Generated idata fails to be deserialized

[carrascomj]

I have two idata.json outputs generated by maud sample (maud-metabolic-models==0.4.0.2) and none of them can be read back to memory.

Traceback:

>>> idata = az.from_json("idata.json")
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/jorge/.virtualenvs/maud/lib/python3.9/site-packages/arviz/data/io_json.py", line 29, in from_json
    idata_dict = json.load(file)
ujson.JSONDecodeError: Unexpected character found when decoding 'Infinity'

Without ujson (using the stdlib json module), it also fails but with a more cryptic error.

[carrascomj]

As a workaround

from glob import glob
from maud.loading_maud_inputs import load_maud_input
from maud.getting_idatas import get_idata

mi = load_maud_input("user_input")
idata = get_idata(glob("samples/*csv"), mi, "train")

works fine.

[teddygroves]

I think this might be the cause of the problem: ultrajson/ultrajson#562.

The ujson library relatively recently changed how it encodes infinity from 'Inf' to 'Infinity'. When it tries to read a file with the old 'Inf's it raises the JSONDecodeError above.

I had a look in the idata.json file and it had 'Inf's (i.e. not 'Infinity's) in it, so I suspect the ujson reading the file might be incompatible with the one that made it.

I tried (with ujson 5.7.0 and arviz 0.14.0) doing the following and it worked ok:

import arviz as az
import numpy as np

t = az.from_dict({"bla": np.inf})
t.to_json("t.json")
t_loaded = az.from_json("t.json")

and this a) didn't raise an error and b) produced a file containing 'Infinity'.

I think we should maybe require a recent version of ujson in Maud in order to stop this issue coming up again.