PyTorch Implementation of Improved Techniques for Traning GANs

This is the extension of my PyTorchDCGAN repository, adopting techniques and suggested metric from Salimans et al. [1]. Escpecially, I'll test my model on the following datasets:

• MNIST [2]
• CelebA [3], cropped
• LSUN [4]
• Danboru2019 [5], SFW, downscaled

and compare the output with naive DCGAN [6], both qualitatively and quantitatively. However, due to the limitation on my access to the hardware, some of the planned experiments may not be performed.

Disclaimer

As I'm an active learning student, this implementation may not be complete or accurate. Therefore, I recommend you to use other reliable implementations if you're willing to use it in your project.

To-Do

Though I marked minibatch discrimination as implemented, it needs more optimizations. For example, in matrix subtraction, it requires a lot of memories.

• Feature Matching
• Minibatch Discrimination
• Historical Averaging
• One-sided Label Smoothing
• Virtual Batch Normalization

Model Outputs

MNIST

Single epoch with Feature Matching

10 epochs with Feature Matching

Unfortunately, as my vanilla DCGAN implementation did, the generator still collapsed on a small dataset of MNIST [2].

Single epoch with Minibatch Discrimination

Compared to my vanilla DCGAN implementation, images look more connected/filled, though it looks less like the actual numbers.

10 epochs with Minibatch Discrimination

Much better! For the first time, the model doesn't collapse on long triaining!

References

• [1] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen, "Improved Techniques for Training GANs," in NIPS 2016, D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, 2016, pp. 2234-2242.
• [2] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, "Gradient-based learning applied to document recognition." in Proc. IEEE, Nov. 1998, pp. 2278-2324.
• [3] Z. Liu, P. Luo, X. Wang, and X. Tang, "Deep Learning Face Attributes in the Wild," in Proc. ICCV, Dec. 2015.
• [4] F. Yu, Y. Zhang, S. Song, A. Seff, and J. Xiao, "LSUN: Construction of a Large-scale Image Dataset using Deep Learning with Humans in the Loop," 2015, arXiv:1506.03365.
• [5] Anonymous, and Danbooru Community, and G. Branwen, Jan. 2020, "Danbooru2019: A Large-Scale Crowdsourced and Tagged Anime Illustration Dataset," Web. [Online]. Available: https://www.gwern.net/Danbooru2019.
• [6] A. Radfold, L. Metz, and S. Chintala, "Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks," 2015, arXiv:1511.06434.