[BigGAN] Large Scale GAN Training for High Fidelity Natural Image Synthesis

1、目的

        在大型数据集上训练class-conditional GAN,并探索相关的trick

2、贡献

        1)数据集的扩大使得GAN的表现也随之提升。文章的网络参数量是之前工作的2~4倍,batch size是之前的8倍。文章分别从两方面对performance进行提升:scalability - architectural change;conditioning - regularization scheme

        2)发现了truncation trick,通过sampling technique来平衡variaty和fidelity的trade-off

        3)要想实现完全的训练stability,就必须大幅度牺牲performance

3、网络结构

        1)SA-GAN

        2)class information

                G:class-conditional BatchNorm(所有BatchNorm层使用shared embedding,linearly projected to each layer's gains and biases,以节省computation和memory损耗)

                D:projection

        3)2 steps D, 1 step G

        4)evaluation时,对G的weights moving average

        5)增加depth和width可以显著提升实验结果

        6)skip-z将noise向量z添加到网络多层中。BigGAN通过将z分为chunks,然后和conditional vector c并联;BigGAN-deep则直接将z和conditional vector c并联

4、创新性

        1)truncated normal

                z sampling: values fall outside a range are resampled to fall inside that range

                trade-off: threshold ↑,sample variety ↑,quality ↓

        2)Orthogonal Initialization

                truncated normal在一些模型下性能不好(这里我没看懂为啥不好,文章说会导致训练和测试的distribution shift?),可以通过让G smooth来中和该问题

                                        

                最终文章采用了改良版

                                                        

5、局限性                

        1)会出现training collapse。训练不稳定的来源是G和D共同作用的结果。为了使得训练稳定,就需要牺牲最终的performance,因此不如直接用early stopping

[BigGAN] Large Scale GAN Training for High Fidelity Natural Image Synthesis_第1张图片

         -> Generator

                每个weight matrix的top three singular values \sigma _{0}, \sigma _{1}, \sigma _{2}对mode collapse是最informative的

                                        ​​​​​​​

                 spectral normalization可以防止\sigma _{0}或者\frac{\sigma _{0}}{\sigma _{1}}逐步增加和爆发,并且在一些情况下还能略微提升实验结果,但仍然无法完全制止training collapse

        -> Discriminator       

                G会周期性的产生会严重干扰D的batches

                                        [BigGAN] Large Scale GAN Training for High Fidelity Natural Image Synthesis_第2张图片

                 R_{1} 以0为中心的梯度惩罚使得训练更加stable,但是最终的performance严重下降

                D的loss逐渐接近0,但是在collapse时会迅速增加。因为D在训练集上过拟合了

        2)出现了新的failure类型。以往的生成模型的failure主要包括local artifacts、images consisting of texture blobs instead of objects、canonical mode collapse。文章发现了class leakage

                        ​​​​​​​        ​​​​​​​        ​​​​​​​        [BigGAN] Large Scale GAN Training for High Fidelity Natural Image Synthesis_第3张图片

                                 

你可能感兴趣的:(paper阅读笔记,生成对抗网络,人工智能,神经网络)