HeatMap values don't match between backends

[jbednar]

HoloViews 1.14.8 (current release) shows different output depending on the backend:

import numpy as np, holoviews as hv
hv.extension('bokeh', 'matplotlib')

data = [(i, chr(97+j),  i*j) for i in range(5) for j in range(5) if i!=j]
hm = hv.HeatMap(data).sort()

The Bokeh version looks suspicious to me. Off by one error in indexing the y dimension?

[marcbernot]

The inconsistencies disappear if you define hm = hv.HeatMap(data).sort(by='y'). Does calling sort() is expected to yield the same result?

BUT, if we don't sort hm, they are some inconsistencies with the matplotlib backend.

The graphic and values are correct, but the letters of the y-axis are incorrect. Besides, the shared_axes option is not taken into account when displaying layoutin the following code (this may be tagged as another bug).

import numpy as np, holoviews as hv
import pandas as pd

hv.extension('bokeh', 'matplotlib')

data = [(i, chr(97+j),  i*j) for i in range(5) for j in range(5) if i!=j]
hm = hv.HeatMap(data)

df = pd.DataFrame(data,columns = ['x','y','val']).set_index(['y','x']).sort_index()
hm2 = hv.HeatMap(df,['x','y'])
df.unstack().sort_index(ascending=False)

The data in the dataframe df is

	val
x	0	1	2	3	4
y
e	0.0	4.0	8.0	12.0	NaN
d	0.0	3.0	6.0	NaN	12.0
c	0.0	2.0	NaN	6.0	8.0
b	0.0	NaN	2.0	3.0	4.0
a	NaN	0.0	0.0	0.0	0.0

[marcbernot]

At the moment when element.py calls get_data(self, element, ranges, style), the error is already present since it results in out of sync data and yfactors with yfactors being ['b', 'c', 'd', 'e', 'a']. I'll try to understand and fix this (maybe my first commit :))

[marcbernot]

Ok, the problem is that the data is accessed through hm.gridded.data, i.e.

{'x': array([0, 1, 2, 3, 4]),
 'y': array(['a', 'b', 'c', 'd', 'e'], dtype=object),
 'z': array([[nan,  0.,  0.,  0.,  0.],
        [ 0., nan,  2.,  3.,  4.],
        [ 0.,  2., nan,  6.,  8.],
        [ 0.,  3.,  6., nan, 12.],
        [ 0.,  4.,  8., 12., nan]])}

and the y-axis is made from hm.dimension_values('y',expanded=False) which is array(['b', 'c', 'd', 'e', 'a'], dtype=object) (notice that it is not in the same order as the y value of hm.gridded.data.

Here is a smaller and troubling exemple of the discrepancy.

data_ok = [(1,'b',0),(0,'c',1),(1,'a',2)]
data_nok = [(0,'b',0),(0,'c',1),(1,'a',2)]

h_ok = hv.HeatMap(data_ok)
h_nok = hv.HeatMap(data_nok)
h_ok+h_nok

Strangely, h_ok.gridded.data['y'] is ['b','c','a'] and h_nok.gridded.data['y'] is ['a','b','c'] (which results in the shuffled data since the correct indexing is not taken into account).

I see two options to fix this : 1) Making changes in `hm.gridded so that it enforces that hm.gridded.data['y'] and hm.dimension_values('y',expanded=False) are identical (this is the implicit hypothesis that was broken) 2) Making changes in get_data so as to correctly extract the data (make no implicit hypothesis on hm.gridded.data)

Any opinion on the way to go?