内容简介:This is the main component of the GAN. Here a generative network makes an attempt to learn the trends of the distribution and generate data belonging to this distribution.So how do we do this? Well it depends. In case of our example, for generating dog ima
This is the main component of the GAN. Here a generative network makes an attempt to learn the trends of the distribution and generate data belonging to this distribution.
So how do we do this? Well it depends. In case of our example, for generating dog images, our noise will undergo some De-Convolution Layers and finally, the output will be an image, hoping to be a dog’s. The generated image will be classified as real or fake by the discriminator. The generator will be penalized with more loss if the generated image is far from ‘being real’ and the loss will decrease as it makes progress in fooling the discriminator.
This is the basic idea behind a generator network. The architecture may differ with the use case. There are GAN architectures that are conditional i.e. there is some notion of a condition for generating data alongside the noise. Such additions may alter the architecture a bit, but the fundamental idea of the adversaries stand.
We now move on to the collective training of the generator and the discriminator.
Combined Training of the Generator and Discriminator as ‘The GAN’
So, GANs are great! But combined training of the generator and discriminator poses a problem for convergence. Let’s see how the GAN is trained.
The generator should be isolated from the discriminator and vice versa while training. In short, first the discriminator is trained for several epochs then the generator, and this sequence continues back and forth. This ensures that the training is in the right direction. Otherwise, training both at the same time would be like hitting a moving target and would possess higher chances of failure.
If you can’t train a classifier to tell the difference between real and generated data even for the initial random generator output, you can’t get the GAN training started.
Moreover, as the generator gets better at fooling the discriminator, as a corollary, the discriminator’s accuracy decreases. There comes a point where the discriminator reaches accuracy 0.5 which suggests that the discriminator has started predicting randomly i.e. on a coin toss (which it is obviously not supposed to do). This does not help the generator training in any way and may deteriorate its performance instead.
Hence, convergence in GANs is not stable and is a major issue.
Training Loss
The GANs paper defines the training loss as:
This is the Minimax loss, the discriminator tries to maximize it and the generator tries to minimize log(1 – D(G(z))) as it cannot touch log(D(x)).
However, the paper also suggests modifying this loss for the generator to maximize log(D(G(z))).
A Few Use Cases (to get you thinking)
- Vanilla GANs (the ones described in the GAN paper ) can be used to augment data for training in case of imbalanced or less data.
- Deep Convolutional Generative Adversarial Networks or DCGAN are vanilla GANs with Convolutional Layers for image generation
- The pix2pix model can be used to take a wireframe of a structure as an input and generate the complete structure as output. Other applications include coloring black-white images.
- CycleGAN is used to transform images from one domain to another without any paired training samples
- DeepFake Generation and Detection is one of the latest research topics in GANs which essentially generates edited or tampered images that look realistic to the naked eye. This can be easily done manually with the help of photo/video editing softwares and hence, is vulnerable to misuse. DeepFake detection can be derived as a corollary to generation.
I will try to cover the implementation of some of these GANs in future articles. So stay tuned!
Conclusion
The main takeaway from this article is the concept of adversarial training evaluation of a generative model. This gives rise to many branches for researchers to come up with similar evaluators for better model performance.
We have also seen the idea of generative and discriminative modeling, and how they differ from each other.
We studied how GANs work and the idea behind it. We’ve had a pragmatic approach overall; we addressed the training aspect of the GANs and the issues that arise while actually training the network.
Finally, we had a word on some well-known flavors of GANs.
References:
Original paper: https://arxiv.org/abs/1406.2661
For more on the mathematical transformational aspect of GANs, refer:
以上就是本文的全部内容,希望对大家的学习有所帮助,也希望大家多多支持 码农网
猜你喜欢:本站部分资源来源于网络,本站转载出于传递更多信息之目的,版权归原作者或者来源机构所有,如转载稿涉及版权问题,请联系我们。
计算机组成:结构化方法
坦嫩鲍姆 / 刘卫东 / 人民邮电出版社 / 2006-1 / 65.00元
本书采用结构化方法来介绍计算机系统,书的内容完全建立在“计算机是由层次结构组成的,每层完成规定的功能”这一概念之上。作者对本版进行了彻底的更新,以反映当今最重要的计算机技术以及计算机组成和体系结构方面的最新进展。书中详细讨论了数字逻辑层、微体系结构层、指令系统层、操作系统层和汇编语言层,并涵盖了并行体系结构的内容,而且每一章结尾都配有丰富的习题。 本书适合作为计算机专业本科生计算机组......一起来看看 《计算机组成:结构化方法》 这本书的介绍吧!