Regularization of polynomial networks for image recognition

Official implementation of the image classification experiments in the CVPR'23 paper "**Regularization of polynomial networks for image recognition**" .

Browsing the folders

The folder structure is the following:

• models: The folder contains the neural network architectures of R-PolyNets, D-PolyNets, R-PDC and D-PDC.
• configs: The folder contains the yml files for the configuration, e.g., epochs to run, learning rate changes. Most of the options for hyper-parameters can be changed here and are propagated to the network. For instance, you can change the dataset from CIFAR10 to CIFAR100 by changing the respective name in yml.
• utils: Misc functions required for training.

Train the network

To train the network for image classification, you can execute the following command:

python train_main_label_smooth.py --config configs/R_PolyNets_no_activation_functions_cifar10.yml --label any-name-you-want-as-label

To train the network for audio classification, you can execute the following command:

python train_main_speech_label_smooth.py --config configs/R_PolyNets_no_activation_functions_speech.yml --label any-name-you-want-as-label

You can choose any yml file in utils.

Changing the dataset: You can change the dataset by changing the name of the dataset/db field (in the yml). The datasets that exist by default in PyTorch, e.g., CIFAR10/CIFAR100/MNIST, are automatically downloaded if they do not exist. They are exported in the path dataset/root.

Changing the model: You have several options to change the model, e.g., by changing a) the model/name (in the yml) to the model you want, or b) by changing the model/name and specify the number of of blocks with the argument num_blocks. Overall, to change the model, the best way to do this is to modify the yml arguments.

Package Dependencies

Apart from PyTorch, we use a number of standard packages. For instance, pandas, pyyaml and also dropblock. All of those can be install with the pip install command.

Citing

If you use this code, please cite [1]:

BibTeX:

@inproceedings{chrysos2023regularization,
   title={Regularization of polynomial networks for image recognition},
   author={Chrysos, Grigorios and Wang, Bohan and Deng, Jiankang and Cevher, Volkan},
   booktitle={Conference on Computer Vision and Pattern Recognition (CVPR)},
   year={2023}
}

References

[1]	Grigorios G. Chrysos, Bohan Wang, Jiankang Deng, and Volkan Cevher, Regularization of polynomial networks for image recognition, Conference on Computer Vision and Pattern Recognition (CVPR), 2023.