Introduced a format for displaying the DCSR_matrix

[Sai-Suraj-27]

Description

Added a __str__() method for printing the DCSR_matrix.
PyTorch displays the sparse tensors in this format:

The introduced format displays the following:

Issue/s resolved: #1040

Changes proposed:

• Added a method for printing the DCSR_matrix.

Type of change

• New feature (non-breaking change which adds functionality)

Due Diligence

• All split configurations tested
• Multiple dtypes tested in relevant functions
• Documentation updated (if needed)
• Title of PR is suitable for corresponding CHANGELOG entry

Does this change modify the behaviour of other functions? If so, which?

No

[github-actions]

Thank you for the PR!

[github-actions]

Thank you for the PR!

[mrfh92]

@Sai-Suraj-27 Regarding the introduced format I have a question: your picture of the PyTorch print result shows that indices are given by a 2D-tensor (which makes sense because specifying the position of a non-zero entry in a matrix requires us to state row- and column-index). In your print result for DCSR, the index-tensor is 1D. Is this due to the CSR-format?

[Sai-Suraj-27]

@Sai-Suraj-27 Regarding the introduced format I have a question: your picture of the PyTorch print result shows that indices are given by a 2D-tensor (which makes sense because specifying the position of a non-zero entry in a matrix requires us to state row- and column-index). In your print result for DCSR, the index-tensor is 1D. Is this due to the CSR-format?

@mrfh92 Yes, sir. In PyTorch's sparse tensor representation, the indices are stored as a 2D tensor. On the other hand, In CSR format, the indices are stored as a 1D tensor, where each element represents the column index of a non-zero entry.
We can also add and display the indptr array which will tell the number of non-zero entries in each row and from that and the indices tensor one can easily find out the (row, col) of all non-zero elements. Please tell me your opinion and if you want me to change anything else in this format of displaying the DCSR_matrix.

[mrfh92]

Hi @Sai-Suraj-27,
thanks for your fast anwer :) I would like to suggest to include the indptr as well such that the output of print gives the user full information needed to "reconstruct" the tensor. However, as I remember @ClaudiaComito has mentored the implementation of the sparse-module so far; so maybe its a better idea to wait until next week when she is back since she can give a more detailed feedback.

[Sai-Suraj-27]

Hi @Sai-Suraj-27, thanks for your fast anwer :) I would like to suggest to include the indptr as well such that the output of print gives the user full information needed to "reconstruct" the tensor. However, as I remember @ClaudiaComito has mentored the implementation of the sparse-module so far; so maybe its a better idea to wait until next week when she is back since she can give a more detailed feedback.

Ok, sir. I will wait for the feedback.

[github-actions]

Thank you for the PR!

[github-actions]

Thank you for the PR!

[mrfh92]

Hi @Sai-Suraj-27,
I think it would be the best option to leave indices and data completely away when displaying a sparse DNDarray. So is suffices to display that it is an object of class DCSR_matrix, its size, split and the number of non-zero entires nzz. The reason is that usually sparse arrays are used in a context where the data is so large that displaying single entries does not make sense anymore.

[Sai-Suraj-27]

Hi @Sai-Suraj-27, I think it would be the best option to leave indices and data completely away when displaying a sparse DNDarray. So is suffices to display that it is an object of class DCSR_matrix, its size, split and the number of non-zero entires nzz. The reason is that usually sparse arrays are used in a context where the data is so large that displaying single entries does not make sense anymore.

@mrfh92 sir, so do you want me to update the function such that we will be printing like this:
DCSR_matrix(size=(3, 3), nnz=8, split=None) (without indptr, data, and indices)

[github-actions]

Thank you for the PR!

[Mystic-Slice]

It would be better if this function is moved to the core/printing module. There are cases like local print vs global print to be handled. See the __str__ method of the DNDarray class for more information.

And it is usually best to use the printing format used by PyTorch since it gives more information to the user. The information outputted by __str__ should prioritize user-readability more than anything else. Just the indptr and indices would be too difficult to understand. I, personally, would prefer to see the exact coordinates being printed out. I say this because the users of Heat shouldn't be expected to know about the CSR matrix and its format. They should just be able to use this more-efficient data structure without having to learn too much about it.

[Mystic-Slice]

What @mrfh92 says also seems acceptable to me. Maybe the users won't want to see the actual data points. But my only problem with this solution is, the users wouldn't be able to see the data points even if they wanted to. There is no other way for them to get the coordinates if we don't show it in the print output.

[github-actions]

This pull request is stale because it has been open for 60 days with no activity.

[github-actions]

This pull request was closed because it has been inactive for 60 days since being marked as stale.