GenFormer - Generated Images are All You Need to Improve Robustness of Transformers on Small Datasets
Official repository of the ICPR24 paper "GenFormer - Generated Images are All You Need to Improve Robustness of Transformers on Small Datasets"
Sven Oehri, Nikolas Ebert, Ahmed Abdullah, Didier Stricker & Oliver WasenmΓΌller
CeMOS - Research and Transfer Center, University of Applied Sciences Mannheim
- Installation
- Main Results and Models
- Data preparation
- Evaluation
- Training
- Data Generation
- Acknowledgements
- Citing
The code was tested with CUDA 11.8, Python 3.10 PyTorch 2.2.0. The code was not optimized for multi-GPU setups and various models only support 1 GPU.
git clone https://github.com/CeMOS-IS/GenFormer
cd GenFormer
To set up an Anaconda environment, run the following commands from:
conda create -n genformer python=3.10
conda activate genformer
pip install torch torchvision torchaudio
pip install -r requirements.txt
pip install robust-minisets
python3 setup.py develop
To build a docker image, run the following
docker build -t genformer .
Adjust visible GPUs, shared memory size and mounted local directory according to your setup, then run the container.:
docker run -it --name genformer --rm \
--shm-size 100G --gpus '"device=0"' \
-v /path/to/repo/:/genformer \
genformer bash
Results and Models on Tiny ImagNet
Method | Train-Strategy | Gen. Images | T-IN Err. | T-INv2 Err. | T-IN-R-Err. | T-IN-C CE | T-IN-A Err. | model |
---|---|---|---|---|---|---|---|---|
DeiT-Ti | Baseline | False | 50.3 | 68.0 | 92.5 | 80.6 | 80.6 | model |
DeiT-Ti | Baseline | True | 44.1 | 65.3 | 89.6 | 77.7 | 78.0 | model |
DeiT-Ti | Cutmix + MixUp | False | 44.4 | 65.0 | 89.7 | 74.5 | 78.3 | model |
DeiT-Ti | Cutmix + MixUp | True | 38.5 | 57.8 | 86.4 | 71.8 | 74.4 | model |
DeiT-Ti | AugMix | False | 40.4 | 61.3 | 88.4 | 72.8 | 76.2 | model |
DeiT-Ti | AugMix | True | 36.4 | 57.8 | 85.6 | 71.0 | 73.1 | model |
PVT-T | Baseline | False | 46.5 | 67.2 | 91.5 | 78.9 | 78.9 | model |
PVT-T | Baseline | True | 42.8 | 64.8 | 87.4 | 76.8 | 77.7 | model |
PVT-T | Cutmix + MixUp | False | 41.9 | 64.6 | 87.8 | 74.1 | 76.8 | model |
PVT-T | Cutmix + MixUp | True | 37.5 | 60.1 | 84.6 | 70.1 | 74.1 | model |
PVT-T | AugMix | False | 39.9 | 62.1 | 87.9 | 73.0 | 75.9 | model |
PVT-T | AugMix | True | 36.4 | 58.3 | 86.2 | 71.0 | 73.6 | model |
Results and Models on CIFAR-10 / CIFAR-100
Method | Gen. Images | C-10 Err. | C-10.1 Err. | C-10-C CE | C-100-C Err | C-100-C CE | model |
---|---|---|---|---|---|---|---|
ResNet18 | False | 4.6 | 11.2 | 15.1 | 20.5 | 37.9 | model-C10 / model-C100 |
ResNet18 | True | 4.1 | 10.1 | 14.6 | 20.3 | 38.0 | model-C10 / model-C100 |
DeiT-Ti | False | 10.5 | 22.2 | 23.0 | 35.3 | 51.8 | model-C10 / model-C100 |
DeiT-Ti | True | 4.0 | 9.8 | 12.0 | 24.7 | 39.4 | model-C10 / model-C100 |
PVT-T | False | 6.9 | 14.8 | 17.8 | 29.7 | 48.0 | model-C10 / model-C100 |
PVT-T | True | 3.6 | 9.7 | 11.7 | 21.1 | 35.9 | model-C10 / model-C100 |
PVTv2-B0 | False | 5.0 | 11.1 | 14.2 | 23.1 | 41.2 | model-C10 / model-C100 |
PVTv2-B0 | True | 3.5 | 8.8 | 12.6 | 19.5 | 34.2 | model-C10 / model-C100 |
ConViT | False | 5.6 | 13.2 | 14.0 | 25.5 | 40.6 | model-C10 / model-C100 |
ConViT | True | 3.2 | 7.8 | 9.6 | 18.2 | 31.5 | model-C10 / model-C100 |
Results and Models on MedMNIST
Method | Gen. Images | B-MNIST Err. | B-MNIST-C CE | P-MNIST Err. | P-MNIST-C CE | OS-MNIST Err. | OS-MNIST-C CE | model |
---|---|---|---|---|---|---|---|---|
DeiT-Ti | False | 23.1 | 24.5 | 8.0 | 12.8 | 26.9 | 45.9 | model-B / model-P / model-OS |
DeiT-Ti | True | 9.0 | 14.7 | 5.1 | 10.5 | 21.3 | 32.5 | model-B / model-P / model-OS |
PVT-T | False | 22.4 | 23.9 | 13.1 | 17.8 | 20.3 | 34.4 | model-B / model-P / model-OS |
PVT-T | True | 9.0 | 13.9 | 5.4 | 10.0 | 17.1 | 24.7 | model-B / model-P / model-OS |
- For evaluation, the robustness test sets are automatically generated and downloaded using the robust_minisets library. No further customization is required.
