Hi @malmaud ,

Thanks for the PR! Here are a couple of thoughts

1. In addition to the DataArrayValidator (which handles properties like scatter.x), several of the other validator types also accept arrays (e.g. the ColorValidator can sometimes accept numeric arrays for properties like scatter.marker.color). So to be consistent we should try to handle all array cases if we can.
2. To support (1), I'd like to work out a way to do the coercion to numpy arrays inside the copy_to_readonly_numpy_array function, this way we can reuse the kind, force_numeric, and 'WRITEABLE' logic.
3. In order to accomplish (1) and (2), while preserving our current error messages, I think we would need to find a way to have is_homogeneous_array determine whether it's appropriate to convert non-numpy non-pandas objects into numpy arrays.

I haven't looked into this too much yet, but do PyTorch tensors and xarray DataArrays support the numpy array interface? (https://docs.scipy.org/doc/numpy-1.15.1/reference/arrays.interface.html). If so, we might be able to handle this usecase across all array types by adding a simple check to is_homogeneous_array to check for the presence of an __array_interface__ method.

Thanks for the feedback! I'll implement 1) and 2).

3. is a bit subtle - there are a few different ways that numpy.array tries to convert an object to an array - it could, as you said, use the array interface. It will fall back to just seeing if the object has an __array__ method that directly returns a numpy array. Or it can also use the Python buffer interface. Or an earlier version of the array interface. I wouldn't want to re-implement that entire logic. PyTorch tensors, for instance, have an __array__ method but don't support the array interface.

I think just checking for __array_interface__ or __array__ methods on objects is probably sufficient for 99% of cases though - the coercion could always be made more general later if we see objects in the wild that this heuristic isn't sufficient for.

I think just checking for __array_interface__ or __array__ methods on objects is probably sufficient for 99% of cases though - the coercion could always be made more general later if we see objects in the wild that this heuristic isn't sufficient for.

👍

OK, I made the changes.

Thanks @malmaud,

In terms of testing, it looks like something went wrong in plotly.tests.test_optional.test_figure_factory.test_figure_factory.TestTrisurf. Could you have a look?

======================================================================
ERROR: test_trisurf_all_args (plotly.tests.test_optional.test_figure_factory.test_figure_factory.TestTrisurf)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/circleci/project/plotly/tests/test_optional/test_figure_factory/test_figure_factory.py", line 889, in test_trisurf_all_args
    x, y, z, simplices
  File "/home/circleci/project/plotly/figure_factory/_trisurf.py", line 466, in create_trisurf
    edges_color=edges_color, plot_edges=plot_edges)
  File "/home/circleci/project/plotly/figure_factory/_trisurf.py", line 165, in trisurf
    showlegend=False
  File "/home/circleci/project/plotly/graph_objs/_scatter3d.py", line 1939, in __init__
    self['marker'] = marker if marker is not None else _v
  File "/home/circleci/project/plotly/basedatatypes.py", line 2820, in __setitem__
    self._set_compound_prop(prop, value)
  File "/home/circleci/project/plotly/basedatatypes.py", line 3129, in _set_compound_prop
    val = validator.validate_coerce(val, skip_invalid=self._skip_invalid)
  File "/home/circleci/project/_plotly_utils/basevalidators.py", line 2117, in validate_coerce
    v = self.data_class(v, skip_invalid=skip_invalid)
  File "/home/circleci/project/plotly/graph_objs/scatter3d/_marker.py", line 1064, in __init__
    self['color'] = color if color is not None else _v
  File "/home/circleci/project/plotly/basedatatypes.py", line 2829, in __setitem__
    self._set_prop(prop, value)
  File "/home/circleci/project/plotly/basedatatypes.py", line 3065, in _set_prop
    val = validator.validate_coerce(val)
  File "/home/circleci/project/_plotly_utils/basevalidators.py", line 1140, in validate_coerce
    invalid_els = self.find_invalid_els(v, validated_v)
  File "/home/circleci/project/_plotly_utils/basevalidators.py", line 1169, in find_invalid_els
    self.find_invalid_els(orig_el, validated_el, invalid_els)
  File "/home/circleci/project/_plotly_utils/basevalidators.py", line 1167, in find_invalid_els
    for orig_el, validated_el in zip(orig, validated):
TypeError: zip argument #1 must support iteration

Before merging this, we'll also need some tests that exercise the new __array_interface__ or __array__ logic. I think the easiest thing here might be to create a test file that mirrors _plotly_utils/tests/validators/test_pandas_series_input.py but for a library that plotly.py doesn't know about like xarray. To do this, we'd need to add xarray to

• optional-requirements.txt
• As an optional: dependency in tox.ini
• To the conda create command in .circleci/create_conda_optional_env.sh

Are you interested in adding these tests? If not, we can just leave the PR open until I have a chance to go through and add them (which would hopefully be by the end of the month).

Alright, I'll take a look and add the tests.

Tests have been added. I had to make some other changes to make the tests pass, so apparently that was a worthwhile exercise.

This is great work @malmaud, thanks for taking this on and especially for adding the tests! I had two inline questions, but everything else lgtm.

Merged! Thanks again for the solid PR.

Please see #1427