search

aicenter / GenerativeModels.jl

Generative Models with trainable conditional distributions in Julia!
MIT License
31 stars 3 forks source link

Clean up mmd #67

Closed vitskvara closed 4 years ago

vitskvara commented 4 years ago

It was fixed and is now used from IPMeasures.jl, so there is some unnecessary stuff.

vitskvara commented 4 years ago

I deleted the old stuff and reimplemented mmd via IPMeasures for VAEs. I implemented 2 versions of mmd - mmd_mean and mmd_rand, depending on whether only the mean of encoder is used or if it is actually sampled from it, but I am not sure what difference it makes in training.

vitskvara commented 4 years ago

Nevertheless, training via mmd is still not working due to https://github.com/JuliaLang/julia/issues/31427

nmheim commented 4 years ago

mmd training fixed with 1d48d8e0065c59e9c6cc7a3974dec66598a5413e . if we are happy with the mmd interface I will close this. @smidl ?

vitskvara commented 4 years ago

It's fine with me now.