- For training with and without generated data, the
train
andval
sets will be downloaded and prepared automatically without any further customization needed. - For training with generated data, ensure that all generated data is stored in the
data/
folder. The folder structure for all data sets should be as follows:
$ tree data
generated dataset
βββ class1
β βββ img1.jpeg
β βββ img2.jpeg
β βββ ...
βββ class2
β βββ img3.jpeg
β βββ ...
βββ ...
Note: The folder of the synthetic images must be specified in the config
-file under SYNTHETIC_DATA.DIR_NAME
(e.g. configs/genformer/vanilla_genformer/deit_vanilla_tinyim_genformer.yaml
line 24 to 26)
Note: More information on how to create the generated datasets and the necessary models can be found in Section Data Generation.
To evaluate a pre-trained Model on Tiny ImageNet
, Tiny ImageNet-C
, Tiny ImageNet-A
and Tiny ImageNet-R
run:
python3 tools/eval.py --cfg path/to/config.yaml --ckpt path/to/weights.pth -d dataset
For example, to evaluate the DeiT-Tiny with on Tiny ImageNet-R
:
python3 tools/eval.py --cfg configs/baseline/vanilla/deit_vanilla_tinyim.yaml --ckpt checkpoints/deit_tinyimagenet_vanilla.pth -d tinyimagenet-r
In addition to Tiny-ImagNet
, numerous other datasets such as CIFAR
, MedMNIST
and EuroSAT
can be evaluated. A list of all supported datasets can be found with the following command:
python3 tools/eval.py -h
To train a Model on a dataset of choice from scratch, run:
python3 tools/train.py --cfg path/to/config.yaml
Options:
--resume
: loads the last checkpoint of the modelopts
: Manipulate config options
e.g. Deit-Tiny
on Tiny ImageNet
w/o GenFormer
:
python3 tools/train.py --cfg configs/baseline/vanilla/deit_vanilla_tinyim.yaml
e.g. Deit-Tiny
on Tiny ImageNet
w/ GenFormer
:
python3 tools/train.py --cfg configs/genformer/vanilla_genformer/deit_vanilla_tinyim_genformer.yaml
e.g. Deit-Tiny
on Tiny ImageNet
w/ GenFormer
an wandB
-logging modified via command line:
python3 tools/train.py --cfg configs/genformer/vanilla_genformer/deit_vanilla_tinyim_genformer.yaml LOG.WANDB True
To adjust all training parameters, the corresponding .yaml
-file in the configs/
-folder must be adjusted. An overview of all options can be found in genformer/engine/cfg.py
.
We utilized the code from Elucidating the Design Space of Diffusion-Based Generative Models (EDM) to create the generated datasets. Here, we are only providing our trained models used for data generation. For more information on the use of the diffusion model, please refer to the authors' documentation.
Natural Images
MedMNIST
Aerial Images
This repository is built using the timm library and the tiny-transformers repository.
This research was partly funded by Albert and Anneliese Konanz Foundation, the German Research Foundation under grant INST874/9-1 and the Federal Ministry of Education and Research Germany in the project M2Aind-DeepLearning (13FH8I08IA).
If you have used GenFormer in your research, please cite our work. π
@inproceedings{oehri2024genformer,
title = {GenFormer β Generated Images are All You Need to Improve Robustness of Transformers on Small Datasets},
author = {Oehri, Sven and Ebert, Nikolas and Abdullah, Ahmed and Stricker, Didier and Wasenm{\"u}ller, Oliver},
booktitle = {International Conference on Pattern Recognition (ICPR)},
year = {2024},
}