pdfcoffee

soumyasankar99
from soumyasankar99 More from this publisher
09.05.2023 Views

Chapter 6The preceding images were the initial attempts by the GAN. As it learned throughthe following 5000 epochs, the quality of digits generated improved manyfold:The complete code is available in DCGAN.ipynb in the GitHub repo. We can takethe concepts discussed here and apply them to images in other domains. One of theinteresting work on images was reported in the paper, Unsupervised RepresentationLearning with Deep Convolutional Generative Adversarial Networks, Alec Radford, LukeMetz, Soumith Chintala, 2015. Quoting the abstract:"In recent years, supervised learning with convolutional networks (CNNs) hasseen huge adoption in computer vision applications. Comparatively, unsupervisedlearning with CNNs has received less attention. In this work we hope to help bridgethe gap between the success of CNNs for supervised learning and unsupervisedlearning. We introduce a class of CNNs called deep convolutional generativeadversarial networks (DCGANs), that have certain architectural constraints,and demonstrate that they are a strong candidate for unsupervised learning.Training on various image datasets, we show convincing evidence that our deepconvolutional adversarial pair learns a hierarchy of representations from objectparts to scenes in both the generator and discriminator. Additionally, we use thelearned features for novel tasks - demonstrating their applicability as general imagerepresentations."(Radford et al., 2015).[ 207 ]

Generative Adversarial NetworksFollowing are some of the interesting results of applying DCGANs to a celebrityimage dataset:[ 208 ]

Generative Adversarial Networks

Following are some of the interesting results of applying DCGANs to a celebrity

image dataset:

[ 208 ]

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!