Repository navigation
I broke imshow(<signed integer array>) :-( #6825
Description
mpl2.0b3
In [4]: plt.imshow(np.random.randint(-128, 128, dtype=np.int8, size=(3, 3)))
Out[4]: <matplotlib.image.AxesImage at 0x7fe00ba2f0f0>
In [5]: /usr/lib/python3.5/site-packages/matplotlib/colors.py:848: RuntimeWarning: overflow encos
resdat /= (vmax - vmin)
and a fully black image appears.
Pretty sure this is due to #6700 and can be fixed by restoring the casts of vmin and vmax to a floating type. However that type should be determined as in process_value (i.e., np.float32 if dtype.itemsize <= 2 else np.promote_types(dtype, float) -- see the issue for the discussion).
Not making a PR yet because this (as well as the original #6700) may or may not affect the other Normalize classes too? Perhaps that's something that should be factored out.
Activity
efiring commentedon Jul 25, 2016
@anntzer, it looks like a more general solution to these problems is needed, starting with input normalization: float128 should be converted to float64 as early as possible, float32 should be preserved and used when possible, etc. Maybe the thing to do is revert #6700, and develop a more comprehensive solution, which need not block 2.0. Does this make sense to you? And if so, can you lead it?
anntzer commentedon Jul 25, 2016
Sure, I'll give it a try. Continuing the discussion in that thread.
Merge pull request #6837 from anntzer/fix-integer-normalization