Issue with vEMdiffuse-a not functioning correctly

[Shogosuga]

Hello. I'm using vEMdiffuse-a, but I'm getting poor inference results. Could you please advise on what I should do?

I followed the steps in training.ipynb to train the model, and then used prediction.ipynb to obtain inference results. During inference, I modified resume='./experiments/vEMDiffuse-a/best' to resume='./experiments/train_vEMDiffuse-a_240303_001236/checkpoint/500' to use my trained model as described in your notebook. However, I couldn't achieve the results shown in your original notebook.

Could you please advise if there's something wrong with my code?

[Luchixiang]

The demo dataset in training.ipynb only consists of 24 layers from MICrONS multi-area volume, which is too small to train a diffusion model. To improve the performance, you should download more data from the MICrONS multi-area volume and train the diffusion model for a longer time (about 1000 epochs).

[Shogosuga]

Thank you for your quick response!

I want to set the number of epochs, so could you tell me which variable in the vEMDiffuse-a.json sets the number of epochs? ["train"]["n_epoch"] or ["train"]["n_iter"]?

[Luchixiang]

Both works for setting the training duration. n_iter controls the number of iteration and n_epoch controls the number of epochs

[Shogosuga]

I see. Your paper states, 'we terminated the training after 1200 epochs'. Did you achieve this by simply setting the ['train']['n_epoch'] variable in the config to 1200?

[Luchixiang]

yes