基于多条件对抗和梯度优化的生成对抗网络

该文针对模式崩溃的问题，从多生成器博弈强迫每个生成器生成不同模式数据的思路出发，提出了一种基于多生成器的生成对抗网络(IMGAN)。IMGAN在多个生成器之间采用参数共享的方式来加速训练，同时采用最后一层独立训练的方式来弱化参数同一性所带来的影响；引入一个正则惩罚项使得损失函数可以更好地满足Lipschitz连续，一定程度上避免了梯度消失带来的影响；引入一个超参数来解决多重损失函数带来的差异性问题，避免过度偏向其中某一种梯度方向。最后，通过在多个数据集上的对比实验验证了该文模型的表现和性能。

A Generation Adversarial Network Based on Multi-Condition Confrontation and Gradient Optimization

Aiming at the problem of pattern collapse, this paper starts from the idea of forcing each generator to generate different pattern data in a multi-generator game, and proposes a multi-generator-based generation confrontation network, named improved multi-generator generative adversarial nets (IMGAN). IMGAN uses parameter sharing between multiple generators to speed up training, and at the same time uses the last layer of independent training to weaken the impact of parameter identity; introduces a regular penalty term to make the loss function better satisfy Lipschitz continuousness, which avoids the effect of gradient disappearance to a certain extent; and introduces a hyperparameter to solve the disparity problem caused by multiple loss functions and avoid excessive bias toward one of the gradient directions. At last, we verify the performance of our model through comparative experiments on multiple data sets.