HI2I: Hypercomplex Image-to-Image Translation

Eleonora Grassucci, Luigi Sigillo, Aurelio Uncini, and Danilo Comminiello

The official repository is: ispamm/HI2I.

Abstract 📑

Image-to-image translation (I2I) aims at transferring the content representation from an input domain to an output one, bouncing along different target domains. Recent I2I generative models which gain outstanding results in this task comprise a set of diverse deep networks each with tens of million parameters. Moreover, images are usually three-dimensional being composed of RGB channels and common neural models do not take dimensions correlation into account, losing beneficial information. In this paper, we propose to leverage hypercomplex algebra properties to define lightweight I2I generative models capable of preserving pre-existing relations among images dimensions, thus exploiting additional input information. On manifold I2I benchmarks, we show how the proposed Quaternion StarGANv2 and parameterized hypercomplex StarGANv2 (PHStarGANv2) reduce parameters and storage memory amount while ensuring high domain translation performance and good image quality as measured by FID and LPIPS score.

Model Architecture (from StarGANv2) 🎬

Results 📊

Model	Params	Storage Mem	Savings	FID Reference	LPIPS Reference	FID Latent	LPIPS Latent	Checkpoint
StarGANv2	87M	307MB	0%	21.24	0.24	17.16	0.25	Link
Quaternion StarGANv2	22M	76MB	75%	23.09	0.22	27.90	0.12	Link
PHStarGANv2 n=3	29M	137MB	67%	28.11	0.29	16.63	0.33	Link
PHStarGANv2 n=4	22M	76MB	75%	24.33	0.27	16.54	0.29	Link

How to run experiments 💻

First, please install the requirements:

pip install -r requirements.txt

We upload a sample config.json, to run different experiments please edit this file, i.e:

• datasets can be either afhq and celeba_hq.
• num_domains depends on the number of domains of the chosen dataset: 3 for afhq, 2 for celeba_hq.
• w_hpf is set to 0 for afhq since it does not employ the support network, while is set to 1 for celeba_hq.
• lambda_ds is equal to 2 for afhq and 1 for celeba_hq.
• phm=True for PHStarGANv2, False for running the real-valued baseline.
• N is the hyperparameter for parameterized hypercomplex layers, we test N=2,3,4.
• seed=777 is to reproduce our experiments.

Then, run python main.py. The experiment will be directly tracked on Weight&Biases (we strongly suggest to use W&B!).

Experiments require approximately 80 hours on a single NVIDIA Tesla V100-32GB GPU. For a fast training one can edit the code involving torch.nn.parallel.DistributedDataParallel employing more than one GPU.

Cite

Please cite our work if you found it useful:

@inproceedigns{grassucci2022HI2I,
      title={Hypercomplex Image-to-Image Translation}, 
      author={Grassucci, Eleonora and Sigillo, Luigi and Uncini, Aurelio and Comminiello, Danilo},
      year={2022},
      month={July},
      booktitle={{IEEE} International Joint Conference on Neural Networks (IJCNN)}
}

Interested in Quaternion and Hypercomplex Generative Models?

Check also:

• Lightweight Convolutional Neural Network by Hypercomplex Parameterization, Under Review, 2021 [Paper] [GitHub].
• Quaternion-Valued Variational Autoencoder, ICASSP, 2021 [Paper] [GitHub].
• An Information-Theoretic Perspective on Proper Quaternion Variational Autoencoders, Entropy, 2021 [Paper] [GitHub].
• Quaternion Generative Adversarial Networks, Generative Adversarial Learning: Architectures and Applications, editors: Dr Roozbeh Razavi-Far, Dr Ariel Ruiz-Garcia, Professor Vasile Palade, Professor Jürgen Schmidhuber, Springer, Jan 2022. [Paper][GitHub].