cuda out of memory

[yuling999666]

Hi @YifanLu2000, when you run spateo for lymph nodes, did you write codes to solve the problem of CUDA out of memory? It stuck on 50% progress on my server. It would be nice if you can give some suggestions. Thanks.

[YifanLu2000]

Hi @yuling999666, I’d like to know the size of your GPU memory and the number of spots/cells in your dataset. You can refer to this tutorial to reduce GPU memory usage: Improve efficiency and scalability.

[yuling999666]

Hi @yuling999666, I’d like to know the size of your GPU memory and the number of spots/cells in your dataset. You can refer to this tutorial to reduce GPU memory usage: Improve efficiency and scalability.

Hi @YifanLu2000 , thanks for your reply. Our GPU server installed a Nvidia A6000 48GB graphic card there. It also has 256 GB RAM. I have tried the tutorial of sparse computation, CUDA still out of memory. The lymph nodes have spots number about 50k-80k each slice. Thank you for your time.

[YifanLu2000]

Hi @yuling999666, did you try using the sparse calculation?

[yuling999666]

Hi @YifanLu2000, Thank you for the reply. I tried but it still out of memory. It seems that only downsampling works out right now.

[YifanLu2000]

Hi @yuling999666, when performing alignment, are you using the entire transcriptome (e.g., 20k features)? Using too many gene expression features can significantly increase GPU memory usage. If this is the case, you might want to try using HVG (highly variable genes) or PCA features instead.

[yuling999666]

Hi @YifanLu2000 , thank you for your reply. Yes, I used the entire transcriptome. I will try to use PCA/HVG instead. One more question for Figure 3A in paper. Did you provide colormap for plotting this figure in your material? I have checked GitHub of spateo-notebooks, but I didn't find it. I think you guys did great work and it's a nice visualization. If you didn't provide in the material, could you give me the color map for plotting Hemibrain data? Thanks!

[YifanLu2000]

Hi @yuling999666, I use "Spectral" colormap for the Hemibrain data:

pc, _ = st.tdr.construct_pc(adata=adata, spatial_key=spatial_key, groupby='cell_type', key_added="tissue", colormap='Spectral',)

[yuling999666]

Hi @YifanLu2000. Got you. Thank you! I appreciate it.

[yuling999666]

Hi @YifanLu2000 , thank you for providing the colormap. Did you provide the colormap for Lymph nodes instead of Hemibrain? I have used the colormap but it is quite different from the plot in figure 3A for Hemibrain. I think the colormap is very close to figure 3B. Could you check? Thank you for your time and consideration.

[YifanLu2000]

Hi @yuling999666, I have confirmed that I am using the Spectral colormap for Hemibrain, and also the same for Lymph nodes. You may check if the annotation is the same.

[yuling999666]

Hi @YifanLu2000, thank you for providing annotation. I will check. I appreciate it!