JSONLab: portable, robust JSON/binary-JSON encoder/decoder

A toolbox to encode/decode JSON/UBJSON/MessagePack files in MATLAB/Octave
Updated 28 Mar 2024

View License

Editor's Note: Popular File 2018

This file was selected as MATLAB Central Pick of the Week

JSONLab: compact, portable, robust JSON/binary-JSON encoder/decoder for MATLAB/Octave
We strongly encourage all users help us fill out a brief community survey to help developers better integrate JSONLab with our JSON-based scientific data resources at https://neurojson.io, registered users can receive future critical updates
Compatible Specifications
What's New
We are excited to announce that the JSONLab project, as the official reference library for both JData and BJData specifications, has been funded by the US National Institute of Health (NIH) as part of the NeuroJSON project (https://neurojson.org and https://neurojson.io).
The goal of the NeuroJSON project is to develop scalable, searchable, and reusable neuroimaging data formats and data sharing platforms. All data produced from the NeuroJSON project will be using JSON/Binary JData formats as the underlying serialization standards and the lightweight JData specification as language-independent data annotation standard, all of which have been evolved from the over a decade development of JSONLab.
JSONLab v2.9.8 - code named "Micronus Prime - beta" - is the beta-release of the next milestone (v3.0), containing a number of key feature enhancements and bug fixes. The major new features include
  1. exporting JSON Memory-Map (jsonget,jsonset) for rapid disk-map like reading/writing of JSON/binary JSON files and writing, implementing JSON-Mmap spec v1 Draft 1
  2. supporting JSONPath query (jsonpath) to MATLAB data and JSON/binary JSON file and streams, including deep-scan operators,
  3. (breaking) upgrading the supported BJData spec to V1 Draft 2 where the default numerical data byte order changed from Big-Endian to Little-Endian,
  4. adding initial support to JData _DataLink_ decoding to link multiple JSON/binary JSON files
  5. dynamically cache linked data files (jsoncache, jdlink) to permit on-demand download and processing of complex JSON-encoded datasets such as neuroimaging datasets hosted on https://neurojson.io
  6. support high-performance Blosc2 meta-compressor for storing large N-D array data,
  7. savejson/loadjson can use MATLAB/Octave built-in jsonencode/jsondecode using the BuiltinJSON option
  8. automatically switch from struct to containers.Map when encoded key-length exceeds 63
  9. provide fall-back zlib/gzip compression/decompression function (octavezmat) on Octave when ZMat is not installed
  10. include built-in .nii/.nii.gz/.jnii/.h5/.snirf/.tsv/.csv parsers to allow loadjd.m to read wide range of files
  11. include json2couch from jbids (https://github.com/NeuroJSON/jbids) to allow uploading json files to CouchDB server
There have been many major updates added to this release since the previous release v2.0 in June 2020. A list of the major changes are summarized below (with key features marked by *), including the support to BJData Draft-2 specification, new interface functions savejd/loadjd, and options to use MATLAB/Octave built-in jsonencode/jsondecode functions. The octave-jsonlab package has also been included in the official distributions of Debian Bullseye and Ubuntu 21.04 or newer.
  • 2024-03-28 [b39c374] [feat] add json2couch from jbids toolbox
  • 2024-03-27*[2e43586] [feat] merge nii/jnii/hdf5/tsv reading functions self-contained
  • 2024-03-27 [b482c8f] [test] pass all tests on matlab R2010b
  • 2024-03-27 [2008934] [doc] additional documentations on decompression functions
  • 2024-03-27 [0a582fb] [doc] add documentations for jsonpath, jsoncache, jdlink, and maxlinklevel
  • 2024-03-27 [5dba1de] [bug] .. searches deep level of struct, make jdlink work for Octave 4.2 and 5.2
  • 2024-03-27 [fea481e] [doc] add line-by-line comment on examples, add jsonset/jsonget
  • 2024-03-26 [e1d386d] [feat] support saving dictionary to json and bjdata
  • 2024-03-26 [dfc744b] [feat] support caching data from any URL using hash, add NO_ZMAT flag
  • 2024-03-24 [22d297e] [doc] fix README.rst formatting issues
  • 2024-03-24 [7e27db5] [doc] update documentation, preparing for v2.9.8 release
  • 2024-03-24 [1227a0b] [format] reformat
  • 2024-03-24 [67f30ca] [feat] support using . or [] in JSONPath to escape dots in key names
  • 2024-03-24 [ee830cd] [bug] fix error_pos error when giving a non-existant input file
  • 2024-03-24 [d69686d] [feat] add jdlink to dynamically download and cache linked data
  • 2024-03-22 [772a1ef] [ci] fix octave failed test
  • 2024-03-22*[cff529a] [test] add jsonpath test, refine jsonpath syntax support
  • 2024-03-22 [22435e4] [bug] fix jsonpath handling of recursive deep scans
  • 2024-03-21 [c9f8a20] [bug] support deep scan in cell and struct, merge struct/containers.Map
  • 2024-03-21 [394394a] [bug] improve jsonpath cell with deep scan
  • 2024-03-20 [a599e71] [feat] add jsoncache to handle _DataLink_ download cache, rename jsonpath
  • 2024-02-19*[4f2edeb] [feat] support .. jsonpath operator for deep scan
  • 2024-01-11 [c43a758] [bug] fix missing index_esc reset, add test for automap
  • 2024-01-11*[ef5b472] [feat] automatically switch to map object when key length > 63
  • 2023-11-17 [ee24122] use sprintf to replace unescapejsonstring
  • 2023-11-12 [abe504f] [ci] test again on macos-12
