HDF5.jl

Overview

HDF5 stands for Hierarchical Data Format v5 and is closely modeled on file systems. In HDF5, a "group" is analogous to a directory, a "dataset" is like a file. HDF5 also uses "attributes" to associate metadata with a particular group or dataset. HDF5 uses ASCII names for these different objects, and objects can be accessed by Unix-like pathnames, e.g., "/sample1/tempsensor/firsttrial" for a top-level group "sample1", a subgroup "tempsensor", and a dataset "firsttrial".

For simple types (scalars, strings, and arrays), HDF5 provides sufficient metadata to know how each item is to be interpreted. For example, HDF5 encodes that a given block of bytes is to be interpreted as an array of Int64, and represents them in a way that is compatible across different computing architectures.

However, to preserve Julia objects, one generally needs additional type information to be supplied, which is easy to provide using attributes. This is handled for you automatically in the JLD/JLD2. These specific formats (conventions) provide "extra" functionality, but they are still both regular HDF5 files and are therefore compatible with any HDF5 reader or writer.

Language wrappers for HDF5 are often described as either "low level" or "high level." This package contains both flavors: at the low level, it directly wraps HDF5's functions, thus copying their API and making them available from within Julia. At the high level, it provides a set of functions built on the low-level wrap which may make the usage of this library more convenient.

Installation

julia>]
pkg> add HDF5

Starting from Julia 1.3, the HDF5 binaries are by default downloaded using the HDF5_jll package.

Using custom or system provided HDF5 binaries

To use system-provided HDF5 binaries instead, set the preferences libhdf5 and libhdf5_hl, see also Preferences.jl. These need to point to the local paths of the libraries libhdf5 and libhdf5_hl.

For example, to use HDF5 (libhdf5-mpich-dev) with MPI using system libraries on Ubuntu 20.04, you would run

$ sudo apt install mpich libhdf5-mpich-dev

If your system HDF5 library is compiled with MPI, you need to tell MPI.jl to use the same locally installed MPI implementation. This can be done in Julia by running:

using MPIPreferences
MPIPreferences.use_system_binary()

to set the MPI preferences, see the documentation of MPI.jl. You can set the path to the system library using Preferences.jl by:

using Preferences, HDF5

set_preferences!(
    HDF5,
    "libhdf5" => "/usr/lib/x86_64-linux-gnu/hdf5/mpich/libhdf5.so",
    "libhdf5_hl" => "/usr/lib/x86_64-linux-gnu/hdf5/mpich/libhdf5_hl.so", force = true)

Alternatively, HDF5.jl provides a convenience function HDF5.API.set_libraries! that can be used as follows:

using HDF5

HDF5.API.set_libraries!("/usr/lib/x86_64-linux-gnu/hdf5/mpich/libhdf5.so", "/usr/lib/x86_64-linux-gnu/hdf5/mpich/libhdf5_hl.so")

Going back to the default, i.e. deleting the preferences again, can be done by calling HDF5.API.set_libraries!(). If HDF5 cannot be loaded, it may be useful to use the UUID to change these settings:

using Preferences, UUIDs

set_preferences!(
    UUID("f67ccb44-e63f-5c2f-98bd-6dc0ccc4ba2f"), # UUID of HDF5.jl
    "libhdf5" => "/usr/lib/x86_64-linux-gnu/hdf5/mpich/libhdf5.so",
    "libhdf5_hl" => "/usr/lib/x86_64-linux-gnu/hdf5/mpich/libhdf5_hl.so", force = true)

Also see the file test/configure_packages.jl for an example.

