:
| | Model Size | Pretrained Model | ImageNet | ImageNet-A | ImageNet-C | Stylized-ImageNet |
| -------- | :--------: | :-------------------------------------------------------------------------------------------------: | :------: | :--------: | :--------: | :---------------: |
| Res18\* | 11.69M | [download link](https://drive.google.com/file/d/1Q5gj330KoCkNShr_y9mkvFZ5sUcHtfSn/view?usp=sharing) | 67.83 | 1.92 | 64.14 | 7.92 |
| Res50\* | 25.56M | [download link](https://drive.google.com/file/d/1SE8u1jctOM5dsbhmrcYW9Mwo3yWfNAaw/view?usp=sharing) | 76.28 | 4.53 | 55.62 | 8.17 |
| Res101\* | 44.55M | [download link](https://drive.google.com/file/d/1dk640r6Y504Swhs2lUi-NiPoeNITe6ix/view?usp=sharing) | 77.97 | 8.84 | 49.19 | 11.60 |
- ResNets best model (for Out-of-Distribution (OOD) generalization), denoted as `res-best`:
| | Model Size | Pretrained Model | ImageNet | ImageNet-A | ImageNet-C | Stylized-ImageNet |
| ----------- | :--------: | :-------------------------------------------------------------------------------------------------: | :------: | :--------: | :--------: | :---------------: |
| Res18-best | 11.69M | [download link](https://drive.google.com/file/d/16mtiO-04UaIb19BFFKptlThaE7Surac4/view?usp=sharing) | 66.81 | 2.03 | 62.65 | 9.45 |
| Res50-best | 25.56M | [download link](https://drive.google.com/file/d/12g6Gwn-KXwLBrscjvWgjrK_kXM0twgiS/view?usp=sharing) | 75.74 | 6.32 | 52.25 | 10.77 |
| Res101-best | 44.55M | [download link](https://drive.google.com/file/d/13HbBPkFVijP8VClBQBCAzdKcGc6o7WZy/view?usp=sharing) | 77.83 | 11.49 | 47.35 | 13.28 |
DeiTs:
| | Model Size | Pretrained Model | ImageNet | ImageNet-A | ImageNet-C | Stylized-ImageNet |
| ---------- | :--------: | :-------------------------------------------------------------------------------------------------: | :------: | :--------: | :--------: | :---------------: |
| DeiT-Mini | 9.98M | [download link](https://drive.google.com/file/d/1FzuZP_eH2Vrb0hohIH-NUnSTuI6MG-Eg/view?usp=sharing) | 72.89 | 8.19 | 54.68 | 9.88 |
| DeiT-Small | 22.05M | [download link](https://drive.google.com/file/d/1VorHupVJxnBOMS79gYbIhcOaL00h_KTr/view?usp=sharing) | 76.82 | 12.21 | 47.99 | 12.98 |
#### Model Distillation
| | Architecture | Pretrained Model | ImageNet | ImageNet-A | ImageNet-C | Stylized-ImageNet |
| ------- | :-------------: | :-------------------------------------------------------------------------------------------------: | :------: | :--------: | :--------: | :---------------: |
| Teacher | DeiT-Small | [download link](https://drive.google.com/file/d/1iNEvIPKYgi1ivfL4v4dKg8j1QN5bgVpn/view?usp=sharing) | 76.8 | 12.2 | 48.0 | 13.0 |
| Student | Res50\*-Distill | [download link](https://drive.google.com/file/d/1MK0TQXoEAFfrhEC1Edm7FrcnemW8Zp7Y/view?usp=sharing) | 76.7 | 5.2 | 54.2 | 9.8 |
| Teacher | Res50\* | [download link](https://drive.google.com/file/d/1SE8u1jctOM5dsbhmrcYW9Mwo3yWfNAaw/view?usp=sharing) | 76.3 | 4.5 | 55.6 | 8.2 |
| Student | DeiT-S-Distill | [download link](https://drive.google.com/file/d/1IrOowURrFbdZGe_FK87_6UvZWjz2Y-9n/view?usp=sharing) | 76.2 | 10.9 | 49.3 | 11.9 |
### Adversarial Training
| | Pretrained Model | Clean Acc | PGD-100 | Auto Attack |
| ---------- | :-------------------------------------------------------------------------------------------------: | :-------: | :-----: | :---------: |
| Res50-ReLU | [download link](https://drive.google.com/file/d/1q8VxQuMWGVpFeU0OZay989dmDjrPrf3d/view?usp=sharing) | 66.77 | 32.26 | 26.41 |
| Res50-GELU | [download link](https://drive.google.com/file/d/1IPExDTAAuxIhUSYrmlweTQaKNscTw24-/view?usp=sharing) | 67.38 | 40.27 | 35.51 |
| DeiT-Small | [download link](https://drive.google.com/file/d/1U5XmAUQkSlw5Q1ZhsVriBEOQEk-bPfFU/view?usp=sharing) | 66.50 | 40.32 | 35.50 |
## Vanilla Training
### Data preparation
Download and extract ImageNet train and val images from http://image-net.org/.
The directory structure is the [standard layout](https://pytorch.org/docs/stable/torchvision/datasets.html#imagefolder) for the torchvision, and the training and validation data is expected to be in the `train` folder and `val` folder respectively:
```
/path/to/imagenet/
train/
class1/
img1.jpeg
class2/
img2.jpeg
val/
class1/
img3.jpeg
class/2
img4.jpeg
```
### Environment
Install dependencies:
```bash
pip3 install -r requirements.txt
```
### Training Scripts
To train a ResNet model on ImageNet run:
```bash
bash script/res.sh
```
To train a DeiT model on ImageNet run:
```bash
bash script/deit.sh
```
## Generalization to Out-of-Distribution Sample
### Data Preparation
Download and extract [ImageNet-A](https://github.com/hendrycks/natural-adv-examples), [ImageNet-C](https://github.com/hendrycks/robustness), [Stylized-ImageNet](https://github.com/rgeirhos/Stylized-ImageNet) val images:
```
/path/to/datasets/
val/
class1/
img1.jpeg
class/2
img2.jpeg
```
### Evaluation Scripts
To evaluate pre-trained models, run:
```bash
bash script/generation_to_ood.sh
```
It is worth noting that for ImageNet-C evaluation, the error rate is calculated based on the Noise, Blur, Weather and Digital categories.
## Adversarial Training
To perform adversarial training on ResNet run:
```bash
bash script/advres.sh
```
To do adversarial training on DeiT run:
```sh
bash scripts/advdeit.sh
```
## Robustness to Adversarial Example
### PGD Attack Evaluation
To evaluate the pre-trained models, run:
```bash
bash script/eval_advtraining.sh
```
### AutoAttack Evaluation
`./autoattack` contains the [AutoAttack](https://github.com/fra31/auto-attack) public package, with a little modification to best support ImageNet evaluation.
```sh
cd autoattack/
bash autoattack.sh
```
### Patch Attack Evaluation
Please refer to [PatchAttack](https://github.com/Chenglin-Yang/PatchAttack)
## Citation
If you use our code, models or wish to refer to our results, please use the following BibTex entry:
```bibtex
@inproceedings{bai2021transformers,
title = {Are Transformers More Robust Than CNNs?},
author = {Bai, Yutong and Mei, Jieru and Yuille, Alan and Xie, Cihang},
booktitle = {Thirty-Fifth Conference on Neural Information Processing Systems},
year = {2021},
}
```