Can you provide the full example with the timedelta64 array that produces that error? I've encountered similar looking problems before, that depend on when we do our unit conversions, and this might actually be fixed in the master branch.

@dstansby - the full example is:

import numpy as np
import matplotlib.pyplot as plt
x = np.array(range(100), dtype='M8[D]')
ax = plt.subplot(1,1,1)
y = np.random.random(100)
ax.bar(x[:-1], y[:-1], width=np.diff(x))

The problem is our unit conversion. We do

x = self.convert_xunits(x)
width = self.convert_xunits(width)

However, we don't have a conversion implemented for timedelta64. A result, x is our internal float-based date format while width is still a timedelta64, which cannot be subtracted from each other.

The solution would be to implement a converter for timedelta64.

Ah, so this would be fixed by something like #9120 I think.

Sort of. It needs a bit more infrastructure:

• implement a converter analogously to DateConverter
• add it to matplotlib.units.registry

I'm having a look at this now, but having trouble with your example @astrofrog ; when I do

import numpy as np
x = np.array(range(100), dtype='M8[D]')
print((x[:-1] + np.diff(x) / 2.0) - x[:-1])

I just get a bunch of zeros; dividing np.diff(x) by a number seems to just zero it out. Do you know of a way to fix this? Internally Matplotlib has to divide the width by 2, which is why I can't get your example working.

'M8[D]' has a single day resolution, and your array goes up by one, so dividing by two rounds down to 0. You'd need to at least change it to hourly resolution.

@dstansby, not sure of the approach you are taking here, but I think as @timhoffm says, we should add timedelta64 to dates.py, and register it. Then all the math can be done in floating point numbers...

Trying to take that approach, but running into the problem that np.timedelta64 is a subclass of numbers.Number, so the conversion interface assumes it doesn't need converting:

import numpy as np
import matplotlib.pyplot as plt
import matplotlib.units as munits
x = np.array(range(100), dtype='M8[D]')
xdiff = np.diff(x)
print(xdiff.dtype)
print(munits.ConversionInterface.is_numlike(xdiff))

prints

timedelta64[D]
True