-
-
Notifications
You must be signed in to change notification settings - Fork 402
idata to plotting data
Ari Hartikainen edited this page Aug 21, 2020
·
2 revisions
To have a consistent interface for (realtime) plotting, we should have a good idea of how to do the following
- common data structure for data --> plots
- possibility to update data "easily"
This can be done by translating idata --> pandas.DataFrame (wide format)
- On the Bokeh side, dataframe can be inserted into ColumnDataSource (CDS) object.
- On the matplotlib side, dataframe can be used as a
data
source.
Some plots enable us to update plots with raw data, some need a bit more processing and some need a lot of processing. DataFrame contains GroupBy functionality which can be used for grouped plotting (e.g. by-chain).
Here are some code how to translate idata to dataframe. It is still a bit brittle and needs some idea how to access coordinates information. (Maybe a class could be used)
import numpy as np
import pandas as pd
def idata_to_dataframe(idata):
"""InferenceData to DataFrame by groups.
Skips data groups.
"""
idict = idata.to_dict()
dims = idict["dims"]
coords = idict["coords"]
idfs = {}
coords_info = {}
for group, values in idict.items():
if group in ("dims", "coords"):
continue
if group in ("observed_data", "constant_data", "attrs"):
continue
group_dict = {}
chain = draw = None
for variable, array in values.items():
if chain is None:
chain, draw = array.shape[:2]
group_dict["chain"] = (np.ones((chain, draw), dtype=int) * np.arange(chain)[:, None]).ravel()
group_dict["draw"] = (np.ones((chain, draw), dtype=int) * np.arange(draw)).ravel()
if not array.shape[2:]:
group_dict[variable] = array.ravel()
continue
for idx in np.ndindex(array.shape[2:]):
key = f"{variable}[{','.join(map(str, idx))}]"
idx_tuple = (slice(None), slice(None)) + idx
group_dict[key] = array[idx_tuple].ravel()
if variable in dims:
var_dims = dims[variable]
for i, dim in enumerate(var_dims):
if dim in coords:
coord_value = coords[dim][idx[i]]
coords_info[(group, key)] = (dim, coord_value)
group_df = pd.DataFrame.from_dict(group_dict, orient="columns")
idfs[group] = group_df
return idfs, coords_info
Here is a rough idea what different parts different plots have
pair
parallel
density
distcomparison
dist
energy
forest
hdi
joint
kde
posterior
ppc
trace
violin
autocorr
ess
mcse
rank
compare
bpv
elpd
khat
loopit
And categorized in a different way:
-
raw
--> scatter, rug -
density
--> kde / hist autocorr
loo
raw
density
density[contour]
raw
density
density
density
density
density
density[boxplot]
density
raw
density
density
density
density
raw
density
autocorr
density
autocorr
autocorr
autocorr
autocorr
loo
stats
loo
loo
loo