Closed shunk031 closed 3 months ago
It's strange. I might forget to add the config to this new project. But anyway, as you know, it's simple but a parameter of num_bin
should be set to 32
. We found that when num_bin
is set to 128
, the training was unstable, as mentioned in the official issue in LayoutDM repository, https://github.com/CyberAgentAILab/layout-dm/issues/29.
DEBUG=False
ADDITIONAL_ARGS=" \
data.transforms=[image,shuffle] \
++tokenizer.num_bin=32 \
"
EXP_ID="LayoutDM_Shuffle_NumBin32"
If you sent me a pull request if that config works, I would greatly appreciate it
Thank you for your response.
I will experiment with the configuration you provided. If I achieve good results, I will create a PR for it.
Is there a place we can get the clustering weights for cgl? Got the following error with the config above.
FileNotFoundError: [Errno 2] No such file or directory: '~/RALF/cache/PRECOMPUTED_WEIGHT_DIR/clustering/cache/cgl_kmeans_train_clusters.pkl'
I can find the clustering weights for publaynet and rico from layoutDM repo, but not for cgl
I guess I probably need to generate it myself using clustering_coordinates.py
from layoutDM repo
@wandonye
I'm sorry for not getting back to you sooner. I have already uploaded the file. Feel free to use it. Google drive: https://drive.google.com/file/d/1XkYubu123dhwxB3wsEnN-iR-jOc_hGxD/view?usp=sharing One drive: https://1drv.ms/u/s!AjGupxRG928Al1PKN2XGJoB98GIZ?e=RHOhlU
Hi, thanks for sharing this amazing repository. I strongly believe this research will greatly advance the field of layout generation.
I have a question though, I couldn't find the config for LayoutDM in the
configs
directory. Could you please provide more details about this? Thanks in advance.