Help with GAN with rectagular data

Hi all, i’m working on time series data applications of GANs, the dataset is usually non-square (e.g. Time 100 days x 10 features, 1 channel).

Lots of codes (e.g. in github) are designed for square images. If you modify the code and once you change input/output sizes, kernel sizes, strides and padding; tbh i cannot the sizes right different layers of the of the CNN (in particular for generator net).

any help here? tricks? thanks

Hey :slight_smile:
I’m doing actually a similar thing for my bachelorthesis. My GAN has to work with CNNs and an additional context vector, which I haven’t implemented yet.
I tried to build a GAN with STGCN. That are CNNs for spatio temporal data. They use non square kernelsizes for the convolution.

thanks. this a good reference Taxaria:

1 Like