Extra features for Python's JSON: comments, order, numpy, pandas, datetimes, and many more! Simple but customizable.
Latest commit 6100c79 Jan 21, 2017 mark update description


JSON tricks (python)

The pyjson-tricks package brings several pieces of functionality to python handling of json files:

  1. Store and load numpy arrays in human-readable format.
  2. Store and load class instances both generic and customized.
  3. Store and load date/times as a dictionary (including timezone).
  4. Preserve map order {} using OrderedDict.
  5. Allow for comments in json files by starting lines with #.
  6. Sets, complex numbers, Decimal, Fraction, compression, duplicate keys, ...

As well as compression and disallowing duplicate keys.

The 2.0 series added some of the above features and broke backward compatibility. The version 3.0 series is a more readable rewrite that also makes it easier to combine encoders, again not fully backward compatible.

Several keys of the format __keyname__ have special meanings, and more might be added in future releases.

Installation and use

You can install using

pip install json-tricks  # or e.g. 'json-tricks<3.0' for older versions
pip install numpy        # only if you want to use numpy arrays
pip install pytz         # only if you want timezone-aware datetimes

You can import the usual json functions dump(s) and load(s), as well as a separate comment removal function, as follows:

from json_tricks.np import dump, dumps, load, loads, strip_comments

If you do not have numpy, you should ``import from json_tricks.nonp`` instead.

The exact signatures of these functions are in the documentation.

Preserve type vs use primitive

By default, types are encoded such that they can be restored to their original type when loaded with json-tricks. Example encodings in this documentation refer to that format.

You can also choose to store things as their closest primitive type (e.g. arrays and sets as lists, decimals as floats). This may be desirable if you don't care about the exact type, or you are loading the json in another language (which doesn't restore python types). It's also smaller.

To forego meta data and store primitives instead, pass primitives to dump(s). This is available in version 3.8 and later. Example:

data = [
        arange(0, 10, 1, dtype=int).reshape((2, 5)),
        datetime(year=2017, month=1, day=19, hour=23, minute=00, second=00),
        1 + 2j,
        Fraction(1, 3),
        MyTestCls(s='ub', dct={'7': 7}),  # see later
# Encode with metadata to preserve types when decoding
// (comments added and indenting changed)
        // numpy array
                "__ndarray__": [
                        [0, 1, 2, 3, 4],
                        [5, 6, 7, 8, 9]],
                "dtype": "int64",
                "shape": [2, 5],
                "Corder": true
        // datetime (naive)
                "__datetime__": null,
                "year": 2017,
                "month": 1,
                "day": 19,
                "hour": 23
        // complex number
                "__complex__": [1.0, 2.0]
        // decimal & fraction
                "__decimal__": "42"
                "__fraction__": true
                "numerator": 1,
                "denominator": 3,
        // class instance
                "__instance_type__": [
                "attributes": {
                  "s": "ub",
                  "dct": {"7": 7}
        // set
                "__set__": [0, 1, 2, 3, 4, 5, 6]
# Encode as primitive types; more simple but loses type information
print(dumps(data, primitives=True))
// (comments added and indentation changed)
        // numpy array
        [[0, 1, 2, 3, 4],
        [5, 6, 7, 8, 9]],
        // datetime (naive)
        // complex number
        [1.0, 2.0],
        // decimal & fraction
        // class instance
                "s": "ub",
                "dct": {"7": 7}
        // set
        [0, 1, 2, 3, 4, 5, 6]

Note that valid json is produced either way: json-tricks stores meta data as normal json, but other packages probably won't interpret it.


Numpy arrays

The array is encoded in sort-of-readable and very flexible and portable format, like so:

arr = arange(0, 10, 1, dtype=uint8).reshape((2, 5))
print(dumps({'mydata': arr}))

this yields:

        "mydata": {
                "dtype": "uint8",
                "shape": [2, 5],
                "Corder": true,
                "__ndarray__": [[0, 1, 2, 3, 4], [5, 6, 7, 8, 9]]

which will be converted back to a numpy array when using json_tricks.loads. Note that the memory order (Corder) is only stored in v3.1 and later and for arrays with at least 2 dimensions.

As you've seen, this uses the magic key __ndarray__. Don't use __ndarray__ as a dictionary key unless you're trying to make a numpy array (and know what you're doing).

Numpy scalars are also serialized (v3.5+). They are represented by the closest python primitive type. A special representation was not feasible, because Python's json implementation serializes some numpy types as primitives, without consulting custom encoders. If you want to preverse the exact numpy type, use encode_scalars_inplace.

Performance: this method has slow write times similar to other human-readable formats, although read time is worse than csv. File size (with compression) is high on a relative scale, but it's only around 30% above binary. See this benchmark (it's called JSONGzip). A binary alternative might be added, but is not yet available.

This implementation is inspired by an answer by tlausch on stackoverflow that you could read for details.

Class instances

json_tricks can serialize class instances.

If the class behaves normally (not generated dynamic, no __new__ or __metaclass__ magic, etc) and all it's attributes are serializable, then this should work by default.

