Improving MMD-GAN Training with Repulsive Loss Function
Authors: Wei Wang, Yuan Sun, Saman Halgamuge
ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The proposed methods are applied to the unsupervised image generation tasks on CIFAR-10, STL-10, Celeb A, and LSUN bedroom datasets. Results show that the repulsive loss function significantly improves over the MMD loss at no additional computational cost and outperforms other representative loss functions. The proposed methods achieve an FID score of 16.21 on the CIFAR-10 dataset using a single DCGAN network and spectral normalization. |
| Researcher Affiliation | Academia | Wei Wang University of Melbourne Yuan Sun RMIT University Saman Halgamuge University of Melbourne |
| Pseudocode | No | The paper does not contain any pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at: https://github.com/richardwth/MMD-GAN |
| Open Datasets | Yes | Dataset: The loss functions were evaluated on four datasets: 1) CIFAR-10 (50K images, 32 × 32 pixels) (Krizhevsky & Hinton (2009)); 2) STL-10 (100K images, 48 × 48 pixels) (Coates et al. (2011)); 3) Celeb A (about 203K images, 64 × 64 pixels) (Liu et al. (2015)); and 4) LSUN bedrooms (around 3 million images, 64 × 64 pixels) (Yu et al. (2015)). |
| Dataset Splits | No | The paper states 'All models were trained for 100K iterations' and refers to evaluation metrics calculated using 'randomly generated samples' and 'real samples', but it does not specify explicit training, validation, and test dataset splits or percentages for the datasets used in training. |
| Hardware Specification | No | The paper mentions 'LIEF HPC-GPGPU Facility hosted at the University of Melbourne' but does not provide specific details on the GPU models, CPU models, or memory specifications used for the experiments. |
| Software Dependencies | No | The paper mentions 'Adam optimizer' and 'spectral normalization' but does not provide specific version numbers for any software, libraries, or frameworks used. |
| Experiment Setup | Yes | We used Adam optimizer (Kingma & Ba (2015)) with momentum parameters β1 = 0.5, β2 = 0.999; two-timescale update rule (TTUR) (Heusel et al. (2017)) with two learning rates (ρD, ρG) chosen from {1e-4, 2e-4, 5e-4, 1e-3} (16 combinations in total); and batch size 64. All models were trained for 100K iterations on CIFAR-10, STL-10, Celeb A and LSUN bedroom datasets, with ndis = 1, i.e., one discriminator update per generator update. |