search

N3PDF / ganpdfs

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

Impose Sum Rules at the Evolution Basis #23

Closed Radonirinaunimi closed 3 years ago

Radonirinaunimi commented 4 years ago

Test if further improvement can be gained:

Radonirinaunimi commented 4 years ago

Hi @scarlehoff! When you have some spare time, would you mind checking the following layers that computes impose the sum rules restrictions? My previous implementation was bugged and tried a new implementation here. But it looks like this breaks the graph and eventually outputs errors and I am not quite sure how to solve it. https://github.com/N3PDF/ganpdfs/blob/a468b28f456d245492eb2834501312f120280da9/src/ganpdfs/custom.py#L43-L102 PS: If you run it, you can just use the appropriate runcard by ganpdfs smr_runcard.yaml -o <output_folder_name> --force.

scarlehoff commented 4 years ago

I have problems when I try to install this with pip install -e . or python setup.py develop, is it normal?

scarlehoff commented 4 years ago

With this changes I can run ok. For me (maybe your error was different) it was crashing due to a dimensional mismatch, I think basically the last permutation was wrong. Please have a look at my commit and change it if you see anything wrong! I was not sure whether the ggnorm was the `A{g}` norm in the paper but I have treated it as such all the ones that are not multiplied get just a "1". I did it for my own benefit since I was a bit confused about what the previous concatenate was doing. Also changed to stack so you can skip at least one of the permutations!

Radonirinaunimi commented 4 years ago

I have problems when I try to install this with pip install -e . or python setup.py develop, is it normal?

What errors do you get? Is it about package conflicts?

Radonirinaunimi commented 4 years ago

With this changes I can run ok. For me (maybe your error was different) it was crashing due to a dimensional mismatch, I think basically the last permutation was wrong. Please have a look at my commit and change it if you see anything wrong! I was not sure whether the ggnorm was the `A{g}` norm in the paper but I have treated it as such all the ones that are not multiplied get just a "1". I did it for my own benefit since I was a bit confused about what the previous concatenate was doing. Also changed to stack so you can skip at least one of the permutations!

Immense thanks! I'll have a look now.

scarlehoff commented 4 years ago

What errors do you get? Is it about package conflicts?

It is a problem at the level of the installation of the script. When is linked it doesn't seem to find tensorflow? I was wondering whether you saw it as well as I didn't look too deeply into it, my environment is a bit of a mess (I tried to install tf 2.4 before so you can imagine :P).

Radonirinaunimi commented 4 years ago

What errors do you get? Is it about package conflicts?

It is a problem at the level of the installation of the script. When is linked it doesn't seem to find tensorflow? I was wondering whether you saw it as well as I didn't look too deeply into it, my environment is a bit of a mess (I tried to install tf 2.4 before so you can imagine :P).

I see :D On my laptop, DOM, and boogiepop, it works fine. I'll re-install the system on my other laptop with a new fresh environment and check whether I experience any issue.

Radonirinaunimi commented 4 years ago

The first test of the evolBasis generation and/or sum rules restrictions was not successful. The gans just outputted noise. I will have to narrow down the problem to see whether the issue persists at the evolBasis itself or when sum rules are imposed. There might also be some conceptual bugs somewhere.