Datasets are very similar to NumPy arrays. They are homogeneous collections of data elements, with an immutable datatype and (hyper)rectangular shape. Unlike NumPy arrays, they support a variety of transparent storage features such as compression, error-detection, and chunked I/O.
They are represented in h5py by a thin proxy class which supports familiar NumPy operations like slicing, along with a variety of descriptive attributes:
- shape attribute
- size attribute
- dtype attribute
To make an empty dataset, all you have to do is specify a name, shape, and
optionally the data type (defaults to
>>> dset = f.create_dataset("default", (100,)) >>> dset = f.create_dataset("ints", (100,), dtype='i8')
You may initialize the dataset to an existing NumPy array:
>>> arr = np.arange(100) >>> dset = f.create_dataset("init", data=arr)
dtype may be specified along with
data; if so,
they will override
data.dtype. It’s required that
(1) the total number of points in
shape match the total number of points
data.shape, and that (2) it’s possible to cast
An HDF5 dataset created with the default settings will be contiguous; in other words, laid out on disk in traditional C order. Datasets may also be created using HDF5’s chunked storage layout. This means the dataset is divided up into regularly-sized pieces which are stored haphazardly on disk, and indexed using a B-tree.
Chunked storage makes it possible to resize datasets, and because the data is stored in fixed-size chunks, to use compression filters.
To enable chunked storage, set the keyword
chunks to a tuple indicating
the chunk shape:
>>> dset = f.create_dataset("chunked", (1000, 1000), chunks=(100, 100))
Data will be read and written in blocks with shape (100,100); for example,
the data in
dset[0:100,0:100] will be stored together in the file, as will
the data points in range
Chunking has performance implications. It’s recommended to keep the total size of your chunks between 10 KiB and 1 MiB, larger for larger datasets. Also keep in mind that when any element in a chunk is accessed, the entire chunk is read from disk.
Since picking a chunk shape can be confusing, you can have h5py guess a chunk shape for you:
>>> dset = f.create_dataset("autochunk", (1000, 1000), chunks=True)
Auto-chunking is also enabled when using compression or
if a chunk shape is not manually specified.
In HDF5, datasets can be resized once created up to a maximum size,
Dataset.resize(). You specify this maximum size when creating
the dataset, via the keyword
>>> dset = f.create_dataset("resizable", (10,10), maxshape=(500, 20))
Any (or all) axes may also be marked as “unlimited”, in which case they may
be increased up to the HDF5 per-axis limit of 2**64 elements. Indicate these
>>> dset = f.create_dataset("unlimited", (10, 10), maxshape=(None, 10))
Resizing an array with existing data works differently than in NumPy; if any axis shrinks, the data in the missing region is discarded. Data does not “rearrange” itself as it does when resizing a NumPy array.
Chunked data may be transformed by the HDF5 filter pipeline. The most common use is applying transparent compression. Data is compressed on the way to disk, and automatically decompressed when read. Once the dataset is created with a particular compression filter applied, data may be read and written as normal with no special steps required.
Enable compression with the
compression keyword to
>>> dset = f.create_dataset("zipped", (100, 100), compression="gzip")
Options for each filter may be specified with
>>> dset = f.create_dataset("zipped_max", (100, 100), compression="gzip", compression_opts=9)
Lossless compression filters¶
- GZIP filter (
- Available with every installation of HDF5, so it’s best where portability is
required. Good compression, moderate speed.
compression_optssets the compression level and may be an integer from 0 to 9, default is 4.
- LZF filter (
- Available with every installation of h5py (C source code also available). Low to moderate compression, very fast. No options.
- SZIP filter (
- Patent-encumbered filter used in the NASA community. Not available with all installations of HDF5 due to legal reasons. Consult the HDF5 docs for filter options.
Custom compression filters¶
In addition to the compression filters listed above, compression filters can be
dynamically loaded by the underlying HDF5 library. This is done by passing a
filter number to
Group.create_dataset() as the
compression_opts parameter will then be passed to this filter.
The underlying implementation of the compression filter will have the
H5Z_FLAG_OPTIONAL flag set. This indicates that if the compression
filter doesn’t compress a block while writing, no error will be thrown. The
filter will then be skipped when subsequently reading the block.
Filters enabled with the
compression keywords are _lossless_; what comes
out of the dataset is exactly what you put in. HDF5 also includes a lossy
filter which trades precision for storage space.
Works with integer and floating-point data only. Enable the scale-offset
filter by setting
scaleoffset to an
For integer data, this specifies the number of bits to retain. Set to 0 to have HDF5 automatically compute the number of bits required for lossless compression of the chunk. For floating-point data, indicates the number of digits after the decimal point to retain.
Currently the scale-offset filter does not preserve special float values (i.e. NaN, inf), see https://lists.hdfgroup.org/pipermail/hdf-forum_lists.hdfgroup.org/2015-January/008296.html for more information and follow-up.
Block-oriented compressors like GZIP or LZF work better when presented with runs of similar values. Enabling the shuffle filter rearranges the bytes in the chunk and may improve compression ratio. No significant speed penalty, lossless.
Enable by setting
shuffle to True.
Reading & writing data¶
HDF5 datasets re-use the NumPy slicing syntax to read and write to the file. Slice specifications are translated directly to HDF5 “hyperslab” selections, and are a fast and efficient way to access data in the file. The following slicing arguments are recognized:
- Indices: anything that can be converted to a Python long
- Slices (i.e.
