Improving GAN with Neighbors Embedding and Gradient Matching

Authors: Ngoc-Trung Tran, Tuan-Anh Bui, Ngai-Man Cheung5191-5198

AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform experiments to demonstrate that our proposed techniques are computationally simple and easy to be incorporated in existing systems. When Gradient matching and Neighbour embedding are applied together, our GN-GAN achieves outstanding results on 1D/2D synthetic, CIFAR-10 and STL-10 datasets, e.g. FID score of 30.80 for the STL-10 dataset.
Researcher Affiliation Collaboration Ngoc-Trung Tran, Tuan-Anh Bui, Ngai-Man Cheung ST Electronics SUTD Cyber Security Laboratory Singapore University of Technology and Design
Pseudocode Yes Algorithm 1 Our GN-GAN model
Open Source Code Yes Our code is available at: https://github.com/tntrung/gan
Open Datasets Yes When Gradient matching and Neighbour embedding are applied together, our GN-GAN achieves outstanding results on 1D/2D synthetic, CIFAR-10 and STL-10 datasets
Dataset Splits No The paper discusses training data and generated samples for evaluation (e.g., '10K real samples and 5K generated samples' for FID), but does not explicitly provide percentages or counts for a distinct validation dataset split from the original data that would be used during training for hyperparameter tuning or early stopping.
Hardware Specification Yes To examine the computational time of gradient matching of our proposed generator objective, we measure its training time for one mini-batch (size 64) with/without GM (Computer: Intel Xeon Octa-core CPU E5-1260 3.7GHz, 64GB RAM, GPU Nvidia 1080Ti) with CNN for CIFAR-10.
Software Dependencies No Tensorflow is mentioned as the framework for computation ('can be computed easily in Tensorflow'), but no specific version number for Tensorflow or any other software dependencies is provided.
Experiment Setup Yes The number of neurons for each hidden layer is 4, the learning rate is 0.001, λp = 0.1 for both method, λ1 m = λ2 m = 0.1 for our generator objective. We use Adam optimizer with learning rate lr = 0.001, and the exponent decay rate of first moment β1 = 0.8. The parameters of our model are: λp = 0.1, λ1 m = λ2 m = 0.1. The learning rate is decayed every 10K steps with a base of 0.99. The mini-batch size is 128. The training stops after 500 epochs. Our default parameters are used for all experiments λp = 1.0, λr = 1.0, λ1 m = λ2 m = 1.0. Learning rate, β1, β2 for Adam is (lr = 0.0002, β1 = 0.5, β2 = 0.9). The generator is trained with 350K updates for logarithm loss version (Eq. 7) and 200K for hinge loss version (Eq. 8) to converge better.