Creating a nested dictionary from a flattened dictionary

I have a flattened dictionary which I want to make into a nested one, of the form

flat = {'X_a_one': 10,
        'X_a_two': 20, 
        'X_b_one': 10,
        'X_b_two': 20, 
        'Y_a_one': 10,
        'Y_a_two': 20,
        'Y_b_one': 10,
        'Y_b_two': 20}

I want to convert it to the form

nested = {'X': {'a': {'one': 10,
                      'two': 20}, 
                'b': {'one': 10,
                      'two': 20}}, 
          'Y': {'a': {'one': 10,
                      'two': 20},
                'b': {'one': 10,
                      'two': 20}}}

The structure of the flat dictionary is such that there should not be any problems with ambiguities. I want it to work for dictionaries of arbitrary depth, but performance is not really an issue. I've seen lots of methods for flattening a nested dictionary, but basically none for nesting a flattened dictionary. The values stored in the dictionary are either scalars or strings, never iterables.

So far I have got something which can take the input

test_dict = {'X_a_one': '10',
             'X_b_one': '10',
             'X_c_one': '10'}

to the output

test_out = {'X': {'a_one': '10', 
                  'b_one': '10', 
                  'c_one': '10'}}

using the code

def nest_once(inp_dict):
    out = {}
    if isinstance(inp_dict, dict):
        for key, val in inp_dict.items():
            if '_' in key:
                head, tail = key.split('_', 1)

                if head not in out.keys():
                    out[head] = {tail: val}
                else:
                    out[head].update({tail: val})
            else:
                out[key] = val
    return out

test_out = nest_once(test_dict)

But I'm having trouble working out how to make this into something which recursively creates all levels of the dictionary.

Any help would be appreciated!

(As for why I want to do this: I have a file whose structure is equivalent to a nested dict, and I want to store this file's contents in the attributes dictionary of a NetCDF file and retrieve it later. However NetCDF only allows you to put flat dictionaries as the attributes, so I want to unflatten the dictionary I previously stored in the NetCDF file.)

Here is my take:

def nest_dict(flat):
    result = {}
    for k, v in flat.items():
        _nest_dict_rec(k, v, result)
    return result

def _nest_dict_rec(k, v, out):
    k, *rest = k.split('_', 1)
    if rest:
        _nest_dict_rec(rest[0], v, out.setdefault(k, {}))
    else:
        out[k] = v

flat = {'X_a_one': 10,
        'X_a_two': 20, 
        'X_b_one': 10,
        'X_b_two': 20, 
        'Y_a_one': 10,
        'Y_a_two': 20,
        'Y_b_one': 10,
        'Y_b_two': 20}
nested = {'X': {'a': {'one': 10,
                      'two': 20}, 
                'b': {'one': 10,
                      'two': 20}}, 
          'Y': {'a': {'one': 10,
                      'two': 20},
                'b': {'one': 10,
                      'two': 20}}}
print(nest_dict(flat) == nested)
# True

output = {}

for k, v in source.items():
    # always start at the root.
    current = output

    # This is the part you're struggling with.
    pieces = k.split('_')

    # iterate from the beginning until the second to last place
    for piece in pieces[:-1]:
       if not piece in current:
          # if a dict doesn't exist at an index, then create one
          current[piece] = {}

       # as you walk into the structure, update your current location
       current = current[piece]

    # The reason you're using the second to last is because the last place
    # represents the place you're actually storing the item
    current[pieces[-1]] = v

Here's one way using collections.defaultdict, borrowing heavily from this previous answer. There are 3 steps:

1. Create a nested defaultdict of defaultdict objects.
2. Iterate items in flat input dictionary.
3. Build defaultdict result according to the structure derived from splitting keys by _, using getFromDict to iterate the result dictionary.

This is a complete example:

from collections import defaultdict
from functools import reduce
from operator import getitem

def getFromDict(dataDict, mapList):
    """Iterate nested dictionary"""
    return reduce(getitem, mapList, dataDict)

# instantiate nested defaultdict of defaultdicts
tree = lambda: defaultdict(tree)
d = tree()

# iterate input dictionary
for k, v in flat.items():
    *keys, final_key = k.split('_')
    getFromDict(d, keys)[final_key] = v

{'X': {'a': {'one': 10, 'two': 20}, 'b': {'one': 10, 'two': 20}},
 'Y': {'a': {'one': 10, 'two': 20}, 'b': {'one': 10, 'two': 20}}}

As a final step, you can convert your defaultdict to a regular dict, though usually this step is not necessary.

def default_to_regular_dict(d):
    """Convert nested defaultdict to regular dict of dicts."""
    if isinstance(d, defaultdict):
        d = {k: default_to_regular_dict(v) for k, v in d.items()}
    return d

# convert back to regular dict
res = default_to_regular_dict(d)

The other answers are cleaner, but since you mentioned recursion we do have other options.

def nest(d):
    _ = {}
    for k in d:
        i = k.find('_')
        if i == -1:
            _[k] = d[k]
            continue
        s, t = k[:i], k[i+1:]
        if s in _:
            _[s][t] = d[k]
        else:
            _[s] = {t:d[k]}
    return {k:(nest(_[k]) if type(_[k])==type(d) else _[k]) for k in _}

You can use itertools.groupby:

import itertools, json
flat = {'Y_a_two': 20, 'Y_a_one': 10, 'X_b_two': 20, 'X_b_one': 10, 'X_a_one': 10, 'X_a_two': 20, 'Y_b_two': 20, 'Y_b_one': 10}
_flat = [[*a.split('_'), b] for a, b in flat.items()]
def create_dict(d): 
  _d = {a:list(b) for a, b in itertools.groupby(sorted(d, key=lambda x:x[0]), key=lambda x:x[0])}
  return {a:create_dict([i[1:] for i in b]) if len(b) > 1 else b[0][-1] for a, b in _d.items()}

print(json.dumps(create_dict(_flat), indent=3))

Output:

{
 "Y": {
    "b": {
      "two": 20,
      "one": 10
    },
    "a": {
      "two": 20,
      "one": 10
    }
 },
  "X": {
     "b": {
     "two": 20,
     "one": 10
   },
    "a": {
     "two": 20,
     "one": 10
   }
 }
}