I am currently trying to reimplement the paper and I stumbled upon a problem when computing the GSVDD classifier loss.
This is what I have:
Latent/encoded vector of shape=(1, Z), the input from the current batch;
Mean vector of shape=(1, Z), calculated in the E step from all samples;
Standard deviation vector of shape=(1, Z), calculated in the E step from all samples.
Now when I calculate the GSVDD classifier loss I end up with a loss that is also (1, Z), should I average it to get a single value? I'm probably missing a key point here.
Hello, amazing work!
I am currently trying to reimplement the paper and I stumbled upon a problem when computing the GSVDD classifier loss.
This is what I have:
Now when I calculate the GSVDD classifier loss I end up with a loss that is also (1, Z), should I average it to get a single value? I'm probably missing a key point here.