Clarification on input normalization parameters when reloading trained scPred model

[bigbunny222]

Hi, thank you for the great scPred model implementation!

I have a question regarding the input normalization step during model training and prediction.
In the training script, standardization is done manually by computing the mean and standard deviation of the training data as follows:
train_mean = train_epi.mean(dim=0)
train_std = train_epi.std(dim=0)
train_epi = (train_epi - train_mean) / train_std
However, in Section 5* of the tutorial (“Re-load the trained model”), it is mentioned:

“When use the model to do new predictions, if you run the input normalization during the model training, use the same scaler of training data to normalize your new input data.”

Could you please clarify:

Is it necessary to manually save train_mean and train_std during training to apply the same normalization when making new predictions?

If so, do you have any recommended way to store and reuse those parameters (e.g., torch.save() or pickle)?

Would you consider adding this saving/loading process into the codebase or tutorial for reproducibility?

Thanks again for your help and for providing this model to the community!

Best regards,
Bunny

[Yichao-Zhou]

Hi Bunny,

Sorry for the late response!
I tested the model without the normalization, it worked as well as with normalization. The main reason can be that the reference sequence-derived epigenomics features have almost identical distribution as the personalized sequence-derived epigenomic features. Thus, feel free to skip the normalization process, just use the raw epigenomic feature as the input for training. When comes to the prediction task, use the raw epigenomic features too, much more handy.
If you have more questions, feel free to reach out to me via my email: yichaozhou@uchicago.edu.

Thanks!
Charles