  • 2023-11-12 [d2ff26a] [ci] install octave via conda on macos to avoid hanged install
  • 2023-11-07 [33263de] completely reformat m-files using miss_hit
  • 2023-11-07 [3ff781f] make octavezmat work on matlab
  • 2023-10-29 [ea4a4fd] make test script run on MATLAB R2010b
  • 2023-10-27 [ca91e07] use older matlab due to matlab-actions/run-command#43
  • 2023-10-27 [4bf8232] add NO_ZMAT flag, fix fread issue
  • 2023-10-27*[ce3c0a0] add fallback zlib/glib support on Octave via file-based zip/unzip
  • 2023-10-26 [7ab1b6e] fix error for expecting an ending object mark when count is given
  • 2023-09-08 [6dfa58e] Fix typos found by codespell
  • 2023-06-27 [7d7e7f7] fix typo of compression method
  • 2023-06-27*[c25dd0f] support blosc2 codecs in save and load data, upgrade jsave/jload
  • 2023-06-19 [b23181a] test root-level indentation
  • ...
  • detailed update logs can be found at https://github.com/fangq/jsonlab/releases/tag/v2.9.8
Please note that the savejson/loadjson in both JSONLab v2.0-v3.0 are compliant with JData Spec Draft 3; the savebj/loadbj in JSONLab v3.0 is compatible to BJData spec Draft 2, which contains breaking feature changes compared to those in JSONLab v2.0.
The BJData spec was derived from UBJSON spec Draft 12, with the following breaking differences:
  • BJData adds 4 new numeric data types: uint16 [u], uint32 [m], uint64 [M] and float16 [h] (supported in JSONLab v2.0 or newer)
  • BJData supports an optimized ND array container (supported in JSONLab since 2013)
  • BJData does not convert NaN/Inf/-Inf to null (supported in JSONLab since 2013)
  • BJData Draft 2 changes the default byte order to Little-Endian instead of Big-Endian (JSONLab 3.0 or later)
  • BJData only permits non-zero-fixed-length data types as the optimized array type, i.e. only UiuImlMLhdDC are allowed
To avoid using the new features, one should attach 'UBJSON',1 and 'Endian','B' in the savebj command as
savebj('',data,'FileName','myfile.bjd','UBJSON',1, 'Endian','B');
To read BJData data files generated by JSONLab v2.0, you should call
You are strongly encouraged to convert all pre-v2.9 JSONLab generated BJD or .pmat files using the new format.
JSONLab is an open-source JSON/UBJSON/MessagePack encoder and decoder written completely in the native MATLAB language. It can be used to convert most MATLAB data structures (array, struct, cell, struct array, cell array, and objects) into JSON/UBJSON/MessagePack formatted strings and files, or to parse a JSON/UBJSON/MessagePack file into a MATLAB data structure. JSONLab supports both MATLAB and GNU Octave (a free MATLAB clone).
Compared to other MATLAB/Octave JSON parsers, JSONLab is uniquely lightweight, ultra-portable, producing dependable outputs across a wide-range of MATLAB (tested on R2008) and Octave (tested on v3.8) versions. It also uniquely supports BinaryJData/UBJSON/MessagePack data files as binary-JSON-like formats, designed for efficiency and flexibility with loss-less binary storage. As a parser written completely with the native MATLAB language, it is surprisingly fast when reading small-to-moderate sized JSON files (1-2 MB) with simple hierarchical structures, and is heavily optimized for reading JSON files containing large N-D arrays (known as the "fast array parser" in loadjson).
JSON (JavaScript Object Notation) is a highly portable, human-readable and "fat-free" text format to represent complex and hierarchical data, widely used for data-exchange in applications. UBJSON (Universal Binary JSON) is a binary JSON format, designed to specifically address the limitations of JSON, permitting the storage of binary data with strongly typed data records, resulting in smaller file sizes and fast encoding and decoding. MessagePack is another binary JSON-like data format widely used in data exchange in web/native applications. It is slightly more compact than UBJSON, but is not directly readable compared to UBJSON.
We envision that both JSON and its binary counterparts will play important roles for storage, exchange and interoperation of large-scale scientific data among the wide-variety of tools. As container-formats, they offer both the flexibility and generality similar to other more sophisticated formats such as HDF5, but are significantly
simpler with a much greater software ecosystem.
Towards this goal, we have developed the JData Specification (http://github.com/NeuroJSON/jdata) to standardize serializations of complex scientific data structures, such as N-D arrays, sparse/complex-valued arrays, trees, maps, tables and graphs using JSON/binary JSON constructs. The text and binary formatted JData files are syntactically compatible with JSON/UBJSON formats, and can be readily parsed using existing JSON and UBJSON parsers. JSONLab is not just a parser and writer of JSON/UBJSON data files, but one that systematically converts complex scientific data structures into human-readable and universally supported JSON forms using the standardized JData data annotations.
The installation of JSONLab is no different from installing any other MATLAB toolbox. You only need to download/unzip the JSONLab package to a folder, and add the folder's path to MATLAB/Octave's path list by using the following command:
If you want to add this path permanently, you can type pathtool, browse to the JSONLab root folder and add to the list, then click "Save". Then, run rehash in MATLAB, and type which savejson, if you see an output, that means JSONLab is installed for MATLAB/Octave.
