👏 Credit is made by this Repository toy_gradlogp. The author's blog is worth reading too, for example this article: [Tuto] 夢魘の CUDA: 使用 Preconditioned Conjugate Gradient 輕鬆解決大型稀疏線性方程組
What I did is just modifying the data pipeline from tensorflow to pytorch's dataloader and re-factored the folder/config structure with hydra config tool.
If your are interested in the theory behind Score Matching, maybe read the note I made: Estimation of Non-Normalized Statistical Models by Score Matching
Configuration is following hydra usage.
For the simplest case, you can modify conf/train_config.yaml then execute:
python train.pySame as toy_gradlogp