Both, the MPI preferences and the preferences for HDF5.jl write to a file called LocalPreferences.toml in the project directory. After performing the described steps this file could look like the following:

[MPIPreferences]
_format = "1.0"
abi = "MPICH"
binary = "system"
libmpi = "/software/mpi/lib/libmpi.so"
mpiexec = "/software/mpi/bin/mpiexec"

[HDF5]
libhdf5 = "/usr/lib/x86_64-linux-gnu/hdf5/mpich/libhdf5.so"
libhdf5_hl = "/usr/lib/x86_64-linux-gnu/hdf5/mpich/libhdf5_hl.so"

If you want to switch to another HDF5 library or the library moved, you can call the set_preferences! commands again (or manually edit LocalPreferences.toml) to set the new paths. Using the default implementation provided by HDF5_jll can be done by simply manually deleting the LocalPreferences.toml file.

Opening and closing files

"Plain" (i.e., with no extra formatting conventions) HDF5 files are created and/or opened with the h5open command:

fid = h5open(filename, mode)

The mode can be any one of the following:

For example

julia> using HDF5
julia> fname = tempname(); # temporary file
julia> fid = h5open(fname, "w")🗂️ HDF5.File: (read-write) /tmp/jl_Duv8HqLgHI

This produces an object of type HDF5File, a subtype of the abstract type DataFile. This file will have no elements (groups, datasets, or attributes) that are not explicitly created by the user.

When you're finished with a file, you should close it:

close(fid)

Closing a file also closes any other open objects (e.g., datasets, groups) in that file. In general, you need to close an HDF5 file to "release" it for use by other applications.

Creating a group or dataset

Groups can be created via the function create_group

julia> create_group(fid, "mygroup")📂 HDF5.Group: /mygroup (file: /tmp/jl_Duv8HqLgHI)

We can write the "mydataset" by indexing into fid. This also happens to write data to the dataset.

julia> fid["mydataset"] = rand()0.919100834621411

Alternatively, we can call create_dataset, which does not write data to the dataset. It merely creates the dataset.

julia> create_dataset(fid, "myvector", Int, (10,))🔢 HDF5.Dataset: /myvector (file: /tmp/jl_Duv8HqLgHI xfer_mode: 0)

Creating a dataset within a group is as simple as indexing into the group with the name of the dataset or calling create_dataset with the group as the first argument.

julia> g = fid["mygroup"]📂 HDF5.Group: /mygroup (file: /tmp/jl_Duv8HqLgHI)
julia> g["mydataset"] = "Hello World!""Hello World!"
julia> create_dataset(g, "myvector", Int, (10,))🔢 HDF5.Dataset: /mygroup/myvector (file: /tmp/jl_Duv8HqLgHI xfer_mode: 0)

The do syntax is also supported. The file, group, and dataset handles will automatically be closed after the do block terminates.

julia> h5open("example2.h5", "w") do fid
           g = create_group(fid, "mygroup")
           dset = create_dataset(g, "myvector", Float64, (10,))
           write(dset,rand(10))
       end

Opening and closing objects

If you have a file object fid, and this has a group or dataset called "mygroup" at the top level of a file, you can open it in the following way:

julia> obj = fid["mygroup"]📂 HDF5.Group: /mygroup (file: /tmp/jl_Duv8HqLgHI)
├─ 🔢 mydataset
└─ 🔢 myvector

This does not read any data or attributes associated with the object, it's simply a handle for further manipulations. For example:

julia> g = fid["mygroup"]📂 HDF5.Group: /mygroup (file: /tmp/jl_Duv8HqLgHI)
├─ 🔢 mydataset
└─ 🔢 myvector
julia> dset = g["mydataset"]🔢 HDF5.Dataset: /mygroup/mydataset (file: /tmp/jl_Duv8HqLgHI xfer_mode: 0)

or simply

julia> dset = fid["mygroup/mydataset"]🔢 HDF5.Dataset: /mygroup/mydataset (file: /tmp/jl_Duv8HqLgHI xfer_mode: 0)

When you're done with an object, you can close it using close(obj). If you forget to do this, it will be closed for you anyway when the file is closed, or if obj goes out of scope and gets garbage collected.