If you use MATLAB in a shared environment such as a Linux server, the best way to add path is to type
mkdir ~/matlab/
nano ~/matlab/startup.m
and type addpath('/path/to/jsonlab') in this file, save and quit the editor. MATLAB will execute this file every time it starts. For Octave, the file you need to edit is ~/.octaverc, where ~ is your home directory.
To use the data compression features, please download the ZMat toolbox from https://github.com/NeuroJSON/zmat/releases/latest and follow the instruction to install ZMat first. The ZMat toolbox is required when compression is used on MATLAB running in the -nojvm mode or GNU Octave, or 'lzma/lzip/lz4/lz4hc' compression methods are specified. ZMat can also compress large arrays that MATLAB's Java-based compression API does not support.
Install JSONLab on Fedora 24 or later
JSONLab has been available as an official Fedora package since 2015. You may install it directly using the below command
sudo dnf install octave-jsonlab
To enable data compression/decompression, you need to install octave-zmat using
sudo dnf install octave-zmat
Then open Octave, and type pkg load jsonlab to enable jsonlab toolbox.Install JSONLab on Debian
JSONLab is currently available on Debian Bullseye. To install, you may run
sudo apt-get install octave-jsonlab
One can alternatively install matlab-jsonlab if MATLAB is available.Install JSONLab on Ubuntu
JSONLab is currently available on Ubuntu 21.04 or newer as package octave-jsonlab. To install, you may run
sudo apt-get install octave-jsonlab
For older Ubuntu releases, one can add the below PPA
To install, please run
sudo add-apt-repository ppa:fangq/ppa
sudo apt-get update
to add this PPA, and then use
sudo apt-get install octave-jsonlab
to install the toolbox. octave-zmat will be automatically installed.Install JSONLab on Arch Linux
JSONLab is also available on Arch Linux. You may install it using the below command
sudo pikaur -S jsonlab
Using JSONLab
JSONLab provides a pair of functions, loadjson -- a JSON parser, and savejson -- a MATLAB-to-JSON encoder, to read/write the text-based JSON; it also provides three equivalent pairs -- loadbj/savebj for binary JData, loadubjson/saveubjson for UBJSON and loadmsgpack/savemsgpack for MessagePack. The load* functions for the 3 supported data formats share almost the same input parameter format, similarly for the 3 save* functions (savejson/saveubjson/savemsgpack). These encoders and decoders are capable of processing/sharing almost all data structures supported by MATLAB, thanks to jdataencode/jdatadecode - a pair of in-memory data converters translating complex MATLAB data structures to their easy-to-serialized forms according to the JData specifications. The detailed help information can be found in the Contents.m file.
In JSONLab 2.9.8 and later versions, a unified file loading and saving interface is provided for JSON, binary JSON and HDF5, including loadjd and savejd for reading and writing below files types:
  • JSON based files: .json, .jdt (text JData file), .jmsh (text JMesh file), .jnii (text JNIfTI file), .jnirs (text JSNIRF file)
  • BJData based files: .bjd, .jdb (binary JData file), .bmsh (binary JMesh file), .bnii (binary JNIfTI file), .bnirs (binary JSNIRF file), .pmat (MATLAB session file)
  • UBJSON based files: .ubj
  • MessagePack based files: .msgpack
  • HDF5 based files: .h5, .hdf5, .snirf (SNIRF fNIRS data files) - require EasyH5 toolbox
In the below section, we provide a few examples on how to us each of the core functions for encoding/decoding JSON/Binary JSON/MessagePack data.
jsonmesh=struct('MeshNode',[0 0 0;1 0 0;0 1 0;1 1 0;0 0 1;1 0 1;0 1 1;1 1 1],...
'MeshElem',[1 2 4 8;1 3 4 8;1 2 6 8;1 5 6 8;1 5 7 8;1 3 7 8],...
'MeshSurf',[1 2 4;1 2 6;1 3 4;1 3 7;1 5 6;1 5 7;...
2 8 4;2 8 6;3 8 4;3 8 7;5 8 6;5 8 7],...
'MeshCreator','FangQ','MeshTitle','T6 Cube',...
'SpecialData',[nan, inf, -inf]);
% convert any matlab variables to JSON (variable name is used as the root name)
% convert matlab variables to JSON with a root-name "jmesh"
% an empty root-name directly embed the data in the root {}
% the compact=1 flag prints JSON without white-space in a single-line
% if 3 inputs are given, the 3rd parameter defines the output file name
% param/value pairs can be provided after the 2nd input to customize outputs
% if you want to use params/values and save JSON to a file, you must use the 'filename' to set output file
% jsonlab utilizes JData annotations to encode complex/sparse ND-arrays
% when setting 'BuiltinJSON' to 1, savejson calls jsonencode.m in MATLAB (R2016+)
% or Octave (v7+) to convert data to JSON; this is typically faster, but does not
% support all features native savejson offers
savejson('cpxrand',eye(5)+1i*magic(5), 'BuiltinJSON', 1)
% JData annotations also allows one to compress binary strongly-typed data and store in the JSON
% gzip/zlib are natively supported in MATLAB and Octave; using ZMat toolbox, one can use lz4, lzma, blosc2 etc compressors
% 'ArrayToStruct' flag forces all arrays to use the JData ND array annotations to preserve types
% JData supports compact storage of special matrices using the '_ArrayShape_' annotation
% loadjson can directly parse a JSON string if it starts with "[" or "{", here is an empty object
% loadjson can also parse complex JSON objects in a string form
% if the input is a file name, loadjson reads the file and parse the data inside
dat=loadjson(['examples' filesep 'example1.json'])
