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!

[YifanLu2000]

Hi @forrwill, thank you again for raising this issue. I have fixed the bugs related to CPU + sparse calculation in Spateo alignment. I tested the following configuration, and the results look good:

If you encounter any other problems, please feel free to raise an issue.

[YifanLu2000]

Hi @forrwill , I have updated the code in the latest main branch