Reading and writing data

Suppose you have a group g which contains a dataset with path "mydataset", and that you've also opened this dataset as dset = g["mydataset"]. You can read information in this dataset in any of the following ways:

A = read(dset)
A = read(g, "mydataset")
Asub = dset[2:3, 1:3]

The last syntax reads just a subset of the data array (assuming that dset is an array of sufficient size). libhdf5 has internal mechanisms for slicing arrays, and consequently if you need only a small piece of a large array, it can be faster to read just what you need rather than reading the entire array and discarding most of it.

Datasets can be created with either

g["mydataset"] = rand(3,5)
write(g, "mydataset", rand(3,5))

One can use the high level interface load and save from FileIO, where an optional OrderedDict can be passed (track_order inferred). Note that using track_order=true or passing an OrderedDict is a promise that the read file has been created with the appropriate ordering flags.

julia> using OrderedCollections, FileIO
julia> save("track_order.h5", OrderedDict("z"=>1, "a"=>2, "g/f"=>3, "g/b"=>4))
julia> load("track_order.h5"; dict=OrderedDict())
OrderedDict{Any, Any} with 4 entries:
  "z"   => 1
  "a"   => 2
  "g/f" => 3
  "g/b" => 4

Passing parameters

It is often required to pass parameters to specific routines, which are collected in so-called property lists in HDF5. There are different property lists for different tasks, e.g. for the access/creation of files, datasets, groups. In this high level framework multiple parameters can be simply applied by appending them at the end of function calls as keyword arguments.

g["A"] = A  # basic
g["A", chunk=(5,5)] = A # add chunks

B = h5read(fn,"mygroup/B", # two parameters
  fapl_mpio=(ccomm,cinfo), # if parameter requires multiple args use tuples
  dxpl_mpio=HDF5.H5FD_MPIO_COLLECTIVE )

This will automatically create the correct property lists, add the properties, and apply the property list while reading/writing the data. The naming of the properties generally follows that of HDF5, i.e. the key fapl_mpio returns the HDF5 functions h5pget/set_fapl_mpio and their corresponding property list type H5P_FILE_ACCESS. The complete list if routines and their interfaces is available at the H5P: Property List Interface documentation. Note that not all properties are available. When searching for a property check whether the corresponding h5pget/set functions are available.

Chunking and compression

You can also optionally "chunk" and/or compress your data. For example,

A = rand(100,100)
g["A", chunk=(5,5)] = A

stores the matrix A in 5-by-5 chunks. Chunking improves efficiency if you write or extract small segments or slices of an array, if these are not stored contiguously.

A = rand(100,100)
g1["A", chunk=(5,5), compress=3] = A
g2["A", chunk=(5,5), shuffle=(), deflate=3] = A
using H5Zblosc # load in Blosc
g3["A", chunk=(5,5), blosc=3] = A

Standard compression in HDF5 ("compress") corresponds to ("deflate") and uses the deflate/zlib algorithm. The deflate algorithm is often more efficient if prefixed by a "shuffle" filter. Blosc is generally much faster than deflate – however, reading Blosc-compressed HDF5 files require Blosc to be installed. This is the case for Julia, but often not for vanilla HDF5 distributions that may be used outside Julia. (In this case, the structure of the HDF5 file is still accessible, but compressed datasets cannot be read.) Compression requires chunking, and heuristic chunking is automatically used if you specify compression but don't specify chunking.

It is also possible to write to subsets of an on-disk HDF5 dataset. This is useful to incrementally save to very large datasets you don't want to keep in memory. For example,

dset = create_dataset(g, "B", datatype(Float64), dataspace(1000,100,10), chunk=(100,100,1))
dset[:,1,1] = rand(1000)

creates a Float64 dataset in the file or group g, with dimensions 1000x100x10, and then writes to just the first 1000 element slice. If you know the typical size of subset reasons you'll be reading/writing, it can be beneficial to set the chunk dimensions appropriately.