% param/value pairs can be used following the 1st input to customize the parsing behavior
dat=loadjson(['examples' filesep 'example1.json'],'SimplifyCell',0)
% if a URL is provided, loadjson reads JSON data from the URL and return the parsed results,
% similar to webread, except loadjson calls jdatadecode to decode JData annotations
% using the 'BuildinJSON' flag, one can use the built-in jsondecode.m in MATLAB (R2016+)
% or Octave (7.0+) to parse the JSON data for better speed, note that jsondecode encode
% key names differently compared to loadjson
dat=loadjson('{"_obj":{"string":"value","array":[1,2,3]}}', 'builtinjson', 1)
% when the JSON data contains long key names, one can use 'UseMap' flag to
% request loadjson to store the data in a containers.Map instead of struct (key name limited to 63)
dat=loadjson('{"obj":{"an object with a key longer than 63":"value","array":[1,2,3]}}', 'UseMap', 1)
% loadjson can further download the linked data pointed by `_DataLink_` tag, and merge with the parent
% a JSONPath can be attached to the URL to retrieve a sub element
% loadjson can optionally return a JSON-memory-map object, which defines each JSON element's
% memory buffer offset and length to enable disk-map like fast read/write operations
[dat, mmap]=loadjson('{"obj":{"key":"value","array":[1,2,3]}}')
% if set 'mmaponly' to 1, loadjson only returns the JSON-mmap structure
mmap=loadjson('{"obj":{"key":"value","array":[1,2,3]}}', 'mmaponly', 1)
% savebj works almost exactly like savejson, except that the output is the more compact binary JSON
a={single(rand(2)), struct('va',1,'vb','string'), 1+2i};
% customizing the root-name using the 1st input, and the 3rd input setting the output file
% enabling the 'debug' flag to allow printing binary JSON in text-form, helping users to run tests or troubleshoot
savebj('rootname',a, 'debug',1)
% like savejson, savebj also allow data compression for even more compact storage
% binary JSON does not need base64-encoding, therefore, the output can be ~33% smaller than text-based JSON
[length(savebj('magic',magic(100),'Compression','zlib')), length(savejson('magic',magic(100),'Compression','zlib'))]
% savebj can output other popular binary JSON formats, such as MessagePack or UBJSON
savebj('mesh',a,'FileName','meshdata.msgpk','MessagePack',1) % same as calling savemsgpack
savebj('mesh',a,'FileName','meshdata.ubj','UBJSON',1) % same as calling saveubjson
% similarly, loadbj does almost exactly the same as loadjson, but it parses binary JSON instead
obj=struct('string','value','array',single([1 2 3]),'empty',[],'magic',uint8(magic(5)));
% loadbj can load a binary JSON (BJData - a derived format from UBJSON) object from a buffer
% you can test if loadbj parsed object still matches the data saved using savebj
% similarly, savebj/loadbj can compress/decompress binary array data using various compressors
% if given a path to a binary JSON file (.jdb,.bnii,.pmat,.jmsh,...), it opens and parses the file
% loadbj can directly load binary JSON data files from URL, here is a binary-JSON based NIfTI file
% similar to loadjson, loadbj can also return JSON-memory-map to permit disk-map
% like direct reading/writing of specific data elements
[dat, mmap]=loadbj(ubjdata)
mmap=loadbj(ubjdata, 'mmaponly', 1)
% jdataencode transforms complex MATLAB data structures (ND-array, sparse array, complex arrays,
% table, graph, containers.Map etc) into JSON-serializable forms using portable JData annotations
% here, we show how to save a complex-valued sparse array using JSON JData annotations
testdata = struct('a',rand(5)+1i*rand(5),'b',[],'c',sparse(5,5));
% when setting 'annotatearray' to 1, jdataencode uses _ArrayType_/_ArraySize_/_ArrayData_
% JData tags to store ND array to preserve data types; use 'prefix' to customize variable name prefix
% when setting 'usearrayshape' to 1, jdataencode can use _ArrayShape_ to encode special matrices
% jdatadecode does the opposite to jdataencode, it recognizes JData annotations and convert
% those back to MATLAB native data structures, such as ND-arrays, tables, graph etc
% we can test that the decoded data are the same as the original
% if one uses jsondecode to parse a JSON object, the output JData annotation name prefix is different
% jsondecode adds "x_" as prefix
rawdecode_builtin = jsondecode(savejson('',rawdata));
% in comparison, loadjson calls encodevarname.m, producing "x0x5F_" as prefix (hex for '_')
% encodevarname encoded names can be reversed to original decodevarname.m
rawdecode_jsonlab = loadjson(savejson('',rawdata), 'jdatadecode', 0);
% savejd is a unified interface for savejson/savebj/savemsgpack/saveh5 depending on the output file suffix
a={single(rand(2)), struct('va',1,'vb','string'), 1+2i};
savejd('', a, 'test.json')
savejd('', a, 'test.jdb')
savejd('', a, 'test.ubj')
savejd('', a, 'test.h5')
% loadjd is a unified interface for loadjson/loadbj/loadmsgpack/loadh5/load/loadjnifti depending on the input file suffix
% supported types include .json,.jnii,.jdt,.jmsh,.jnirs,.jbids,.bjd,.bnii,.jdb,.bmsh,.bnirs,.ubj,.msgpack,
% .h5,.hdf5,.snirf,.pmat,.nwb,.nii,.nii.gz,.tsv,.tsv.gz,.csv,.csv.gz,.mat,.bvec,.bval; input can be an URL
data = loadjd('test.json');
data = loadjd('test.jdb');
data = loadjd('test.ubj');
data = loadjd('test.h5');
data = loadjd('file:///path/to/test.jnii');
data = loadjd('https://neurojson.org/io/stat.cgi?action=get&db=abide&doc=CMU_b&file=0a429cb9101b733f594eefc1261d6985-zlib.bnii');
% loadjson/loadbj JSON-memory-map (mmap) output returned by loadjson or loadbj
