Open carlito387 opened 3 years ago
Ideally dont destroy the old code both implementation would be nice
I will be looking into good metrics to compare our network trained on different f-divergences. Frédéric shared this paper on evaluation methods for GANs — https://arxiv.org/pdf/1802.03446.pdf — I will start digging into this.
@simondemeule if possible could you implement the all plots on one divergence firsts it would really help for comparison of the rates of divergence. Also if it's not too much troubles would it be possible to add to the graph code something that would plot real and fake stats (https://github.com/simondemeule/IFT6269-project/issues/17) in a graph? Thanks!
Since our critic outputs V_w(x) regardless of the f-divergence chosen, we can easily calculate any divergence from V_w(x) gen and real. Hence we could modify the code to save the V_w(x) at each time step and then we could compare the convergence based on on divergences instead of on different divergences (which I dont think means much).
Assigning it to myself