Repository files navigation

Hypercomplex Neural Networks with PyTorch: this repository would be a container for hypercomplex neural network modules to facilitate research in this topic.

⚠️ PHC layer for 3D convolutions has been released! Check it in the filelayers/ph_layers.py!

⚠️ PHC layer for 1D convolutions has been released! Check it in the filelayers/ph_layers.py!Thanks to Christian Brignone and Gioia Mancini for this contribution.

[ArXiv preprint][IEEEXplore]

Hypercomplex neural networks have proved to reduce the overall number of parameters while ensuring valuable performances by leveraging the properties of Clifford algebras. Recently, hypercomplex linear layers have been further improved by involving efficient parameterized Kronecker products. In this paper, we define the parameterization of hypercomplex convolutional layers to develop lightweight and efficient large-scale convolutional models. Our method grasps the convolution rules and the filters organization directly from data without requiring a rigidly predefined domain structure to follow. The proposed approach is flexible to operate in any user-defined or tuned domain, from 1D to nD regardless of whether the algebra rules are preset.Such a malleability allows processing multidimensional inputs in their natural domain without annexing further dimensions, as done, instead, in quaternion neural networks for 3D inputs like color images.As a result, the proposed method operates with 1/n free parameters as regards its analog in the real domain. We demonstrate the versatility of this approach to multiple domains of application by performing experiments on various image datasets as well as audio datasets in which our method outperforms real and quaternion-valued counterparts.

The core of the approach is the sum of Kronecker products which grasps the convolution rule and the filters organization directly from data. The higlights of our approach is defined in:

defkronecker_product1(self,A,F):siz1=torch.Size(torch.tensor(A.shape[-2:])*torch.tensor(F.shape[-4:-2]))siz2=torch.Size(torch.tensor(F.shape[-2:]))res=A.unsqueeze(-1).unsqueeze(-3).unsqueeze(-1).unsqueeze(-1)*F.unsqueeze(-4).unsqueeze(-6)siz0=res.shape[:1]out=res.reshape(siz0+siz1+siz2)returnoutdefforward(self,input):self.weight=torch.sum(self.kronecker_product1(self.A,self.F),dim=0)input=input.type(dtype=self.weight.type())returnF.conv2d(input,weight=self.weight,stride=self.stride,padding=self.padding)

Te PHC layer, by setting n=4, is able to subsume the Hamilton rule to organize filters in the convolution as:

The foldertutorials contains a set of tutorials to understand the Parameterized Hypercomplex Multiplication (PHM) layer and the Parameterized Hypercomplex Convolutional (PHC) layer. We develop simple toy examples to learn the matrices A that define algebra rules in order to demonstrate the effectiveness of the proposed approach.

• PHM tutorial.ipynb is a simple tutorial which shows how the PHM layer learns the Hamilton product between two pure quaternions.
• PHC tutorial.ipynb is a simple tutorial which shows how the PHC layer learn the Hamilton rule to organize filters in convolution.
• Toy regression examples with PHM.ipynb is a notebook containing some regression tasks.

To reproduce image classification experiments, please refer to theimage-classification folder.

• pip install -r requirements.txt.
• Choose the configurations inconfigs and run the experiment:

python main.py --TextArgs=config_name.txt.

The experiment will be directly tracked onWeight&Biases.

To reproduce sound event detection experiments, please refer to thesound-event-detection folder.

• pip install -r requirements.txt.

We follow the instructions in the original repository for theL3DAS21 dataset:

• Download the dataset:

python download_dataset.py --task Task2 --set_type train --output_path DATASETS/Task2

python download_dataset.py --task Task2 --set_type dev --output_path DATASETS/Task2

• Preprocess the dataset:

python preprocessing.py --task 2 --input_path DATASETS/Task2 --num_mics 1 --frame_len 100

Specifynum_mics=2 andoutput_phase=True to perform experiments up to 16-channel inputs.

• Run the experiment:

python train_baseline_task2.py

Specify the hyperparameters options.We perform experiments withepochs=1000,batch_size=16 andinput_channels=4/8/16 on a single Tesla V100-32GB GPU.

• Run the evaluation:

python evaluate_baseline_task2.py

Specify the hyperparameters options.

⚠️ Work in progress...

• PHC for transposed convolutions
• PHC in TensorFlow

Please cite our work if you found it useful:

@article{grassucci2021PHNNs,      title={{PHNN}s: Lightweight Neural Networks via Parameterized Hypercomplex Convolutions},       author={Grassucci, E. and Zhang, A. and Comminiello, D.},      year={2022},      journal={IEEE Transactions on Neural Networks and Learning Systems},       year={2022},      pages={1-13},      doi={10.1109/TNNLS.2022.3226772}}

• Quaternion Generative Adversarial Networks.
• A Quaternion-Valued Variational Autoencoder.
• PHC-GNN.
• Compacter.

Quaternion layers are borrowed from:

• Quaternion-Neural-Networks.