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

-> 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

$R_{1}$ 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相关推荐

Paper之BigGAN：ICLR 2019最新论文《LARGE SCALE GAN TRAINING FOR HIGH FIDELITY NATURAL IMAGE SYNTHESIS》(未完待续)
Paper之BigGAN:ICLR 2019最新论文<LARGE SCALE GAN TRAINING FOR HIGH FIDELITY NATURAL IMAGE SYNTHESIS> ...
Paper之BigGAN：《Large Scale Gan Training For High Fidelity Natural Image Synthesis》翻译与解读
Paper之BigGAN:<Large Scale Gan Training For High Fidelity Natural Image Synthesis>翻译与解读目录效果 1 ...
【PaddlePaddle论文复现】LARGE SCALE GAN TRAINING FOR HIGH FIDELITY NATURAL IMAGE SYNTHESIS阅读心得
[PaddlePaddle论文复现]LARGE SCALE GAN TRAINING FOR HIGH FIDELITY NATURAL IMAGE SYNTHESIS阅读心得作者:Andrew B ...
Large scale GAN training for high fidelity natural image synthesis解读
<Large scale GANtraining for high fidelity natural image synthesis>这篇文章对训练大规模生成对抗网络进行了实验和理论分析, ...
《Large scale GAN training for high fidelity natural image synthesis》论文阅读记录
论文连接地址:https://arxiv.org/pdf/1809.11096.pdf Github地址:https://github.com/AaronLeong/BigGAN-pytorch 目录 ...
用于高保真自然图像合成的大规模GAN训练（Large Scale GAN Training For High Fidelity Natural Images）论文 pdf
下载地址:https://u20150046.ctfile.com/fs/20150046-376632643 By Andrew Brock,Jeff Donahue,Karen Simonyan( ...
【食品图像识别】Large Scale Visual Food Recognition
1 引言视觉智能部与中科院计算所于2020-2021年度展开了<细粒度菜品图像识别和检索>科研课题合作,本文系双方联合在IEEE T-PAMI2023发布论文<Large Scal ...
Machine Learning week 10 quiz: Large Scale Machine Learning
Large Scale Machine Learning 5 试题 1. Suppose you are training a logistic regression classifier using ...
【论文阅读 NeurIPS 2022】A Large Scale Search Dataset for Unbiased Learning to Rank
文章目录前言 Abs Intro 2.Preliminary 2.1.Ubiased Learning to Rank 2.2.Existion ULTR Datasets 3.Dataset De ...

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