Revers Unrolled

GAN

Abstract

2. RELATED WORK

3. METHOD

Session

1. INTRODUCTION

Abstract

在生成對抗網路（GAN）中，模式坍塌是 GAN 常見的訓練阻礙之一，其導致 GAN 生成的內容大量集中在真實分布中的特定一種模式。本研究運用訓練過程中 GAN 會在多個模式中變換的特性改善訓練流程，將當前 Step 的生成網路不僅須受當前判別網路之檢驗還須受 Step-k 的判別網路檢驗，提升規避判別網路盲點模式的難度，強制其擬和真實分布。相較於舊有方法將當前 Step 的判別網路展開，我們的方法重新利用已有的權重，降低了運算資源的損耗。

Abstract

In Generative Adversarial Networks (GANs), mode collapse is one of the common training obstacles, which results in the generation of a large number of contents focusing on a specific pattern in the real distribution. In this study, we improve the training process by utilizing the characteristic of GANs switching between multiple modes during training. The current step's generator network is not only examined by the current discriminator network, but also by the discriminator network of step-k, which increases the difficulty of avoiding the blind spot mode of the discriminator network and forces it to mimic the real distribution. Compared to the previous method of expanding the current step's discriminator network, our method reuses the existing weights, reducing the consumption of computational resources.