# json_tricks/test_class.py
class MyTestCls:
        def __init__(self, **kwargs):
                for k, v in kwargs.items():
                        setattr(self, k, v)

cls_instance = MyTestCls(s='ub', dct={'7': 7})

json = dumps(cls_instance, indent=4)
cls_instance_again = loads(json)

You'll get your instance back. Here the json looks like this:

        "__instance_type__": [
        "attributes": {
                "s": "ub",
                "dct": {
                        "7": 7

As you can see, this stores the module and class name. The class must be importable from the same module when decoding (and should not have changed). If it isn't, you have to manually provide a dictionary to cls_lookup_map when loading in which the class name can be looked up. Note that if the class is imported, then globals() is such a dictionary (so try loads(json, cls_lookup_map=glboals())). Also note that if the class is defined in the 'top' script (that you're calling directly), then this isn't a module and the import part cannot be extracted. Only the class name will be stored; it can then only be deserialized in the same script, or if you provide cls_lookup_map.

If the instance doesn't serialize automatically, or if you want custom behaviour, then you can implement __json__encode__(self) and __json_decode__(self, **attributes) methods, like so:

class CustomEncodeCls:
        def __init__(self):
                self.relevant = 42
                self.irrelevant = 37

        def __json_encode__(self):
                # should return primitive, serializable types like dict, list, int, string, float...
                return {'relevant': self.relevant}

        def __json_decode__(self, **attrs):
                # should initialize all properties; note that __init__ is not called implicitly
                self.relevant = attrs['relevant']
                self.irrelevant = 12

As you've seen, this uses the magic key __instance_type__. Don't use __instance_type__ as a dictionary key unless you know what you're doing.

Date, time, datetime and timedelta

Date, time, datetime and timedelta objects are stored as dictionaries of "day", "hour", "millisecond" etc keys, for each nonzero property.

Timezone name is also stored in case it is set. You'll need to have pytz installed to use timezone-aware date/times, it's not needed for naive date/times.

        "__datetime__": null,
        "year": 1988,
        "month": 3,
        "day": 15,
        "hour": 8,
        "minute": 3,
        "second": 59,
        "microsecond": 7,
        "tzinfo": "Europe/Amsterdam"

This approach was chosen over timestamps for readability and consistency between date and time, and over a single string to prevent parsing problems and reduce dependencies. Note that if primitives=True, date/times are encoded as ISO 8601, but they won't be restored automatically.

Don't use __date__, __time__, __datetime__, __timedelta__ or __tzinfo__ as dictionary keys unless you know what you're doing, as they have special meaning.


Given an ordered dictionary like this (see the tests for a longer one):

ordered = OrderedDict((
        ('elephant', None),
        ('chicken', None),
        ('tortoise', None),

Converting to json and back will preserve the order:

from json_tricks import dumps, loads
json = dumps(ordered)
ordered = loads(json, preserve_order=True)

where preserve_order=True is added for emphasis; it can be left out since it's the default.

As a note on performance, both dicts and OrderedDicts have the same scaling for getting and setting items (O(1)). In Python versions before 3.5, OrderedDicts were implemented in Python rather than C, so were somewhat slower; since Python 3.5 both are implemented in C. In summary, you should have no scaling problems and probably no performance problems at all, especially for 3.5 and later. Python 3.6+ preserve order of dictionaries by default making this redundant, but this is an implementation detail that should not be relied on.


This package uses # and // for comments, which seem to be the most common conventions, though only the latter is valid javascript.

For example, you could call loads on the following string:

{ # "comment 1
        "hello": "Wor#d", "Bye": "\"M#rk\"", "yes\\\"": 5,# comment" 2
        "quote": "\"th#t's\" what she said", // comment "3"
        "list": [1, 1, "#", "\"", "\\", 8], "dict": {"q": 7} #" comment 4 with quotes
// comment 5

And it would return the de-commented version:

        "hello": "Wor#d", "Bye": "\"M#rk\"", "yes\\\"": 5,
        "quote": "\"th#t's\" what she said",
        "list": [1, 1, "#", "\"", "\\", 8], "dict": {"q": 7}

Since comments aren't stored in the Python representation of the data, loading and then saving a json file will remove the comments (it also likely changes the indentation).

The implementation of comments is not particularly efficient, but it does handle all the special cases I could think of. For a few files you shouldn't notice any performance problems, but if you're reading hundreds of files, then they are presumably computer-generated, and you could consider turning comments off (ignore_comments=False).

Other features

  • Sets are serializable and can be loaded. By default the set json representation is sorted, to have a consistent representation.
  • Save and load complex numbers (version 3.2) with 1+2j serializing as {'__complex__': [1, 2]}.
  • Save and load Decimal and Fraction (including NaN, infinity, -0 for Decimal).
  • json_tricks allows for gzip compression using the compression=True argument (off by default).
  • json_tricks can check for duplicate keys in maps by setting allow_duplicates to False. These are kind of allowed, but are handled inconsistently between json implementations. In Python, for dict and OrderedDict, duplicate keys are silently overwritten.

Usage & contributions

Revised BSD License; at your own risk, you can mostly do whatever you want with this code, just don't use my name for promotion and do keep the license file.

Contributions (ideas, issues, pull requests) are welcome!