search

cossio / RestrictedBoltzmannMachines.jl

Train and sample Restricted Boltzmann machines in Julia
MIT License
15 stars 3 forks source link

where is RestrictedBoltzmannMachines.fpcd! — Method #23

Closed bhomass closed 1 year ago

bhomass commented 1 year ago

in the document page https://docs.juliahub.com/RestrictedBoltzmannMachines/ygJtt/0.21.1/autodocs/

There is RestrictedBoltzmannMachines.fpcd! — Method

listed. but nowhere to be found in the code. Does it exist? maybe in an older version? Isn't it needed to reliably sample synthesized fantasy particles from random input?

cossio commented 1 year ago

It's from an older version. I moved that method to https://github.com/cossio/FastPCD.jl.

Isn't it needed to reliably sample synthesized fantasy particles from random input?

Not sure what you mean.

bhomass commented 1 year ago

what I mean is from the original paper, it is mentioned that using CD-k, you get poor results in sampling using random inputs. So they developed FPCD to improve the quality of the fantasy particles. However, from what I see in your code, you seem to get good results without leveraging FPCD and yet using random inputs.

btw, I went through you binaryRBM example on MNIST data. Have you also obtained good results for MNIST using gaussian-binary layers?