Avoid subsetting cells in NMF initialisation

[krademaker]

In this NMF initialisation, cells are subsetted.

However, I noticed that if cells are actually subsetted, this will lead to inconsistent tensor dimensions when running the actual NMF model. E.g.: RuntimeError: The size of tensor a (270000) must match the size of tensor b (766783) at non-singleton dimension 0.

For a dataset:

         Summary Statistics          
┏━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━┓
┃     Summary Stat Key     ┃ Value  ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━┩
│         n_cells          │ 766783 │
│          n_vars          │  474   │
│         n_batch          │   27   │
│ n_extra_categorical_covs │   1    │
│ n_extra_continuous_covs  │   0    │
│    n_var_categoricals    │   1    │
└──────────────────────────┴────────┘

And the NMF loadings initialised for only 270000 cells.

[AlexanderAivazidis]

I am surprised about this. Didn't it work previously even when you used less cells for subsampling? I think it did for me. But either way in the end I effectively never used subsampling because i set the number of cells per category for subsampling to a very large number.