Hey im currently working on a GAN for generating new Data on Heavy Rain Events.
I have a Datasets of 77 Heavy Rain events from Aachen, Germany as 24x19 Matrices.
Basically i have it set up and running.
Here is my Code so far:
while my in input data looks like this for example:
after 18000 iterations
my output Data looks often more like
and after 30000iterations
it looks more like a overall distribution than single heavy rain events
Is it because my generator topology is to “linear” ?
Is it because i have no labeling on my Dataset?
Since im new to Python and Pytorch, i am grateful for every tip and addition to my Code.
Thank You in advance
I think you model is suffering from what I call “garbage in garbage out”.
you want your generator to generate new data from a randomized vector with no statistical relationship to ground truth? no way.
My suggestion is first to come up with a way to describe your ground truth as a vector of z_dim that is statistically relevant. Then feed that to your generator. Your generator will eventually learn enough prior so that when you do feed it a new randomly generated vector it can output plausible new data.
If you don’t want to do that. Then I suggest changing your Generator network to an Encoder-Decoder set up.
Hey Scott_Hoang , Big Thanks for your reply, I will try to implement your first Suggestion.