CoopInit: Initializing Generative Adversarial Networks via Cooperative Learning

Authors: Yang Zhao, Jianwen Xie, Ping Li

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of the proposed approach on image generation and one-sided unpaired image-to-image translation tasks through extensive experiments. In this section, we extensively evaluate the effectiveness of our proposed initialization strategy, Coop Init, for GANs. We begin by testing our method on image generation and unpaired image-to-image translation, comparing our framework to state-of-the-art models. Then we perform some analysis on our model.
Researcher Affiliation Industry Yang Zhao, Jianwen Xie, Ping Li Cognitive Computing Lab Baidu Research 10900 NE 8th St. Bellevue, WA 98004, USA {yangzhao.eric, jianwen.kenny, pingli98}@gmail.com
Pseudocode Yes Algorithm 1: Cooperative Learning; Algorithm 2: Training a GAN with Coop Init
Open Source Code No The paper does not provide an explicit statement or a link to open-source code for the described methodology.
Open Datasets Yes We evaluate the performance of image generation on four widely used datasets listed below: (i) CIFAR-10 (Krizhevsky 2009): This dataset consists of 60K 32x32 images in 10 evenly distributed classes, including 50K training images and 10K testing images. (ii) Image Net (Russakovsky et al. 2015): To balance the computational budget, we use a down-sampled version of Image Net that consists of 32x32 images. Image Net contains over 10 million natural images of 1,000 classes. (iii) FFHQ (Karras, Laine, and Aila 2019): This dataset consists of 70K high-quality and diverse human facial images. We choose to use a down-sampled version of the data with a resolution of 256x256.
Dataset Splits Yes CIFAR-10 (Krizhevsky 2009): This dataset consists of 60K 32x32 images in 10 evenly distributed classes, including 50K training images and 10K testing images.
Hardware Specification Yes All experiments were conducted on 4 Nvidia Titan Xp (12GB) GPUs and Google Colab.
Software Dependencies No The paper mentions optimizers like 'Adam optimizer' and 'R1 regularization' hyperparameters, but does not specify software dependencies like Python, PyTorch, TensorFlow, or other libraries with version numbers.
Experiment Setup Yes Following Zhao et al. (2020b), we halve the number of channels of feature maps at higher resolution layers (i.e., 16x16 and above) to enable faster computation. We further apply the non-saturating loss, set the learning rate to 0.0025, and use the original connection unless specified otherwise, following the approach of Karras et al. (2020a). We evaluate the impact of hyperparameters, including learning rate lr and R1 regularization strength γ, on the proposed learning algorithm. We find that γ = 0.01 works best...