Gans!

Exploring how Geometric Transforms Map onto the Z-Space Vector

My code for a small GANs research project I did at HT Kung's lab the summer of 2017. I initially implemented a DCGAN in Torch, but Chainer was already configured on the GPU we had, so I ported it over. The purpose of the research project was to see how a geometric transformation in a picture maps onto the noise vector after training.

1 dimensional Z-space of a DCGAN trained on randomely rotated lines:

---

1 dimensional Z-space mappings during training on same dataset:

---

1 dimensional Z-space mappings during training on dataset of resized lines:

---

Final 1 dimensional Z-space mappings of DCGAN trained on dataset of resized and rotated lines:

---

Final 2 dimensional Z-space mappings of DCGAN trained on dataset of resized and rotated lines: