dwromero / att_gconvs Goto Github PK

This repository contains the source code accompanying the paper:

Attentive Group Equivariant Convolutional Networks
David W. Romero, Erik J. Bekkers, Jakub M. Tomczak & Mark Hoogendoorn, ICML 2020.

Although group convolutional networks are able to learn powerful representations based on symmetry patterns, they lack explicit means to learn meaningful relationships among them (e.g., relative positions and poses). In this paper, we present attentive group equivariant convolutions, a generalization of the group convolution, in which attention is applied during the course of convolution to accentuate meaningful symmetry combinations and suppress non-plausible, misleading ones. We indicate that prior work on visual attention can be described as special cases of our proposed framework and show empirically that our attentive group equivariant convolutional networks consistently outperform conventional group convolutional networks on benchmark image datasets. Simultaneously, we provide interpretability to the learned concepts through the visualization of equivariant attention maps.

The folder structure is as follows:

• attgconv contains the main PyTorch library.

• demo includes some short jupyter notebook demo's on how to use the the code.

• experiments contains the experiments described in the paper.

This code as based on PyTorch and has been tested with the following library versions:

• torch==1.4.0

• numpy==1.17.4

• scipy==1.3.2

• matplotlib==3.1.1

• jupyter==1.0.0

The exact specification of our environment is provided in the file environment.yml. An appropriate environment can be easily created via:

conda env create -f environment.yml

or constructed manually with conda via:

conda create --yes --name torch
conda activate torch
# Please check your cudatoolkit version and replace it in the following line
conda install conda install pytorch==1.4.0 torchvision==0.5.0 cudatoolkit=10.1 -c pytorch
conda install numpy==1.17.4 scipy==1.3.2 matplotlib==3.1.1 jupyter==1.0.0 --yes

For the sake of reproducibility, we provide the parameters used in the corresponding baselines hardcoded by default. If you wish to vary these parameters for your own experiments, please modify the corresponding parser.py file in the experiment folder and erase the hard-coded values from the run_experiment.py file.

Remark on Logger in experiments. We provide a logger system that automatically saves any print during the execution of the program into a file named saved/foldername/modellog_i.out. The logger object is created right before the training starts in the run_experiment.py file (line sys.stdout = Logger(args)). We recommend users working in an slurm environment to comment this line, as it will otherwise impede writing into the corresponding slurm_####.out file. Hence, you won't be able to see any prints in the slurm_####.out file. Deactivated here by default.

We provide some pretrained models from our experiments for easy reproducibility. To use these models, utilize the keyword --pretrained and make sure the training parameters as well as the additional --extra_comment argument correspond to those given in the folder name.

The utilized datasets have been uploaded to a repository for reproducibility. Please extract the files in the corresponding experiments/experiment_i/data folder. For our experiments in CIFAR-10, we make use of the dataset provided in torchvision.

Rot-MNIST:

The dataset can be downloaded from: https://drive.google.com/file/d/1PcPdBOyImivBz3IMYopIizGvJOnfgXGD/view?usp=sharing

PCAM:

We use an ImageFolder structure for our experiments. A file containing the entire dataset in this format can be downloaded from: https://drive.google.com/file/d/1THSEUCO3zg74NKf_eb3ysKiiq2182iMH/view?usp=sharing

Code used to transform the .h5 dataset to this format is provided in experiments/pcam/data/.

If you found this work useful in your research, please consider citing:

@article{romero2020attentive,
  title={Attentive Group Equivariant Convolutional Networks},
  author={Romero, David W and Bekkers, Erik J and Tomczak, Jakub M and Hoogendoorn, Mark},
  journal={arXiv preprint arXiv:2002.03830},
  year={2020}
}

The code and scripts in this repository are distributed under MIT license. See LICENSE file.