Recursive Reasoning in Minimax Games: A Level $k$ Gradient Play Method

Authors: Zichu Liu, Lacra Pavel

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental By combining Lv.k GP with Adam optimizer, our algorithm shows a clear advantage in terms of performance and computational overhead compared to other methods. Using a single Nvidia RTX3090 GPU and 30 times fewer parameters than Big GAN on CIFAR-10, we achieve an FID of 10.17 for unconditional image generation within 30 hours, allowing GAN training on common computational resources to reach state-of-the-art performance.
Researcher Affiliation Academia Zichu Liu University of Toronto jieben.liu@mail.utoronto.ca Lacra Pavel University of Toronto pavel@ece.utoronto.ca
Pseudocode Yes Algorithm 1: Level k Adam: proposed Adam with recursive reasoning steps
Open Source Code No We will provide the code needed to reproduce our main experimental results.
Open Datasets Yes In our first experiment, we evaluate Lv.k Adam on generating a mixture of 8-Gaussians... We evaluate the effectiveness of our Lv.k Adam algorithm on unconditional generation of CIFAR-10 [31]. ...we evaluate its performance on the STL-10 dataset [11] with 3 48 48 resolutions.
Dataset Splits No The paper states it trains Lv.k Adam with 'batch size 128 for 600 epochs' and mentions parameters like 'latent dimension of 64' for 8-Gaussians, and specific learning rates and β values for CIFAR-10. While it mentions training details, it does not explicitly provide information on dataset splits for training, validation, or testing using percentages or counts within the main paper content. It refers to Appendix 5-8 for training details, but that content is not provided.
Hardware Specification Yes Using a single Nvidia RTX3090 GPU and 30 times fewer parameters than Big GAN on CIFAR-10, we achieve an FID of 10.17 for unconditional image generation within 30 hours... Each run on CIFAR-10 dataset takes 30 33 hours on a Nvidia RTX3090 GPU. Each experiment on STL-10 takes 48 60 hours on a Nvidia RTX3090 GPU.
Software Dependencies No The paper mentions using the 'Adam optimizer' and 'SNGAN architecture'. It also states 'In Appendix 5-7' that libraries used are described, but no specific software dependencies with version numbers (e.g., Python, PyTorch/TensorFlow, or specific library versions) are provided in the main text.
Experiment Setup Yes We use a two layer multi-layer perceptron with Re LU activations, latent dimension of 64 and batch size of 128. ... For Lv.k Adam, we use β1 = 0 and β2 = 0.9 for all experiments. We use different learning rates for the generator (ηθ = 4e 5) and the discriminator (ηϕ = 2e 4). We train Lv.k Adam with batch size 128 for 600 epochs.