search

N3PDF / ganpdfs

Generative Adversarial Networks (GANs) for MC PDF replicas.
https://n3pdf.github.io/ganpdfs/
0 stars 0 forks source link

Roadmap compression #13

Closed scarrazza closed 4 years ago

scarrazza commented 4 years ago

So, I believe we should to the following:

  1. implement the classical compressor using CMA and compare the results
  2. Implement the GAN model and check its quality: train slice by slice in x and compare the generated sample to the original after binning
  3. Integrate the GAN generator with the compressor.
scarlehoff commented 4 years ago

Should we focus in 2 first and 1 second? I can help with 1 once the kfolding is done and I'm "NNPDF-free"

scarrazza commented 4 years ago

Yes, 2 is the most important here.

Radonirinaunimi commented 4 years ago

Thanks for the roadmap! I'm currently writing a note explaining what I did so far and will directly start working on these.

Radonirinaunimi commented 4 years ago

@scarrazza, @scarlehoff I added a short note that summarizes the whole concept and everything that was implemented. Could you please check if everything makes sense and/or propose ways to improve the model?

scarlehoff commented 4 years ago

I think it's ok. I like very much the jupyter notebook. One general comment though is that when uploading jupyter notebook as demonstration (as in this case) it is better if you upload it to git already "having run" because that way one can just go to https://github.com/N3PDF/ganpdfs/blob/concept/tools/proof_concept.ipynb and see the result (without having to checkout the branch, I'm that lazy :P ).

Radonirinaunimi commented 4 years ago

@scarlehoff Thanks for the feedback.

Indeed, I should have left the outputs in the notebook. I did not realize that I cleaned everything before pushing. I will make sure to have all outputs in next implementations.

Concerning the whole concept of the training, do you have any comments? Does everything somehow make sense?

@scarrazza

implement the classical compressor using CMA and compare the results

Could you please elaborate more on what this steps technically consist?

scarrazza commented 4 years ago

Could you please elaborate more on what this steps technically consist?

is the code you have for the prior estimators plus a function which returns the replica indexes from the prior using the GA/CMA.