Description

alugowski

opened on Aug 25, 2023

Issue body actions

https://pypi.org/project/matrepr/

Benefits:

• output looks like a matrix, not a dataframe
• faster, especially on first run where the pandas import is slow
• small code

See it in action with python-graphblas:
https://nbviewer.org/github/alugowski/matrepr/blob/main/doc/demo-python-graphblas.ipynb

To try yourself, use matrepr's jupyter integration (causes Jupyter to use matrepr's formatter instead of the native one):

import matrepr.jupyter

For the Python REPL, a __repr__ monkey patch does the same:

import matrepr.patch.graphblas

Reacted by Sultan Orazbayev and Erik Welch

Activity

1. eriknw commented on Aug 25, 2023

   

   eriknw

   on Aug 25, 2023

   Member

   More actions

   I like it! I/we will definitely take a closer look. Thanks @alugowski

   It definitely looks more like a matrix, and I was thinking just the other day that pandas is a rather heavy dependency that we only use for our repr.

   How well does (or will or can) it handle more advanced numpy dtypes? For example:

   • np.dtype([("x", np.bool_), ("y", np.int64)], align=True)
     • We currently render an element of this dtype as e.g. (False, 2)
   • np.dtype("(3,)uint16")
     • We currently render an element of this dtype as e.g. [1, 2, 3]
2. alugowski commented on Aug 26, 2023

   

   alugowski

   on Aug 26, 2023

   ContributorAuthor

   More actions

   How well does (or will or can) it handle more advanced numpy dtypes? For example:

   * `np.dtype([("x", np.bool_), ("y", np.int64)], align=True)`
     
     * We currently render an element of this dtype as e.g. `(False, 2)`
   
   * `np.dtype("(3,)uint16")`
     
     * We currently render an element of this dtype as e.g. `[1, 2, 3]`
   

   I had to do a small patch, but both those now render the same in all three output types (html, latex, string). It just calls str(), which I assume is what you do too.
3. alugowski commented on Aug 26, 2023

   

   alugowski

   on Aug 26, 2023

   ContributorAuthor

   More actions

   * `np.dtype("(3,)uint16")`
     
     * We currently render an element of this dtype as e.g. `[1, 2, 3]`
   

   One question about this one.

   Do you prefer it to be rendered like that string, or as a nested submatrix?

   For reference, here is how I currently render some interesting values:
   

   Row 3 is nested matrices of types matrepr supports (lists, lists of lists, scipy matrices. a graphblas matrix would look like that too).

   I ask because I'm considering adding dense numpy arrays as a supported package. Since AFAIK objects of type np.dtype("(3,)uint16") appear just like any other numpy array, one option is to render it like a submatrix instead of str().
4. eriknw commented on Aug 26, 2023

   

   eriknw

   on Aug 26, 2023

   Member

   More actions

   Ooh, nested sub-matrix is pretty cool!

   Right now we just do what is easiest (hence using pandas in the first place). For submatrix dtypes, the easiest is to turn them into lists, vals.tolist(), and let pandas render each entry as a Python list.

   I'm happy to revisit how we render elements and to try what you think is nicest :)
5. alugowski commented on Aug 28, 2023

   

   alugowski

   on Aug 28, 2023

   ContributorAuthor

   More actions

   Sounds good!

   I added ndarray support so numpy submatricies get rendered all pretty now, too:

   

   Let me know how it looks on your datasets. Happy to make tweaks!
6. alugowski commented on Aug 28, 2023

   

   alugowski

   on Aug 28, 2023

   ContributorAuthor

   More actions

   BTW this is all that is required to use MatRepr with python-graphblas:

   def _str_(mat):
       from matrepr import to_str
       from matrepr.adapters.graphblas_driver import GraphBLASDriver
       # Enable terminal width detection
       return to_str(GraphBLASDriver.adapt(mat), width_str=0, max_cols=9999)
   
   
   graphblas.Matrix.__repr__ = _str_
   graphblas.Vector.__repr__ = _str_

   Reacted by Sultan Orazbayev
7. alugowski commented on Sep 7, 2023

   

   alugowski

   on Sep 7, 2023

   ContributorAuthor

   More actions

   One newer feature is to support custom row/column labels. This is handy for something like an adjacency matrix where the matrix elements are edge weights and row/column labels are the vertex weights.

   

   Reacted by Sultan Orazbayev

Issue actions