PART 2
I’ve tried generating faces to see how many images and epochs you need to get something recognizable. For part 1, I’ve done 23 epochs with about 20,000 images, and found that the machine was able to create something recognizable by the 4th epoch but the realness level stayed the same.
Since then, I’ve been running the process for 2 weeks now (hope the GPU won’t blow up..) and was able to finish up to epoch 70. I knew that the precision won’t improve anymore after around epoch 30~40 but you know, just in case. Here are the results.
RESULTS
As you can see the images didn’t improve at all. To be honest, I didn’t think that the improvement would stagnate this much. So, to recap what I’ve learned from this experiment are…
A) Simply running a GAN with 20,000 images is clearly not enough
B) The image get’s a little recognizable by around epoch 5
C) The image’s realness level doesn’t improve much after epoch 30. (Not just by the numbers)
There are so many ways to tune the hyperparameters, so I’ll try not to tweak too much at once. One way I’m thinking is improving image resolution. Since I’ve learned about C) above, I’ll keep the epoch to no more than 30. If I still can’t find any improvement, then I will tweak some of the hyperparameters in my code.
