search

EperLuo / scDiffusion

A model developed for the generation of scRNA-seq data
MIT License
27 stars 4 forks source link

The hyperparameters to reproduce the results #1

Open EddieBio opened 6 months ago

EddieBio commented 6 months ago

Hi,

Could you please provide the hyperparameters for the reproduction of the results? I have run your train.sh, however, I cannot reproduce the unconditional generation results. Beside, different scripts contain different model paths although they all aim for unconditional generation. For example, the SCC is 0.76 and the PCC is 0.71, the UMAP figure does not suggest that the real dataset can be covered by the generated dataset, and the ACC of RF is 1.0.

image

image

Thanks.

EperLuo commented 6 months ago

Hi, sorry for the late reply. The hyperparameters used in this code repository are the same as we used when conducting our experiments. The problem might lie in insufficient training. We trained this version of Autoencoder and Diffusion model for 80e4 steps. You could try to train the model for a longer time and see if the results are different. Besides, we've made some major changes to the original model, which reduced the time for training the Autoencoder and diffusion model significantly. You could try this new training pipeline if you are interested.

Sbs12 commented 2 months ago

Hello, I have encountered the same problem. Can your unconditional generation effect achieve the effect described in the article?