% each mmap contains a pair of JSONPath and two numbers [offset, length] of the object in bytes in the buffer/file
jsonstr = '{"obj":{"string":"value","array":[1,2,3]}}';
mmap=loadjson(jsonstr, 'mmaponly', 1)
% mmap = [ ["$",[1,42]], ["$.obj",[8,34]], ["$.obj.string",[18,7]], ["$.obj.array",[34,7]] ]
% this means there are 4 objects, root '$', with its content starting byte 1, with a length of 42 bytes;
% content of object '$.obj' starts byte 8, with a length of 34 bytes
% using the above mmap, jsonget can return any raw data without needing to reparse jsonstr
% below command returns '[1,2,3]' as a string by following the offset/length data in mmap
jsonget(jsonstr, mmap, '$.obj.array')
% you can request multiple objects by giving multiple JSONPath keys
jsonget(jsonstr, mmap, '$.obj', '$.obj.string')
% you can request multiple objects by giving multiple JSONPath keys
jsonget(jsonstr, mmap, '$.obj', '$.obj.string')
% jsonget not only can fast reading a JSON string buffer, it can also do disk-map read of a file
mmap = loadjson('/path/to/data.json', 'mmaponly', 1);
jsonget('/path/to/data.json', mmap, '$.obj')
% using JSON mmap, one can rapidly modify the content of JSON object pointed by a path
jsonstr = '{"obj":{"string":"value","array":[1,2,3]}}';
mmap=loadjson(jsonstr, 'mmaponly', 1)
% we can rewrite object $.obj.array by changing its value '[1,2,3]' to a string "test"
% this returns the updated jsonstr as '{"obj":{"string":"value","array":"test" }}'
% the new value of a key must not have longer bytes than the original value
jsonset(jsonstr, mmap, '$.obj.array', '"test"')
% one can change multiple JSON objects, below returns '{"obj":{"string":"new" ,"array":[] }}'
jsonset(jsonstr, mmap, '$.obj.string', '"new"', '$.obj.array', '[]')
% if mmap is parsed from a file, jsonset can perform disk-map like fast writing to modify the json content
mmap = loadjson('/path/to/data.json', 'mmaponly', 1);
jsonset('/path/to/data.json', mmap, '$.obj.string', '"new"', '$.obj.array', '[]')
% JSONPath is a widely supported standard to index/search a large struct, such as those loaded from a JSON file
% the jsonpath.m function implements a subset of the features
% the below command returns the value of obj.key subfield, which is "value"
obj = loadjson('{"obj":{"key":"value1","array":[1,2,3],"sub":{"key":"value2","array":[]}}}');
jsonpath(obj, '$.obj.key')
% using [] operator, one can also index array elements, index start from 0; the output below is 2
jsonpath(obj, '$.obj.array[1]')
% [] operator supports range, for example below commands yields [1,2]
jsonpath(obj, '$.obj.array[0:1]')
% a negative index in [] counting elements backwards, -1 means the last element
jsonpath(obj, '$.obj.array[-1]')
% jsonpath.m supports JSONPath's deep-scan operator '..', it traverses through the struct
% and find all keys following .., here the output is {"value1", "value2"}
jsonpath(obj, '$.obj..key')
% you can further concatenate JSONPath operators to select outputs from the earlier ones, this outputs {'value2'}
jsonpath(obj, '$.obj..key[1]')
% instead of .keyname, you can use [keyname], below command is the same as above
jsonpath(obj, '$[obj]..[key][1]')
% one can escape special char, such as ".", in the key using special\.key or [special.key]
jsonpath(obj, '$.obj.special\.key.sub')
% the _DataLink_ annotation in the JData specification permits linking of external data files
% in a JSON file - to make downloading/parsing externally linked data files efficient, such as
% processing large neuroimaging datasets hosted on http://neurojson.io, we have developed a system
% to download files on-demand and cache those locally. jsoncache.m is responsible of searching
% the local cache folders, if found the requested file, it returns the path to the local cache;
% if not found, it returns a SHA-256 hash of the URL as the file name, and the possible cache folders
% When loading a file from URL, below is the order of cache file search paths, ranking in search order
% global-variable NEUROJSON_CACHE | if defined, this path will be searched first
% [pwd '/.neurojson'] | on all OSes
% /home/USERNAME/.neurojson | on all OSes (per-user)
% /home/USERNAME/.cache/neurojson | if on Linux (per-user)
% /var/cache/neurojson | if on Linux (system wide)
% /home/USERNAME/Library/neurojson| if on MacOS (per-user)
% /Library/neurojson | if on MacOS (system wide)
% C:\ProgramData\neurojson | if on Windows (system wide)
% When saving a file from a URL, under the root cache folder, subfolders can be created;
% if the URL is one of a standard NeuroJSON.io URLs as below
% https://neurojson.org/io/stat.cgi?action=get&db=DBNAME&doc=DOCNAME&file=sub-01/anat/datafile.nii.gz
% https://neurojson.io:7777/DBNAME/DOCNAME
% https://neurojson.io:7777/DBNAME/DOCNAME/datafile.suffix
% the file datafile.nii.gz will be downloaded to /home/USERNAME/.neurojson/io/DBNAME/DOCNAME/sub-01/anat/ folder
% if a URL does not follow the neurojson.io format, the cache folder has the below form
% CACHEFOLDER{i}/domainname.com/XX/YY/XXYYZZZZ...
% where XXYYZZZZ.. is the SHA-256 hash of the full URL, XX is the first two digit, YY is the 3-4 digits
% below command searches CACHEFOLDER{i}/io/openneuro/ds000001/sub-01/anat/, and return the path/filename
[cachepath, filename] = jsoncache('https://neurojson.org/io/stat.cgi?action=get&db=openneuro&doc=ds000001&file=sub-01/anat/sub-01_inplaneT2.nii.gz&size=669578')
