Generative Adversarial Network written in TensorFlow

This is a Generative Adversarial Network[1] written in TensorFlow

The model is trained on the dateset SVHN

to run the code, put the file train_32x32.mat (with SVHN) in the dir data/SVHN

then run

python2.7 main.py

the images generated will be put in directory gen/

The following iamges were generated by our model, first one is generated and the second one is mixing generated images with source data (the 400th batch)

Epoch 0

Epoch 65

Epoch 457

Epoch 795

---

Deficency of my apporach

• The result is not stable. For example, in the 547th generation, we get a terriable image as follows

• Sometimes, the two neighboring generations will have nearly no difference, this may due to being the local optimal.

• On the other hand, some output of neighboring generations are very different, that maybe caused by the large stdarand deviation(10.0) of the input noise of the generator.

• The SVHN dataset's distribution has a relatively large deviation, so it is hard for our network to learn it.

• May exists overfitting, no idea what it will cause or how to avoid it

---

Update on Aug 12th, 2017:

Rewrite the model follow WC-GAN[2], which was developed by Alec Radford etc.

All code before were moved to the directory "ignore", please ignore them because they are just something of a mess.

---

Reference

[1] Generative Adversial Nets, Ian J. Goodfellow etc., arxiv 1406.2661

[2] Unsupervised Representation Learning With Deep Convolution Generative Adeversial Networks, Alec Radford etc., ICLR 2016