Pandas series with object type get converted to strings

[cliffckerr]

This error is a slightly different take on #407 and #358.

My issue is that if a dataframe has a column of mixed type (e.g. [4, 'foo']), then it will be converted to strings on unpickling (e.g. ['4', 'foo']):

import pandas as pd
import jsonpickle as jp
import jsonpickle.ext.numpy as jsonpickle_numpy
import jsonpickle.ext.pandas as jsonpickle_pandas
jsonpickle_numpy.register_handlers()
jsonpickle_pandas.register_handlers()

# Create simple data frame of mixed data types
df1 = pd.DataFrame((
  a = [1, 2], 
  b = [4, 'foo']
))

# Convert to jsonpickle
string = jp.dumps(df1)
df2 = jp.loads(string)

# Show that it didn't work
print(df1.b.values)
print(df2.b.values)

assert df1.b[0] == df2.b[0] # False: the first is 4 and the second is '4'

I know there isn't an easy fix for this, but I'm hopeful that there's some fix, since (at least in this case) the actual dataframe data looks a lot like JSON format already! In particular, one suggestion I have is that currently just a single dtype is stored per column:

"meta": {"dtypes": {"a": "int64", "b": "object"}

My suggestion that I believe would solve this particular bug would be that if the dtype is object, to store a list (one for each element) instead. For example, in this case:

"meta": {"dtypes": {"a": "int64", "b": {"object": ["int", "object"]}}

Personally, I also think storing values as a dict of lists (rather than as a string) would be more robust and easier to read/interpret. I feel pd.DataFrame.to_dict() is already pretty close to what would be required!