For fine-grained control of filter and compression pipelines, please use the filters keyword to define a filter pipeline. For example, this can be used to include external filter packages. This enables the use of Blosc, Bzip2, LZ4, ZStandard, or custom filter plugins.

Memory mapping

If you will frequently be accessing individual elements or small regions of array datasets, it can be substantially more efficient to bypass HDF5 routines and use direct memory mapping. This is possible only under particular conditions: when the dataset is an array of standard "bits" types (e.g., Float64 or Int32) and no chunking/compression is being used. You can use the ismmappable function to test whether this is possible; for example,

dset = g["x"]
if HDF5.ismmappable(dset)
    dset = HDF5.readmmap(dset)
end
val = dset[15]

Note that readmmap returns an Array rather than an HDF5 object.

Note: if you use readmmap on a dataset and subsequently close the file, the array data are still available–-and file continues to be in use–-until all of the arrays are garbage-collected. This is in contrast to standard HDF5 datasets, where closing the file prevents further access to any of the datasets, but the file is also detached and can safely be rewritten immediately.

Under the default allocation-time policy, a newly added ismmappable dataset can only be memory mapped after it has been written to. The following fails:

vec_dset = create_dataset(g, "v", datatype(Float64), dataspace(10_000,1))
HDF5.ismmappable(vec_dset)    # == true
vec = HDF5.readmmap(vec_dset) # throws ErrorException("Error mmapping array")

because although the dataset description has been added, the space within the HDF5 file has not yet actually been allocated (so the file region cannot be memory mapped by the OS). The storage can be allocated by making at least one write:

vec_dset[1,1] = 0.0      # force allocation of /g/v within the file
vec = HDF5.readmmap(vec_dset) # and now the memory mapping can succeed

Alternatively, the policy can be set so that the space is allocated immediately upon creation of the data set with the alloc_time keyword:

mtx_dset = create_dataset(g, "M", datatype(Float64), dataspace(100, 1000),
                    alloc_time = HDF5.H5D_ALLOC_TIME_EARLY)
mtx = HDF5.readmmap(mtx_dset) # succeeds immediately

In-memory HDF5 files

It is possible to use HDF5 files without writing or reading from disk. This is useful when receiving or sending data over the network. Typically, when sending data, one might want to

1. Create a new file in memory. This can be achieved by passing Drivers.Core(; backing_store=false) to h5open(...)
2. Add data to the HDF5.File object
3. Get a representation of the file as a byte vector. This can be achieved by calling Vector{UInt8}(...) on the file object.

This is illustrated on the example below

using HDF5

# Creates a new file object without storing to disk by setting `backing_store=false`
file_as_bytes = h5open("AnyName_InMemory", "w"; driver=Drivers.Core(; backing_store=false)) do fid
   fid["MyData"] = randn(5, 5) # add some data
   return Vector{UInt8}(fid) # get a byte vector to send, e.g., using HTTP, MQTT or similar.
end

The same way, when receiving data as a vector of bytes that represent a HDF5 file, one can use h5open(...) with the byte vector as first argument to get a file object. Creating a file object from a byte vector will also by default open the file in memory, without saving a copy on disk.

using HDF5

...
h5open(file_as_bytes, "r"; name = "in_memory.h5") do fid
... # Do things with the data
end
...

Supported data types

HDF5.jl knows how to store values of the following types: signed and unsigned integers of 8, 16, 32, and 64 bits, Float32, Float64; Complex versions of these numeric types; Arrays of these numeric types (including complex versions); String; and Arrays of String. Arrays of strings are supported using HDF5's variable-length-strings facility. By default Complex numbers are stored as compound types with r and i fields following the h5py convention. When reading data, compound types with matching field names will be loaded as the corresponding Complex Julia type. These field names are configurable with the HDF5.set_complex_field_names(real::AbstractString, imag::AbstractString) function and complex support can be completely enabled/disabled with HDF5.enable/disable_complex_support().

