This repo contains code for computing the rate-distortion curve using the Blahut-Arimoto algorithm.
On top of this, using the run_RD or run_RD_CUDA functions will generate the potential critical points along the RD curve as mentioned in this paper.
You may save the results with the save function and read out the file contents with read_RDresult functions.
First clone this repository (with something like git clone https://github.com/TenzinCHW/RDCurves.jl).
Then, after starting the Julia REPL, enter the following:
julia> ]
(@v1.11) pkg> add <path-to-repo>
Or add the following to a script installrd.jl:
import Pkg
Pkg.add("<patch-to-repo>")
and run it using:
julia installrd.jl
To use the implementation of the BA algorithm, say you had a probability distribution p and a distortion matrix d. beta is the Lagrange multiplier and num is the number of iterations to run. For a typical distribution you would run from 1000 to 10000 iterations. Then you would do something like the following:
import RDCurves
p = rand(10)
d = rand(10, 10)
beta = 0.01
num_iter = 1000
Q, R, D = RDCurves.BA(p, d, beta, num_iter)
Q here is the optimal conditional probability distribution.
R here is the achieved rate in bits.
D here is the achieved distortion.
This code was written... a while back. There were some decisions that make the results a little unwieldy due to limitations of typed serialization back in the day. I will try, where I have time to update it to make it nicer to use (kindly post an issue if you think this code will be useful to you and I will gladly refactor).
Additionally the get_possible_peaks function is written in a state machine-like way and is therefore kind of a mess. I wish I had the mind to make it amenable to accepting optimizers from libraries from the likes of Optim.jl or Optimisers.jl but alas, I was a PhD student just trying to graduate. I apologise in advance for the inconvenience. Again, if this looks like it could be useful, kindly post an issue and I will try to help where I can.
I have a super basic test just to make sure the BA function runs. To use it with your GPU, simply put the p vector and d matrix onto the GPU and use the function as per usual. Same with run_RD (hence why I did not include CuBA.jl; it's just there for historical purposes).
As mentioned above, I will put more effort into refactoring if there is actually a demand for it. Thank you for the understanding.