- Field names, in the case of compound data
- At most one
Here are a few examples (output omitted)
>>> dset = f.create_dataset("MyDataset", (10,10,10), 'f') >>> dset[0,0,0] >>> dset[0,2:10,1:9:3] >>> dset[:,::2,5] >>> dset >>> dset[1,5] >>> dset[0,...] >>> dset[...,6]
For compound data, you can specify multiple field names alongside the numeric slices:
>>> dset["FieldA"] >>> dset[0,:,4:5, "FieldA", "FieldB"] >>> dset[0, ..., "FieldC"]
To retrieve the contents of a scalar dataset, you can use the same
syntax as in NumPy:
result = dset[()]. In other words, index into
the dataset using an empty tuple.
For simple slicing, broadcasting is supported:
>>> dset[0,:,:] = np.arange(10) # Broadcasts to (10,10)
Broadcasting is implemented using repeated hyperslab selections, and is safe to use with very large target selections. It is supported for the above “simple” (integer, slice and ellipsis) slicing only.
A subset of the NumPy fancy-indexing syntax is supported. Use this with caution, as the underlying HDF5 mechanisms may have different performance than you expect.
For any axis, you can provide an explicit list of points you want; for a dataset with shape (10, 10):
>>> dset.shape (10, 10) >>> result = dset[0, [1,3,8]] >>> result.shape (3,) >>> result = dset[1:6, [5,8,9]] >>> result.shape (5, 3)
The following restrictions exist:
- List selections may not be empty
- Selection coordinates must be given in increasing order
- Duplicate selections are ignored
- Very long lists (> 1000 elements) may produce poor performance
NumPy boolean “mask” arrays can also be used to specify a selection. The result of this operation is a 1-D array with elements arranged in the standard NumPy (C-style) order. Behind the scenes, this generates a laundry list of points to select, so be careful when using it with large masks:
>>> arr = numpy.arange(100).reshape((10,10)) >>> dset = f.create_dataset("MyDataset", data=arr) >>> result = dset[arr > 50] >>> result.shape (49,)
Length and iteration¶
As with NumPy arrays, the
len() of a dataset is the length of the first
axis, and iterating over a dataset iterates over the first axis. However,
modifications to the yielded data are not recorded in the file. Resizing a
dataset while iterating has undefined results.
On 32-bit platforms,
len(dataset) will fail if the first axis is bigger
than 2**32. It’s recommended to use
Dataset.len() for large datasets.
Creating and Reading Empty (or Null) datasets and attributes¶
HDF5 has the concept of Empty or Null datasets and attributes. These are not
the same as an array with a shape of (), or a scalar dataspace in HDF5 terms.
Instead, it is a dataset with an associated type, no data, and no shape. In
h5py, we represent this as either a dataset with shape
None, or an
h5py.Empty. Empty datasets and attributes cannot be sliced.
To create an empty attribute, use
h5py.Empty as per Attributes:
>>> obj.attrs["EmptyAttr"] = h5py.Empty("f")
Similarly, reading an empty attribute returns
>>> obj.attrs["EmptyAttr"] h5py.Empty(dtype="f")
Empty datasets can be created either by defining a
dtype but no
>>> grp.create_dataset("EmptyDataset", dtype="f")
data to an instance of
>>> grp.create_dataset("EmptyDataset", data=h5py.Empty("f"))
An empty dataset has shape defined as
None, which is the best way of
determining whether a dataset is empty or not. An empty dataset can be “read” in
a similar way to scalar datasets, i.e. if
empty_dataset is an empty
>>> empty_dataset[()] h5py.Empty(dtype="f")
The dtype of the dataset can be accessed via
<dset>.dtype as per normal.
As empty datasets cannot be sliced, some methods of datasets such as
read_direct will raise an exception if used on a empty dataset.
Dataset objects are typically created via
Group.create_dataset(), or by retrieving existing datasets from a file. Call this constructor to create a new Dataset bound to an existing
read_direct(array, source_sel=None, dest_sel=None)¶
Read from an HDF5 dataset directly into a NumPy array, which can avoid making an intermediate copy as happens with slicing. The destination array must be C-contiguous and writable, and must have a datatype to which the source data may be cast. Data type conversion will be carried out on the fly by HDF5.
source_sel and dest_sel indicate the range of points in the dataset and destination array respectively. Use the output of
>>> dset = f.create_dataset("dset", (100,), dtype='int64') >>> arr = np.zeros((100,), dtype='int32') >>> dset.read_direct(arr, np.s_[0:10], np.s_[50:60])
write_direct(source, source_sel=None, dest_sel=None)¶
Write data directly to HDF5 from a NumPy array. The source array must be C-contiguous. Selections must be the output of numpy.s_[<args>]. Broadcasting is supported for simple indexing.
Return a context manager allowing you to read data as a particular type. Conversion is handled by HDF5 directly, on the fly:
>>> dset = f.create_dataset("bigint", (1000,), dtype='int64') >>> with dset.astype('int16'): ... out = dset[:] >>> out.dtype dtype('int16')
Change the shape of a dataset. size may be a tuple giving the new dataset shape, or an integer giving the new length of the specified axis.
Datasets may be resized only up to
Return the size of the first axis.
NumPy-style shape tuple giving dataset dimensions.
NumPy dtype object giving the dataset’s type.
Integer giving the total number of elements in the dataset.
NumPy-style shape tuple indicating the maxiumum dimensions up to which the dataset may be resized. Axes with
String with the currently applied compression filter, or None if compression is not enabled for this dataset. See Filter pipeline.
Setting for the HDF5 scale-offset filter (integer), or None if scale-offset compression is not used for this dataset. See Scale-Offset filter.
Value used when reading uninitialized portions of the dataset, or None if no fill value has been defined, in which case HDF5 will use a type-appropriate default value. Can’t be changed after the dataset is created.
String giving the full path to this dataset.