scvi doublet predictions seem off in comparison to four alternative tools

[robertzeibich]

scvi doublet predictions seem off (too many cells that should not be doublets) in comparison to four alternative tools.

    adata.obs.index.name = None
    scvi.model.SCVI.setup_anndata(adata)
    vae = scvi.model.SCVI(adata)
    vae.train() # batch_size=130

    solo = scvi.external.SOLO.from_scvi_model(vae)
    solo.train(batch_size=130) # batch_size=130
    # vae.train(batch_size=130) and solo.train(batch_size=130) # ValueError: Expected more than 1 value per channel when training, got input size torch.Size([1, 128])
    # change batch_size when error is thrown
    
    doublet_prediction = solo.predict()
    doublet_prediction['prediction'] = solo.predict(soft = False)

    adata.obs = adata.obs.reset_index().merge(doublet_prediction.reset_index(),on='index',how='left').set_index('index')
    adata.obs.index.name = None

[Paste the error output produced by the above code here]

Versions:

1.2.0

[canergen]

Hi, can you show the history of the solo model? Did it converge? I assume the data is from a single batch?

[robertzeibich]

Thank you for getting back to me.

I am not entirely sure what you mean with converge.
Using solo on a single batches only.

I have attached one output file:
log-3713343_20_scvi_doublets.txt

Is there a way to increase the number of workers?

I suspect that this is the same tool (https://github.com/calico/solo/tree/master). Is that correct?