Download to read offline
Uploaded byTechExeter
Generative Adversarial Networks by Tariq Rashid
The document discusses generative adversarial networks (GANs) and how they can be used to generate new images based on training data. It provides examples of GANs trained on the MNIST dataset to generate handwritten digits, and on a dataset of celebrity faces. The examples demonstrate how the generated images improve in quality and level of detail as the GANs are trained for more epochs. Key points are made that GANs can generate plausible new data rather than just copying training examples, and that mode collapse is a major open research question when using GANs.
More Related Content
Similar to Generative Adversarial Networks by Tariq Rashid
Generative Adversarial Networks by Tariq Rashid
- 1. Generative Adversarial Networks withPyTorch Tariq Rashid Tech Exeter, Dec 2019 https://tinyurl.com/wj9dzk7
- 2.
- 3. Art Portrait of EdmondBelamy, 2018 https://www.christies.com/features/A-collaboration-between-two-artists-one-human-one-a-machine-9332-1.aspx
- 4.
- 5.
- 6.
- 7. Machine Learning andNeural Networks
- 8. Machine Learning andNeural Networks
- 9. Machine Learning andNeural Networks
- 10. Machine Learning andNeural Networks http://makeyourownneuralnetwork.blogspot.com
- 11.
- 12.
- 13.
- 14.
- 15.
- 16. GANs - 2-dimensionalMNIST 2 epochs 6 epochs
- 17. GANs - 2-dimensionalMNIST https://arxiv.org/pdf/1708.02556.pdf
- 18. GANs - 2-dimensionalMNIST Improving GAN Performance ... or even just making them work!
- 19. GANs - 2-dimensionalMNIST leaky RELU strong gradient avoids vanishing gradient
- 20. GANs - 2-dimensionalMNIST layer / batch normalisation https://towardsdatascience.com/intuit-and-implement-batch-normalization-c05480333c5b
- 21. GANs - 2-dimensionalMNIST 1 epoch
- 22. GANs - 2-dimensionalMNIST 2 epochs
- 23. GANs - 2-dimensionalMNIST 4 epochs
- 24. GANs - 2-dimensionalMNIST 8 epochs
- 25. GANs - 2-dimensionalMNIST 2 epochs on 60,000 training set remember the generator has not seen the data !
- 26. GANs - full-colourphotos of faces CelebA dataset 202,599 faces (i’m only using 20,000) http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html HDF5 format is useful for large data (and slow io like colab)
- 27. GANs - full-colourphotos of faces 218 x 178 x 3 218 x 178 x 3 100 1 218 x 178 x 3 218 x 178 x 3 300 100 DiscriminatorGenerator
- 28. GANs - full-colourphotos of faces Discriminator Generator simple architecture
- 29. GANs - full-colourphotos of faces will such a simple architecture work?
- 30. GANs - full-colourphotos of faces experiment yourself !
- 31. GANs - full-colourphotos of faces
- 32. Learning Localised ImageFeatures https://devblogs.nvidia.com/deep-learning-nutshell-core-concepts/
- 33. GANs - full-colourphotos of faces epochs = 1
- 34. GANs - full-colourphotos of faces epochs = 2
- 35. GANs - full-colourphotos of faces epochs = 4
- 36. GANs - full-colourphotos of faces epochs = 6
- 37. GANs - full-colourphotos of faces epochs = 8
- 38. GANs - full-colourphotos of faces epochs = 10
- 39. GANs - full-colourphotos of faces epochs = 12 features are composed better
- 40. GANs - full-colourphotos of faces Tutorial Code http://makeyourownalgorithmicart.blogspot.com
- 41. Comparison with PublishedWorks https://github.com/NVlabs/stylegan
- 42.
- 43. Key Points varying inputvector Adversarial training is a cool idea. GANs can make plausible data that isn’t a copy-paste or average of the training data. GANs are new - we’re still learning about how they work and don’t work. Mode collapse is the key open research question!