Expected value argument (Tensor of shape (2500, 12666))

[nine-sarayut]

• I have confirmed this bug exists on the latest version of cell2location. See https://github.com/BayraktarLab/cell2location#installation
• I follow the instructions from the scvi-tools tutorial.

---

Note: Please read this guide detailing how to provide the necessary information for us to reproduce your bug.

Minimal code sample (that we can run without your data, using public data)

adata_ref = sc.read(
    f'./data/2717b0a0-948e-475b-87b2-818af9a68c9d.h5ad', 
    backup_url='https://datasets.cellxgene.cziscience.com/2717b0a0-948e-475b-87b2-818af9a68c9d.h5ad'
)

from cell2location.utils.filtering import filter_genes
selected = filter_genes(adata_ref, cell_count_cutoff=5, cell_percentage_cutoff2=0.03, nonz_mean_cutoff=1.12)

# filter the object
adata_ref = adata_ref[:, selected].copy()

cell2location.models.RegressionModel.setup_anndata(adata=adata_ref, 
                        batch_key='donor_id', 
                        labels_key='cell_type'
)

# create the regression model
from cell2location.models import RegressionModel
mod = RegressionModel(adata_ref) 

# view anndata_setup as a sanity check
mod.view_anndata_setup()

mod.train(max_epochs=250, accelerator='gpu')

ValueError: Expected value argument (Tensor of shape (2500, 12666)) to be within the support (IntegerGreaterThan(lower_bound=0)) of the distribution GammaPoisson(), but found invalid values:
...
         alpha_g_inverse dist            |  1 12666
                        value            |  1 12666
                     log_prob            |         
             data_target dist 2500 12666 |         
                        value 2500 12666 |    

[nine-sarayut]

I can fix it now if you save it before normalizing.

# Saving count data
adata.layers["counts"] = adata.X.copy()

by using this code.

import numpy as np

# Check if counts layer exists
print("Available layers:", adata_ref.layers.keys())

# Copy raw counts from 'counts' layer to .X
adata_ref.X = adata_ref.layers['counts'].copy()

# Verify you now have raw counts
print(f"Max value: {adata_ref.X.max()}")
print(f"Contains integers: {np.all(adata_ref.X.data == np.floor(adata_ref.X.data))}")