As of HDF5.jl version 0.16.13, support was added to map Julia structs to compound HDF5 datatypes.

julia> struct Point3{T}
           x::T
           y::T
           z::T
       end

julia> datatype(Point3{Float64})
HDF5.Datatype: H5T_COMPOUND {
      H5T_IEEE_F64LE "x" : 0;
      H5T_IEEE_F64LE "y" : 8;
      H5T_IEEE_F64LE "z" : 16;
   }

For Arrays, note that the array dimensionality is preserved, including 0-length dimensions:

fid["zero_vector"] = zeros(0)
fid["zero_matrix"] = zeros(0, 0)
size(fid["zero_vector"]) # == (0,)
size(fid["zero_matrix"]) # == (0, 0)

An exception to this rule is Julia's 0-dimensional Array, which is stored as an HDF5 scalar because there is a value to be preserved:

fid["zero_dim_value"] = fill(1.0π)
read(fid["zero_dim_value"]) # == 3.141592653589793, != [3.141592653589793]

HDF5 also has the concept of a null array which contains a type but has neither size nor contents, which is represented by the type HDF5.EmptyArray:

fid["empty_array"] = HDF5.EmptyArray{Float32}()
HDF5.isnull(fid["empty_array"]) # == true
size(fid["empty_array"]) # == ()
eltype(fid["empty_array"]) # == Float32

This module also supports HDF5's VLEN, OPAQUE, and REFERENCE types, which can be used to encode more complex types. In general, you need to specify how you want to combine these more advanced facilities to represent more complex data types. For many of the data types in Julia, the JLD module implements support. You can likewise define your own file format if, for example, you need to interact with some external program that has explicit formatting requirements.

Creating groups and attributes

Create a new group in the following way:

g = create_group(parent, name)

The named group will be created as a child of the parent.

Attributes can be created using

attributes(parent)[name] = value

where attributes simply indicates that the object referenced by name (a string) is an attribute, not another group or dataset. (Datasets cannot have child datasets, but groups can have either.) value must be a simple type: BitsKinds, strings, and arrays of either of these. The HDF5 standard recommends against storing large objects as attributes.

The value stored in an attribute can be retrieved like

read_attribute(parent, name)

You can also access the value of an attribute by indexing, like so:

julia> attr = attributes(parent)[name];
julia> attr[]

Getting information

HDF5.name(obj)

will return the full HDF5 pathname of object obj.

keys(g)

returns a string array containing all objects inside group g. These relative pathnames, not absolute pathnames.

You can iterate over the objects in a group, i.e.,

for obj in g
  data = read(obj)
  println(data)
end

This gives you a straightforward way of recursively exploring an entire HDF5 file.

If you need to know whether group g has a dataset named mydata, you can test that with

if haskey(g, "mydata")
   ...
end
tf = haskey(g, "mydata")

If instead you want to know whether g has an attribute named myattribute, do it this way:

tf = haskey(attributes(g), "myattribute")

If you have an HDF5 object, and you want to know where it fits in the hierarchy of the file, the following can be useful:

p = parent(obj)     # p is the parent object (usually a group)
fn = HDF5.filename(obj)  # fn is a string
g = HDF5.root(obj)       # g is the group "/"

For array objects (datasets and attributes) the following methods work:

dims = size(dset)
nd = ndims(dset)
len = length(dset)

Objects can be created with properties, and you can query those properties in the following way:

p = HDF5.get_create_properties(dset)
chunksz = HDF5.get_chunk(p)

The simpler syntax chunksz = HDF5.get_chunk(dset) is also available.

Finally, sometimes you need to be able to conveniently test whether a file is an HDF5 file:

tf = HDF5.ishdf5(filename)

Mid-level routines

Sometimes you might want more fine-grained control, which can be achieved using a different set of routines. For example,

g = open_group(parent, name)
dset = open_dataset(parent, name[, apl])
attr = open_attribute(parent, name)
t = open_datatype(parent, name)

These open the named group, dataset, attribute, and committed datatype, respectively. For datasets, apl stands for "access parameter list" and provides opportunities for more sophisticated control (see the HDF5 documentation).

