How Much Is Enough?
The last time I used gan(Generating cityscapes), the output was not recognizable due to a lack of data. So, as for my next step I wanted to find out how much data and epochs you need to get something recognizable as the image you are trying to generate. Instead of searching for more data of images of cityscapes, I found a large dataset on Kaggle with images of people so I ran the same code on those.(“https://drive.google.com/uc?id=1O7m1010EJjLE5QxLZiM9Fpjs7Oj6e684”)
Experimenting
The results are the images below. (202599 Images and 23 Epochs)
First of all, I was surprised by the fact that even though I ran the process for a WHOLE WEEK, only 23 epochs were finished.
As you can see, a “person-ish” figure appears by the 4th epoch, but after that, the progress from “person-ish” to “oh-this-is-a-fake-person?” stagnates.
For My Next Step
My 2 goals are, A) How much data is enough? and B) How many epochs do you need?
I haven’t stopped my process yet, so for now, I’ll keep it going for maybe another week to see the optimal amount of epochs. Then maybe I can, decrease the number of images and try it again to check out my goal A).
