**Breaking**: data_kind: Now 'matrix' represents a 2-D numpy array and unrecognized data types fall back to 'vectors'

[seisman]

Need to wait for #3481.

Description of proposed changes

As can be seen in the doctests added in #3480, currently, matrix and vectors kinds are defined like below:

• vectors kind means data=None and required=True. It means that the input is given via a series of vectors (e.g., x/y/z).
• matrix represents any data types not recognized as arg/file/stringio/grid/image/geojson. The most common data types include pd.DataFrame, pd.Series, xr.Dataset, np.ndarray, and sequence of sequences.

So, matrix here is an inclusive concept. However, in both Python and GMT, matrix usually means a homogenous 2-D array. When passing a "matrix" data to GMT, we usually treat the "matrix" as a series of vectors. The only exception is when the "matrix" data is a strict matrix (a 2-D numpy array) with data.dtype.kind in "iuf".

pygmt/pygmt/clib/session.py

Lines 1742 to 1755 in d7560fa

	elifkind=="matrix": # turn 2-D arrays into list of vectors
	ifhasattr(data, "items") andnothasattr(data, "to_frame"):
	# pandas.DataFrame or xarray.Dataset types.
	# pandas.Series will be handled below like a 1-D numpy.ndarray.
	_data= [arrayfor_, arrayindata.items()]
	elifhasattr(data, "ndim") anddata.ndim==2anddata.dtype.kindin"iuf":
	# Just use virtualfile_from_matrix for 2-D numpy.ndarray
	# which are signed integer (i), unsigned integer (u) or
	# floating point (f) types
	_virtualfile_from=self.virtualfile_from_matrix
	_data= (data,)
	else:
	# Python list, tuple, numpy.ndarray, and pandas.Series types
	_data=np.atleast_2d(np.asanyarray(data).T)

I think it makes more sense to use a more strict definition for matrix and let all other data types be vectors. I propose to change the definitions of data kinds to:

• matrix: a 2-D homogenous numpy.ndarray
• vectors: fallbacks to this kind for any unrecognized data types, including a dictionary, pd.DataFrame, pd.Series, xr.Dataset, nested lists/tuples, or any other unrecognized data types.

The refactoring is mainly done in 808755d.

Of course, data=None, required=True is still recognized as "vectors". I also think we should give this case another special kind like "none", which makes more sense and can also simplify the codes. [This is done in a separate PR at #3482, but I'm OK to merge #3482 into this PR first before merging this PR into main].

Other data kinds are unchanged.

This is a breaking change since some previous "matrix" data are now recognized "vectors" and previously "vectors" is "none". But I guess very few users are using this function, so the breaking change should have minimal effects.

[seisman]

We probably can change this line to:

Suggested change

	ifisinstance(data, np.ndarray) anddata.ndim==2:
	ifhasattr(data, "__array_interface__") andlen(data.__array_interface__.shape)==2:

to support more array-like objects that implements the __array_interface__ protocol (https://numpy.org/doc/stable/reference/arrays.interface.html#object.__array_interface__), but it can be done in a future PR.

[seisman]

pandas.DataFrame now is "vectors" kind.

[weiji14]

Two things:

1. Should we extend kind="matrix" to 3-D numpy.ndarray objects?
2. I'm hesitant to match based on np.ndarray() class only, because while it's not widely advertised, checking for data is not None would have caught other array types implementing the __array_function__ protocol (see NEP18), but checking for isinstance(data, np.ndarray) would exclude those. That said, I'm wondering if the fallback to kind="vectors" might just work, but need to test this out more extensively.

[seisman]

• Should we extend kind="matrix" to 3-D numpy.ndarray objects?

It's unclear if we can pass a 3-D numpy array yet. Even if we can, it means more work, since in virtualfile_from_vectors/virtualfile_from_matrix, we explicitly check if the ndim=1 or 2. So, better to revisit the 3-D numpy array support in a separate PR if necessary.

• I'm hesitant to match based on np.ndarray() class only, because while it's not widely advertised, checking for data is not None would have caught other array types implementing the __array_function__ protocol (see NEP18), but checking for isinstance(data, np.ndarray) would exclude those. That said, I'm wondering if the fallback to kind="vectors" might just work, but need to test this out more extensively.

Yes, I was thinking about checking __array_interface__ in #3351 (comment).

[seisman]

Done in 4a4f192, although other array-like objects are not tested.

[weiji14]

Missing test coverage for these lines?

[seisman]

I can add a test for it. Before adding more tests, I'm wondering if we should split the big "test_clib_virtualfiles.py" file (with more than 500 lines) into separate smaller test files, i.e., one test file for each Session method.

• test_clib_virtualfile_in
• test_clib_virtualfile_from_vectors
• test_clib_virtualfile_from_matrix
• test_clib_open_virtualfile

[weiji14]

We've split it before in #2784, so yes, ok to split it again 😆

[seisman]

We've split it before in #2784, so yes, ok to split it again 😆

Done in #3512.

[seisman]

I've added a test in cfa32ed to cover this line.

[seisman]

This test is added to address #3351 (comment).

This new test passes a string dtype numpy array into GMT C API, which contains longitude/latitude strings. The data kind is "matrix", but since the data dtype is not in iuf, PyGMT will use virtualfile_from_vectors rather than virtualfile_from_matrix to pass the data into GMT C API. Ideally, we should check that the virtualfile_from_vectors is called once and virtualfile_from_matrix is not called, but I find it technically complicated with unittest.mock, so I'll leave it untested.

The function is marked as xfail with GMT 6.5 due to a newly found upstream bug at GenericMappingTools/gmt#8600.

[weiji14]

Good catch, and thanks for adding the test!