Pytorch Implementation of Generative Adversarial Nets (GAN)

This is my take on the implementation of GAN [1], using Pytorch [2]. I used MNIST [3] for training and testing.

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.

If you find any bugs or flaws, please let me know. Also, I tried using the IEEE style citation, just for practicing. If you think any of the citations are improper, please let me know.

Model Outputs

Here, I'll present 10 images for each model trained with the different numbers of epochs. I trained them on Google Colab. However, due to some manual editings, codes on Colab may differ from codes on Github. Codes on Github always takes precedence.

10-epochs

100-epochs

Pretrained Models

You can download them from Models folder.

Conclusion

During training, I found that Dropout [4] is essential for preventing models from collapsing. However, I'm not sure about the effects of other modifications, such as maxout activation [5], which may be due to my lack of knowledge.

References

• [1] I. J. Goodfellow et al., "Generative Adversarial Nets," 2014, arXiv:1406.2661v1.
• [2] A. Paszke et al., "PyTorch: An Imperative Style, High-Performance Deep Learning Library," in Advances in Neural Information Processing Systems 32, 2019, pp. 8024-8035.
• [3] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, "Gradient-based learning applied to document recognition," Proceedings of the IEEE, vol. 86, no. 11, pp. 2278-2324, Nov. 1998.
• [4] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, "Dropout: A Simple Way to Prevent Neural Networks from Overfitting," Journal of Machine Learning Research, vol. 15, no. 1, 2014, pp. 1929-1958.
• [5] I. J. Goodfellow, M. Mirza, A. Courville, and Y. Bengio, "Maxout Networks," in International Conference on Machine Learning, 2013, pp. 1319-1327.