一种多条件生成对抗网络模型

A Multi-Conditional Generative Adversarial Network Model

传统生成对抗网络在条件数量增加时,具体条件组合下的训练样本数量相应减少,此时判别器还是只产生一个判断真假样本的结果,条件和样本对于判别器梯度的贡献降低,导致最终符合多个条件的生成样本的多样性下降。因此提出多条件生成对抗网络模型,增加了预处理模块对多个条件进行处理,提高了输入模型中数据条件的维度,并修改了判别器模块的结构,让判别器的权值根据每个条件进行更新,使得梯度值得到充分利用,提高了判别器的性能,因此生成器能得到效果好的生成样本,最后通过实验验证了新模型的有效性。

As the number of conditions in GANs increases,the number of training samples for specific combinations of these conditions decreases.Consequently,the discriminator only produces a single output to classify samples as real or fake.This reduces the gradient contribution to the discriminator,leading to a decrease in the diversity of generated samples that meet multiple conditions.This paper proposes a multi-conditional generative adversarial network model,which incorporates a preprocessing module to handle multiple conditions.This enhancement improves the dimensionality of the conditions in the input data and optimizes the utilization of gradient values.Additionally,the structure of the discriminator module is modified to update the discriminator weights based on each condition.As a result,the generator can produce samples with better performance,supported by the enhanced discriminator.Experimental results demonstrate the effectiveness of the proposed model.