% this searches CACHEFOLDER{i}/raw.githubusercontent.com/55/d2, and the filename is 55d24a4bad6ecc3f5dc4d333be728e01c26b696ef7bc5dd0861b7fa672a28e8e.json
[cachepath, filename] = jsoncache('https://raw.githubusercontent.com/fangq/jsonlab/master/examples/example1.json')
% this searches cachefolder{i}/io/adhd200/Brown folder, and look for file Brown.json
[cachepath, filename] = jsoncache('https://neurojson.io:7777/adhd200/Brown')
% this searches cachefolder{i}/io/openneuro/ds003805 folder, and look for file ds003805.json
[cachepath, filename] = jsoncache('https://neurojson.io:7777/openneuro/ds003805')
% jdlink dynamically downloads, caches and parses data files from one or multiple URLs
% jdlink calls jsoncache to scan cache folders first, if a cache copy exists, it loads the cache first
% here we download a dataset from NeuroJSON.io, containing many linked data files
data = loadjson('https://neurojson.io:7777/openneuro/ds000001');
% we now use jsonpath to scan all linked resources under subfolder "anat"
alllinks = jsonpath(data, '$..anat.._DataLink_')
% let's download all linked nifti files (total 4) for sub-01 and sub-02, and load the files as niidata
niidata = jdlink(alllinks, 'regex', 'sub-0[12]_.*\.nii');
% if you just want to download/cache all files and do not want to parse the files, you can run
Under the examples folder, you can find several scripts to demonstrate the basic utilities of JSONLab. Running the demo_jsonlab_basic.m script, you will see the conversions from MATLAB data structure to JSON text and backward. In jsonlab_selftest.m, we load complex JSON files downloaded from the Internet and validate the loadjson/savejson functions for regression testing purposes. Similarly, a demo_ubjson_basic.m script is provided to test the saveubjson and loadubjson functions for various matlab data structures, and demo_msgpack_basic.m is for testing savemsgpack and loadmsgpack.
Please run these examples and understand how JSONLab works before you use it to process your data.unit testing
Under the test folder, you can find a script to test individual data types and inputs using various encoders and decoders. This unit testing script also serves as a specification validator to the JSONLab functions and ensure that the outputs are compliant to the underlying specifications.
In-memory data compression/decompression
JSONLab contains a set of functions to perform in-memory buffer data compression and decompressionData
Compression: {zlib,gzip,base64,lzma,lzip,lz4,lz4hc,zstd,blosc2}encode.m
% MATLAB running with jvm provides zlib and gzip compression natively
% one can also install ZMat (https://github.com/NeuroJSON/zmat) to do zlib(.zip) or gzip (.gz) compression
output = zlibencode(diag([1,2,3,4]))
[output, info] = zlibencode(uint8(magic(8)))
outputbase64 = char(base64encode(output(:)))
% char, numeric and logical ND-arrays are acceptable inputs to the compression functions
[output, info] = gzipencode(uint8(magic(8)))
% setting a negative integer between -1 to -9 to set compression level: -9 being the highest
[output, info] = zlibencode(uint8(magic(8)), -9)
% other advanced compressions are supported but requires ZMat
% lzma offers the highest compression rate, but slow compresison speed
output = lzmaencode(uint8(magic(8)))
% lz4 offers the fastest compression speed, but slightly low compression ratio
output = lz4encode(peaks(10))
output = lz4hcencode(uint8(magic(8)))
% zstd has a good balanced speed/ratio, similar to zlib
output = zstdencode(peaks(10))
output = zstdencode(peaks(10), -9)
Data Deompression: {zlib,gzip,base64,lzma,lzip,lz4,lz4hc,zstd,blosc2}decode.m
% passing on a compressed byte-array buffer to *decode function decompresses the buffer
[compressed, info] = zlibencode(eye(10));
% the decompressed buffer is a byte-array
decompressd = zlibdecode(compressed);
% to fully recover the original data structure, one most use the info struct returned by the compressor
decompressd = zlibdecode(compressed, info)
% if one passes a zlib compressed buffer to a different decompressor, an error is reported
decompressd = gzipdecode(compressed, info)
outputbase64 = char(base64decode(base64encode('jsonlab test')))
Using jsave/jload to share workspace
Starting from JSONLab v2.0, we provide a pair of functions, jsave/jload to store and retrieve variables from the current workspace, similar to the save/load functions in MATLAB and Octave. The files that jsave/jload reads/writes is by default a binary JData file with a suffix .pmat. The file size is comparable (can be smaller if use lzma compression) to .mat files. This feature is currently experimental.
The main benefits of using .pmat file to share matlab variables include
  • a .pmat file can be 50% smaller than a .mat file when using jsave(..., "compression","lzma"); the only drawback is longer saving time.
  • a .pmat file can be readily read/opened among many programming environments, including Python, JavaScript, Go, Java etc, where .mat file support is not generally available. Parsers of .pmat files are largely compatible with BJData's parsers available at https://neurojson.org/#software
  • a .pmat file is quasi-human-readable, one can see the internal data fields even in a command line, for example using strings -n 2 file.pmat | astyle, making the binary data easy to be understood, shared and reused.
  • jsave/jload can also use MessagePack and JSON formats as the underlying data storage format, addressing needs from a diverse set of applications. MessagePack parsers are readily available at https://msgpack.org/
