Resuming training with scVI

Help scvi-tools
1

Hi all,

First - thanks for all the effort put in to scvi-tools. Working at a biotech startup, it has helped accelerate our efforts. It’s much appreciated.

I wanted to ask about resuming training with a scVI model (or any model really). In PyTorch, one generally needs both the checkpointed model as well as the optimizer state to resume training. When loading a previously trained scVI model, does it also load the optimizer state such that you would be able to continue training?

Thanks in advance.

2

Sorry, just saw this – I had just answered on GitHub.

This is not possible now – it CAN be possible after we move to pytorch lightning.

I’d be interested to hear from others as well if this is a popular feature request.

3

Sorry for the duplicate entry, @adamgayoso . It makes sense; pytorch lighting is great.

The main reason that such a feature would be useful (for us) is to be able to resume training of scvi once we generate scRNAseq data for new cell types that we are interested in.