questions regarding the training results

[Lucas9909]

I am trying to train the CFGen by the following code:
python /CFGen/cfgen/train_sccfm.py dataset=pbmc10k_multimodal +dataset.encoder_type=learnt_autoencoder
dataset.one_hot_encode_features=False denoising_module=resnet_big
denoising_module.is_binarized=True encoder=encoder_multimodal_joint
encoder.is_binarized=True trainer.max_epochs=1000 launcher=slurm_icb
logger.project=fm_resnet_autoencoder_pbmc10k_multimodal_joint
training_config.encoder_ckpt=/CFGen/project_folder/experiments/train_autoencoder_pbmc10k_multimodal_joint/f68e4ca2-ee84-4788-b906-9074847971cc/checkpoints/epoch_39.ckpt

but the output showed:

do you know why the MMD is so large and how to solve this problem? Thanks!

[allepalma]

Hi @Lucas9909, thanks for reaching out. Could you do a qualitative check of the results with the notebook provided? Maybe with the UMAP overlap between generated and real, to see if you are doing everything correctly. Also, I suggest running the metrics from here:

https://github.com/theislab/CFGen/tree/main/notebooks/evaluation_metric_example

and see if you get strange results, let me know :)