jsave % save the current workspace to default.pmat
jsave mydata.pmat
jload % load variables from default.pmat to the current workspace
jload mydata.pmat % load variables from mydata.pmat
vars=jload('mydata.pmat','vars',{'var1','var2'}) % return vars.var1, vars.var2
Sharing JSONLab created data files in Python
Despite the use of portable data annotation defined by the JData Specification, the output JSON files created by JSONLab are 100% JSON compatible (with the exception that long strings may be broken into multiple lines for better readability). Therefore, JSONLab-created JSON files (.json, .jnii, .jnirs etc) can be readily read and written by nearly all existing JSON parsers, including the built-in json module parser in Python.
However, we strongly recommend one to use a lightweight jdata module, developed by the same author, to perform the extra JData encoding and decoding and convert JSON data directly to convenient Python/Numpy data structures. The jdata module can also directly read/write UBJSON/Binary JData outputs from JSONLab (.bjd, .ubj, .bnii, .bnirs, .pmat etc). Using binary JData files are expected to produce much smaller file sizes and faster parsing, while maintaining excellent portability and generality.
In short, to conveniently read/write data files created by JSONLab into Python, whether they are JSON based or binary JData/UBJSON based, one just need to download the below two light-weight python modules:
To install these modules on Python 2.x, please first check if your system has pip and numpy, if not, please install it by running (using Ubuntu/Debian as example)
sudo apt-get install python-pip python3-pip python-numpy python3-numpy
After the installation is done, one can then install the jdata and bjdata modules by
pip install jdata --user
pip install bjdata --user
To install these modules for Python 3.x, please replace pip by pip3. If one prefers to install these modules globally for all users, simply execute the above commands using
sudo pip install jdata
sudo pip install bjdata
The above modules require built-in Python modules json and NumPy (numpy).
Once the necessary modules are installed, one can type python (or python3), and run
import jdata as jd
import numpy as np
where jd.loadt() function loads a text-based JSON file, performs JData decoding and converts the enclosed data into Python dict, list and numpy objects. Similarly, jd.loadb() function loads a binary JData/UBJSON file and performs similar conversions. One can directly call jd.load() to open JSONLab (and derived toolboxes such as jnifti: https://github.com/NeuroJSON/jnifti or jsnirf: https://github.com/NeuroJSON/jsnirf) generated files based on their respective file suffix.
Similarly, the jd.savet(), jd.saveb() and jd.save functions can revert the direction and convert a Python/Numpy object into JData encoded data structure and store as text-, binary- and suffix-determined output files, respectively.
Known Issues and TODOs
JSONLab has several known limitations. We are striving to make it more general and robust. Hopefully in a few future releases, the limitations become less.
Here are the known issues:
  • 3D or higher dimensional cell/struct-arrays will be converted to 2D arrays
  • When processing names containing multi-byte characters, Octave and MATLAB can give different field-names; you can use feature('DefaultCharacterSet','latin1') in MATLAB to get consistentresults
  • savejson can only export the properties from MATLAB classes, but not the methods
  • saveubjson converts a logical array into a uint8 ([U]) array
  • a special N-D array format, as defined in the JData specification, is implemented in saveubjson. You may use saveubjson(...,'NestArray',1) to create UBJSON Draft-12 compliant files
  • loadubjson can not parse all UBJSON Specification (Draft 12) compliant files, however, it can parse all UBJSON files produced by saveubjson.
Contribution and feedback
JSONLab is an open-source project. This means you can not only use it and modify it as you wish, but also you can contribute your changes back to JSONLab so that everyone else can enjoy the improvement. For anyone who want to contribute, please download JSONLab source code from its source code repositories by using the following command:
git clone https://github.com/fangq/jsonlab.git jsonlab
or browsing the github site at
Please report any bugs or issues to the below URL:
Sometimes, you may find it is necessary to modify JSONLab to achieve your goals, or attempt to modify JSONLab functions to fix a bug that you have encountered. If you are happy with your changes and willing to share those changes to the upstream author, you are recommended to create a pull-request on github.
To create a pull-request, you first need to "fork" jsonlab on Github by clicking on the "fork" button on top-right of JSONLab's github page. Once you forked jsonlab to your own directory, you should then implement the changes in your own fork. After thoroughly testing it and you are confident the modification is complete and effective, you can then click on the "New pull request" button, and on the left, select fangq/jsonlab as the "base". Then type in the description of the changes. You are responsible to format the code updates using the same convention (tab-width: 8, indentation: 4 spaces) as the upstream code.
We appreciate any suggestions and feedbacks from you. Please use the following user forum to ask any question you may have regarding JSONLab:

