fail to use the sca.models.SCVI

[melancholy12]

vae = sca.models.SCVI(
adata_ref,
n_layers=2,
encode_covariates=True,
deeply_inject_covariates=False,
use_layer_norm="both",
use_batch_norm="none",
)

ValueError: Expected parameter loc (Tensor of shape (128, 10)) of distribution Normal(loc: torch.Size([128, 10]), scale: torch.Size([128, 10])) to satisfy the constraint Real(), but found invalid values:
tensor([[ nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan],
[ 0.260, 0.540, 0.078, -0.767, 0.169, 0.835, -0.062, 0.543, -0.479,
-0.110],
[ nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan],
[ 0.320, 0.559, 0.204, -0.554, 0.267, 1.067, 0.112, 0.654, -0.443,
-0.260],
[ 0.516, 0.640, -0.024, -0.396, 0.299, 0.797, -0.157, 0.369, -0.568,
-0.112],
[ nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan],
[ nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan],
...,
[ nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan],
[ 0.260, 0.063, -0.155, -0.576, 0.270, 0.842, 0.015, 0.408, -0.723,
-0.015],
[ 0.204, 0.617, 0.028, -0.464, 0.411, 0.688, -0.238, 0.208, -0.392,
0.019],
[-0.025, 0.790, 0.028, -0.647, 0.395, 0.345, -0.034, 0.466, -0.304,
-1.005],
[ 0.406, 0.497, 0.257, -0.780, 0.227, 0.585, 0.059, 0.210, -0.440,
-0.059],
[ nan, nan, nan, nan, nan, nan, nan, nan, nan,
nan],
[ 0.384, 0.563, -0.295, -0.336, 0.161, 0.682, 0.300, 0.584, -0.774,
-0.457]], grad_fn=)

[Koncopd]

Could you provide more information? Maybe your input is not raw counts?

[melancholy12]

@Koncopd No.. It is the processed data.

[Koncopd]

Still could you give me more info? versions of the packages, when exactly this error happens etc.
The data should not be preprocessed, it should contain raw counts.

[melancholy12]

Thanks for your hand. I didn't check the email. Sorry for the delay.
scvi-tools 1.0.2
seaborn 0.12.2
scikit-learn 1.3.0
scanpy 1.9.3
scArches 0.5.9

[sojichld]

Any developments on this issue? I am having the same problem. I have a target dataset with raw data but the atlas does not contain the raw matrix.

[Koncopd]

How do you SCVI.setup_anndata? The raw counts normally should be in adata.X to use with SCVI.

[sojichld]

For me it is sca.models.SCVI.setup_anndata(source_adata, batch_key=condition_key)

This source_data is a reference atlas, that has no adata.raw object. The data was not formatted similarly to the target and some transformation had to be done, so its the best that I have at the moment.

That setup_anndata method will run, the code fails when I run model.train() with the target. From that method I get a similar error to the original issue.

However, it seems that without the raw data available it is not possible to use these methods? I may just have to find alternative methods or backtrack to find some earlier versions of the project.

Update: I was able to do this with just the raw target data. This solved the issue as the atlas has long since been transformed.