Error "TypeError: expected Tensor as element 0 in argument 0, but got numpy.ndarray" when running "st.alignment.morpho_align"

[forrwill]

Hi, I met the error "TypeError: expected Tensor as element 0 in argument 0, but got numpy.ndarray", I had run the function successfully before. But now, It was failed with the error, could you please help me figure out the question.

^M|-----> [Models alignment based on morpho, mode: SN-S.] in progress: 100.0000%|-----> GPU is not available, resorting to torch cpu. |-----> Filtered all samples for common genes. There are 30356 common genes. |-----> Spatial coordinates normalization params: |-----------> Scale: [1357.2966 1357.2966]... |-----------> Scale: [[18417.703 23149.129 241.7397 ] [19898.088 8253.436 331.55087]]... |-----> Preprocess finished. |-----> Performing coarse rigid alignment... |-----> Coarse rigid alignment done.

Traceback (most recent call last): File "./Spateo.test.py", line 60, in align_8W9W, pis_8W9W = st.alignment.morpho_align( File "~/Miniconda3/envs/spateo_env/lib/python3.9/site-packages/spateo_release-1.1.0-py3> P = morpho_model.run() File "~/Miniconda3/envs/spateo_env/lib/python3.9/site-packages/torch/utils/_contextlib.> return func(*args, **kwargs) File "~/Miniconda3/envs/spateo_env/lib/python3.9/site-packages/spateo_release-1.1.0-py3> self._update_assignment_P() File "~/Miniconda3/envs/spateo_env/lib/python3.9/site-packages/spateo_release-1.1.0-py3> self.P, self.K_NA_spatial, self.K_NA_sigma2, sigma2_related = get_P_core( File "~/Miniconda3/envs/spateo_env/lib/python3.9/site-packages/spateo_release-1.1.0-py3> P = _dense_to_sparse( File "~/Miniconda3/envs/spateo_env/lib/python3.9/site-packages/spateo_release-1.1.0-py3> results = _SparseTensor(nx=nx, row=row, col=col, value=val, sparse_sizes=(NA, NB)) File "~/Miniconda3/envs/spateo_env/lib/python3.9/site-packages/spateo_release-1.1.0-py3> return SparseTensor(indices=torch.vstack((row, col)), values=value, size=sparse_sizes) TypeError: expected Tensor as element 0 in argument 0, but got numpy.ndarray

It was confusing that I run the "st.alignment.morpho_align_ref" successfully.

[forrwill]

I reinstalled the spateo-release with "yifan branch". Then , I can run "st.alignment.morpho_align" successfully in cpu mode, But failed with GPU mode,

[YifanLu2000]

Hi, could you provide more information? For example, the function you're running, so I can better understand what's going wrong and help troubleshoot.

[forrwill]

Hi, could you provide more information? For example, the function you're running, so I can better understand what's going wrong and help troubleshoot.

align_8W9W, pis_8W9W = st.alignment.morpho_align( models=[DEOE_8W, DEOE_9W], spatial_key="3d_spatial", key_added="3d_align_spatial", SVI_mode=True, device="0",

sparse_calculation_mode=True,

#use_chunk=True,
#chunk_capacity=2

) This is my code, the error appears when I run "st.alignment.morpho_align"

[YifanLu2000]

Hi, thanks for providing the information. However, the issue seems a bit strange to me since the error appears to be related to sparse calculations. But from the code you've shared, it looks like the sparse_calculation_mode has been commented out. I have two suggestions:

1. For 2D slice alignment: Ensure that the z-axis is not involved. If your z-axis data is in the third dimension, you can handle it like this: slice_a.obsm['spatial_2D'] = slice_a.obsm['spatial_3D'][:, :2].copy() This extracts only the xy-coordinates for alignment.
2. For smaller datasets: If your data isn't too large (e.g., fewer than 10k cells per slice), I recommend not using sparse calculations or chunk operations. In that case, you can comment out the following lines:

   sparse_calculation_mode=True,
   use_chunk=True,
   chunk_capacity=2

   I hope this helps!

[forrwill]

Hi, thanks for providing the information. However, the issue seems a bit strange to me since the error appears to be related to sparse calculations. But from the code you've shared, it looks like the sparse_calculation_mode has been commented out. I have two suggestions:

1. For 2D slice alignment: Ensure that the z-axis is not involved. If your z-axis data is in the third dimension, you can handle it like this: slice_a.obsm['spatial_2D'] = slice_a.obsm['spatial_3D'][:, :2].copy() This extracts only the xy-coordinates for alignment.
2. For smaller datasets: If your data isn't too large (e.g., fewer than 10k cells per slice), I recommend not using sparse calculations or chunk operations. In that case, you can comment out the following lines:

sparse_calculation_mode=True,
use_chunk=True,
chunk_capacity=2

I hope this helps!

when I run the code, I did not commented out the "sparse_calculation_mode=True", and I got the error. Then I reinstalled the spateo and rerun the code. and got the result when I comment out the "sparse_calculation_mode=True"

[YifanLu2000]

Hi, thank you for your feedback! Just to confirm, did this solution work for you?

[forrwill]

Hi, thank you for your feedback! Just to confirm, did this solution work for you?

My data is about 20k~60k cells. if I comment out "sparse_calculation_mode=True". The task would be memory out.

[YifanLu2000]

Hi @forrwill, I apologize for this mistake and thank you for raising the issue. I believe I've found the bug. I noticed that your output includes: GPU is not available, resorting to torch cpu. This means the code is not utilizing the GPU for PyTorch calculations. When sparse_calculation_mode=True is enabled, we rely on torch SparseTensor, which is likely causing the issue. When you used st.alignment.morpho_align_ref, it worked because you likely didn’t enable sparse_calculation_mode=True.

We will need some time to fix this bug since it involves implementing sparse calculations using SciPy’s sparse operations for CPU mode. In the meantime, I suggest checking your PyTorch GPU installation. You can verify if the GPU is set up correctly by running:

import torch
print(torch.cuda.is_available())

Once PyTorch GPU is installed correctly, you should be able to use Spateo alignment without any issues.

In summary, the current issue seems to be with CPU + sparse calculation in Spateo alignment. I hope this helps, and please let me know if you have any other questions.

[forrwill]

Hi @forrwill, I apologize for this mistake and thank you for raising the issue. I believe I've found the bug. I noticed that your output includes: GPU is not available, resorting to torch cpu. This means the code is not utilizing the GPU for PyTorch calculations. When sparse_calculation_mode=True is enabled, we rely on torch SparseTensor, which is likely causing the issue. When you used st.alignment.morpho_align_ref, it worked because you likely didn’t enable sparse_calculation_mode=True.

We will need some time to fix this bug since it involves implementing sparse calculations using SciPy’s sparse operations for CPU mode. In the meantime, I suggest checking your PyTorch GPU installation. You can verify if the GPU is set up correctly by running:

import torch
print(torch.cuda.is_available())

Once PyTorch GPU is installed correctly, you should be able to use Spateo alignment without any issues.

In summary, the current issue seems to be with CPU + sparse calculation in Spateo alignment. I hope this helps, and please let me know if you have any other questions.

In fact, I have tested it in GPU mode, But It was stoped without any error reported, So I am not sure whether it is a bug. Anyway, thank you very much. Please let me know the time point of the update version, Thank you very much!