Cite As

Qianqian Fang (2024). JSONLab: portable, robust JSON/binary-JSON encoder/decoder (https://www.mathworks.com/matlabcentral/fileexchange/33381-jsonlab-portable-robust-json-binary-json-encoder-decoder), MATLAB Central File Exchange. Retrieved .

MATLAB Release Compatibility
Created with R2008a
Compatible with any release
Platform Compatibility
Windows macOS Linux

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!
Version Published Release Notes

-support JSON-mmap for fast read/write of JSON files as disk-map
-support JSONPath
-jdlink/jsoncache download/cache links on-demand
-support Blosc2/ZStd array compression
-set `BuiltinJSON` to call jsonencode/decode
-built-in NIfTI/HDF5/CSV parsers


- fix bjdata spec link


- add unit testing script
- apply data compression to strings
- add jsave and jload for portable data sharing
- compatible with R2008
- add UseMap option output containers.Map
- supports BJData spec Draft 1
- add _ArrayShape_ to save special matrices

- update documentation, finalize 1.0
- fix a bug to strictly respect NoRowBracket in savejson

- fix UBSJON data endianness for all integer types
- fix several compatibility issues for GNU octave
- support 2D cell and struct arrays
- escape special characters in a JSON string
- fix a bug when saving UBSJON files

- update to jsonlab 1.0 beta release
- fix file read and write bugs in loadubjson and saveubjson

update to 1.0 alpha release, included functions (saveubjson and loadubjson) for writing and reading Universal Binary JSON files


- 2012/06/17 *new format for an invalid leading char, unpacking hex code in savejson
- 2012/06/01 support JSONP in savejson
- 2012/05/25 fix the empty cell bug
- 2012/04/05 savejson can save to a file

- fix a bug in saveubjson when saving integer 256

- fix openjdata.org link

- update savejson and saveubjson license statement

-Avoid double expanding arrayzipdata in savejson and saveubjson


-Official reference implementation for JData Spec Draft 2(http://github.com/fangq/jdata)
-Initial support to MessagePack
-ND array serialization changed from column-major->row-major
-Support LZMA/LZ4/LZHC compression
-Support graph,table,class&object



2012/02/28 loadjson quotation mark escape bug, see http://bit.ly/yyk1nS
2012/01/25 patch to handle root-less objects, contributed by Blake Johnson
- unescape strings in loadjson.m
- support zmat base64 encoding
- support data compression
- support function handles
- add option to parse string array or convert to char
- treat string array as cell array in newer matlab
- save 64bit int in ubjson

-2012/01/13 speed up loadjson by 20 fold when parsing large data arrays in matlab
-2012/01/11 remove row bracket if an array has 1 element
-2011/12/22 accept sequence of 'param',value input
-2011/11/18 fix struct array bug


-use Big-endian format to store floating points (d/D) in saveubjson
-speedup parsing large unstructured data
-make parsing independent of whitespace
-allow to parse array of homogeneous elements
-fix a bug that prevents saving to a file in savejson


2011/10/21 fix a bug in loadjson, previous code does not use any of the acceleration
2011/10/20 loadjson supports JSON collections: appendable JSON objects

- 2x speed in savejson (Issue#17)
- fix FileName case bug (SVN rev#495)
- add SingletCell/SingletArray flags (Issue#15,#8)
- fix file name bug (Pull#14)
- fix bug for cell with transposed data (Pull#12)
- export matlab object to JSON (Pull#10,#11)
- 2x speed in savejson (Issue#17)
- fix FileName case bug (SVN rev#495)
- add SingletCell/SingletArray flags (Issue#15,#8)
- fix file name bug (Pull#14)
- fix bug for cell with transposed data (Pull#12)
- export matlab object to JSON (Pull#10,#11)
- 2x speed in savejson (Issue#17)
- fix FileName case bug (SVN rev#495)
- add SingletCell/SingletArray flags (Issue#15,#8)
- fix file name bug (Pull#14)
- fix bug for cell with transposed data (Pull#12)
- export matlab object to JSON (Pull#10,#11)
- 2x speed in savejson (Issue#17)
- fix FileName case bug (SVN rev#495)
- add SingletCell/SingletArray flags (Issue#15,#8)
- fix file name bug (Pull#14)
- fix bug for cell with transposed data (Pull#12)
- export matlab object to JSON (Pull#10,#11)

-massively accelerating loadjson
-parse logicals in loadjson
-make options case insensitive
-reading unicode encoded json files
-allow \uXXXX in a unicode in a string
-handle escape characters in a string
-implement the UBJSON Draft12 new name format