Given that we had bug reports when this as changed, yet the changes passed all the tests, either the tests are incomplete or this is accommodating undocumented use of colours. Can we add a test and or document whatever this is supposed to fix?

The problem is the original 'color comparison' code in the figure legend method didn't allow for colors to be numpy arrays. Long term I think this should be avoided by a general function for comparing colors, which I've put in#10032

Yeah, the problem is that this code didn'tchange in 2.1.1, it just got applied toFigure.legend() as well asaxes.legend(). So we broke things by making them more consistent. But I'm still not understanding the use-case.

If in master I do

importmatplotlib.pyplotaspltimportnumpyasnpfig,ax=plt.subplots()ax.plot(np.arange(10),color=np.array([0.1,0.3,0.4]),label='boo')fig.legend()plt.show()

That works fine. So 1-D numpy arrays work OK in master.

The error comes when the user specifies a 2-D array

importmatplotlib.pyplotaspltimportnumpyasnpfig,ax=plt.subplots()ax.plot(np.arange(10),color=np.array([[0.1,0.3,0.4]]),label='boo')fig.legend()plt.show()

ValueError: Invalid RGBA argument: array([[ 0.1,  0.3,  0.4]])

I'd argue specifying a 2-D array is not correct, and if it worked before, that was just a fortunate happenstance. OTOH,plot lets you put a 2-D array in, so I agree legend should as well.

Anyway, upshot is we should add a test to test for 2-D arrays being allowed.

OK, I take it back; the above 2-D array test fails in 2.1.0 and 2.0.2 as well. So I'm not clear what the failure case is here. What color spec failed in 2.1.1 that this fixes?

ping@skirpichev

@jklymak, sorry. I can't provide better failing test yet.

OK< let us know.

I can reproduce

ValueError: Invalid RGBA argument: array([[ 0.12156863,  0.46666667,  0.70588235,  1.        ]])

but I get that all the way back to 2.0.2, so I don't think this every worked?

If you could just print c before its passed to legend, that would help clarify things.

@dstansby's fix above seems to work for you, but the minimal test I put in abovedoesn't work on@dstansby's change. So I am mystified. Most importantly we should make sure we have a test that covers your case and the case that is breaking forsympy/sympy#13730

Most importantly we should make sure we have a test that covers your case and the case that is breaking forsympy/sympy#13730

Probably, these testcases should be same, as Diofant and Sympy share some codebase.

See here for a link to their efforts:sympy/sympy#13730 (comment)

I'll put in a TST PR for this.