search

CyberAgentAILab / layout-dm

LayoutDM: Discrete Diffusion Model for Controllable Layout Generation [Inoue+, CVPR2023]
https://cyberagentailab.github.io/layout-dm
Apache License 2.0
225 stars 24 forks source link

about the number of parameter adjustment. #11

Closed jdily closed 1 year ago

jdily commented 1 year ago

Hi, thx for sharing the nice work. Recently, I want to reproduce the LayoutGAN++ experiment in Table 1 (using the source code provided by LayoutGAN++). However, I think because the number of parameters are not the same (I think the original LayoutGAN++ parameters are way more than the simplified model used in this paper). I am wondering if you can share the simplified version of LayoutGAN++ as well? Or you can share the simplified architecture so that I can perform experiment on it?

thank you!!

naoto0804 commented 1 year ago

Hi, I'll get you back when I'm back from CVPR

naoto0804 commented 1 year ago

As you guessed, we did experiments for LayoutGAN++ with its codebase. This is because of practical implementation reasons; GANs are complex compared to simple forward-backward workflow and I found it too hard to integrate GANs with this framework.

Below are the instructions, sorry for the inconvenience. (i) We use num_layers=4, d_model=512, nhead=8 for the generator (ii) copy this repo's code for data preprocessing and loading pipeline. (iii) add some codes to dump predictions (if you want to evaluate the result with eval.py)

jdily commented 1 year ago

Hi, thank your for your quick reply. I will try it follow your instructions, and I will ask questions if I still have some problems reproducing that results. :)