How are the built - in models trained?

[Lucyyang1991]

Thank you very much for the built-in model you provided. It is very useful, and the cell types in it are very rich. However, in the actual use process, we often need to create subsets of the model. The subset function you provided is also very useful. But as you said before, just using subset is not enough. The best way is to take the original data of the model, perform subset operation, and then retrain. So, can you nicely provide the training code for the built-in model? In this way, when we perform subset operation, we can try to reproduce it and avoid a large deviation in cell types. In addition, do you consider providing us with functions or code for fast training after subset operation? After all, it is indeed a large amount of work to start retraining from finding the original data.

[ChuanXu1]

@Lucyyang1991, almost all built-in models were trained using simiar code without subsetting: model = celltypist.train(adata, 'anno_column', n_jobs = -1, feature_selection = True, top_genes = 200/300, details = 'NNN', url = 'https://celltypist.cog.sanger.ac.uk/models/NNN/vX/NNN.pkl', source = 'https://doi.org/NNN', version = 'vX')

After subsetting your cells, you can perform exactly the same code above to train your adata subset. If you have further subsetted your genes, you can modify the above code by feature_selection = False, check_expression = False