New objects can be created in the following ways:

g = create_group(parent, name[, lcpl, gcpl]; properties...)
dset = create_dataset(parent, name, data; properties...)
attr = create_attribute(parent, name, data)

creates groups, datasets, and attributes without writing any data to them. You can then use write(obj, data) to store the data. The optional properties and property lists allow even more fine-grained control. This syntax uses data to infer the object's "HDF5.datatype" and "HDF5.dataspace"; for the most explicit control, data can be replaced with dtype, dspace, where dtype is an HDF5.Datatype and dspace is an HDF5.Dataspace.

Analogously, to create committed data types, use

t = commit_datatype(parent, name, dtype[, lcpl, tcpl, tapl])

You can create and write data in one step,

write_dataset(parent, name, data; properties...)
write_attribute(parent, name, data)

You can use extendible dimensions,

d = create_dataset(parent, name, dtype, (dims, max_dims), chunk=(chunk_dims))
HDF5.set_extent_dims(d, new_dims)

where dims is a tuple of integers. For example

b = create_dataset(fid, "b", Int, ((1000,),(-1,)), chunk=(100,)) #-1 is equivalent to typemax(hsize_t)
HDF5.set_extent_dims(b, (10000,))
b[1:10000] = collect(1:10000)

when dimensions are reduced, the truncated data is lost. A maximum dimension of -1 is often referred to as unlimited dimensions, though it is limited by the maximum size of an unsigned integer.

You can copy data from one file to another:

copy_object(source, data_name, target, name)
copy_object(source[data_name], target, name)

Finally, it's possible to delete objects:

delete_object(parent, name)   # for groups, datasets, and datatypes
delete_attribute(parent, name)   # for attributes

Low-level routines

Many of the most commonly-used libhdf5 functions have been wrapped in a submodule API. The library follows a consistent convention: for example, libhdf5's H5Adelete is wrapped with a Julia function called h5a_delete. The arguments are exactly as specified in the HDF5 reference manual. Note that the functions in the API submodule are not exported, so unless you import them specifically, you need to preface them with HDF5.API to use them: for example, HDF5.API.h5a_delete.

HDF5 is a large library, and the low-level wrap is not complete. However, many of the most-commonly used functions are wrapped, and in general wrapping a new function takes only a single line of code. Users who need additional functionality are encouraged to contribute it.

Note that Julia's HDF5 directly uses the "2" interfaces, e.g., H5Dcreate2, so you need to have version 1.8 of the HDF5 library or later.

Language interoperability with row- and column-major order arrays

There are two main methods for storing multidimensional arrays in linear storage row-major order and column-major order. Julia, like Fortran and MATLAB, stores multidimensional arrays in column-major order, while other languages, including C and Python (NumPy), use row-major order. Therefore when reading an array in Julia from row-major order language the dimensions may be inverted.

To read a multidimensional array into the original shape from an HDF5 file written by Python (numpy and h5py) or C/C++/Objective-C, simply reverse the dimensions. For example, one may add the following line after reading the dataset dset:

dset = permutedims(dset, reverse(1:ndims(dset)))

Note that some languages or libraries use both methods, so please check the datset's description for details. For example, NumPy arrays are row-major by default, but NumPy can use either row-major or column-major ordered arrays.

Credits

• Konrad Hinsen initiated Julia's support for HDF5

• Tim Holy and Simon Kornblith (primary authors)

• Tom Short contributed code and ideas to the dictionary-like interface

• Blake Johnson made several improvements, such as support for iterating over attributes

• Isaiah Norton and Elliot Saba improved installation on Windows and OSX

• Steve Johnson contributed the do syntax and Blosc compression

• Mike Nolta and Jameson Nash contributed code or suggestions for improving the handling of HDF5's constants

• Thanks also to the users who have reported bugs and tested fixes