Skip to content


Subversion checkout URL

You can clone with HTTPS or Subversion.

Download ZIP
Fetching contributors…

Cannot retrieve contributors at this time

469 lines (358 sloc) 17.078 kb


This section gives an overview of the operations for storing and retrieving the basic data structures in |project|, such as `NumPy`_ arrays. |project| uses `HDF5`_ format for storing binary coded data. Using the |project| support for `HDF5`_, it is very simple to import and export data.

`HDF5`_ uses a neat descriptive language for representing the data in the HDF5 files, called Data Description Language (DDL).

To perform the functionalities given in this section, you should have `NumPy`_ and |project| loaded into the `Python`_ environment.

HDF5 standard utilities

Before explaining the basics of reading and writing to `HDF5`_ files, it is important to list some `HDF5`_ standard utilities for checking the content of an `HDF5`_ file. These are supplied by the `HDF5`_ project.

Dumps the content of the file using the DDL.
Lists the content of the file using DDL, but does not show the data.
Finds the differences between HDF5 files.

I/O operations using the class

Writing operations

Let's take a look at how to write simple scalar data such as integers or floats.

If after this you use the h5dump utility on the file testfile1.hdf5, you will verify that the file now contains:

HDF5 "testfile1.hdf5" {
GROUP "/" {
  DATASET "my_float" {
     DATASPACE  SIMPLE { ( 1 ) / ( 1 ) }
     DATA {
     (0): 3.1416
  DATASET "my_integer" {
     DATASPACE  SIMPLE { ( 1 ) / ( 1 ) }
     DATA {
     (0): 5


In |project|, when you open a HDF5 file, you can choose one of the following options:

'r' Open the file in reading mode; writing operations will fail (this is the default).

'a' Open the file in reading and writing mode with appending.

'w' Open the file in reading and writing mode, but truncate it.

'x' Read/write/append with exclusive access.

The dump shows that there are two datasets inside a group named / in the file. HDF5 groups are like file system directories. They create namespaces for the data. In the root group (or directory), you will find the two variables, named as you set them to be. The variable names are the complete path to the location where they live. You could write a new variable in the same file but in a different directory like this:

Line 1 opens the file for reading and writing, but without truncating it. This will allow you to access the file contents. Next, the directory /test is created and a new variable is written inside the subdirectory. As you can verify, for simple scalars, you can also force the storage type. Where normally one would have a 64-bit real value, you can impose that this variable is saved as a 32-bit real value. You can verify the dump correctness with h5dump:

GROUP "/" {
 GROUP "test" {
    DATASET "my_float" {
       DATASPACE  SIMPLE { ( 1 ) / ( 1 ) }
       DATA {
       (0): 6.28

Notice the subdirectory test has been created and inside it a floating point number has been stored. Such a float point number has a 32-bit precision as it was defined.


If you need to place lots of variables in a subfolder, it may be better to setup the prefix folder before starting the writing operations on the :py:class:`` object. You can do this using the method :py:meth:``. Look up its help for more information and usage instructions.

Writing arrays is a little simpler as the :py:class:`numpy.ndarray` objects encode all the type information we need to write and read them correctly. Here is an example:

The result of running h5dump on the file testfile3.hdf5 should be:

 DATASET "my_array" {
    DATASPACE  SIMPLE { ( 2, 2 ) / ( 2, 2 ) }
    DATA {
    (0,0): 0, 1,
    (1,0): 2, 3

You don't need to limit yourself to single variables, you can also save lists of scalars and arrays using the function :py:meth:`` instead of :py:meth:``.

Reading operations

Reading data from a file that you just wrote to is just as easy. For this task you should use :py:meth:``. The read method will read all the contents of the variable pointed to by the given path. This is the normal way to read a variable you have written with :py:meth:``. If you decided to create a list of scalar or arrays, the way to read that up would be using :py:meth:`` instead. Here is an example:

Now let's look at an example where we have used :py:meth:`` instead of :py:meth:`` to write data to a file. That is normally the case when you write lists of variables to a dataset.

This is what the h5dump of the file would look like:

HDF5 "testfile4.hdf5" {
GROUP "/" {
   DATASET "arrayset" {
      DATASPACE  SIMPLE { ( 3, 10 ) / ( H5S_UNLIMITED, 10 ) }
      DATA {
      (0,0): 0, 1, 2, 3, 4, 5, 6, 7, 8, 9,
      (1,0): 0, 2, 4, 6, 8, 10, 12, 14, 16, 18,
      (2,0): 0, 3, 6, 9, 12, 15, 18, 21, 24, 27

Notice that the expansion limits for the first dimension have been correctly set by |project| so you can insert an unlimited number of 1D float vectors. Of course, you can also read the whole contents of the arrayset in a single shot:

As you can see, the only difference between :py:meth:`` and :py:meth:`` is on how |project| considers the available data (as a single array with N dimensions or as a set of arrays with N-1 dimensions). In the first example, you would have also been able to read the variable my_array as an arrayset using :py:meth:`` instead of :py:meth:``. In this case, each position readout would return a 1D uint8 array instead of a 2D array.

Array interfaces

What we have shown so far is the generic API to read and write data using HDF5. You will use it when you want to import or export data from |project| into other software frameworks, debug your data or just implement your own classes that can serialize and de-serialize from HDF5 file containers. In |project|, most of the time you will be working with :py:class:``s and :py:class:``s and it is even simpler to load and save those from/to files.

To create an :py:class:`` from a file, just do the following:

Arrays are containers for :py:class:`numpy.ndarray`s or just pointers to a file. When you instantiate an :py:class:`` it does not load the file contents into memory. It waits until you emit another explicit instruction to do so. We do this with the :py:meth:`` method:

Every time you say :py:meth:``, the file contents will be read from the file and into a new array. Try again:

Calling :py:meth:`` will, by default, load the contents of the file into memory. Doing this means that subsequent calls to :py:meth:`` avoid having to read the file again and so do not incur an additional I/O cost.

Notice that, once the array is loaded in memory, a reference to the same array is shared every time you call :py:meth:``.

Saving the :py:class:`` is as easy, just call the :py:meth:`` method:

Numpy ndarray shortcuts

To just load an :py:class:`numpy.ndarray` in memory, you can use a short cut that lives at :py:func:``. This short cut means that you don't have to go through the :py:class:`` container:

You can also directly save :py:class:`numpy.ndarray`s without going through the :py:class:`` container:


Under the hood, we still use the :py:class:`` API to execute the read and write operations. This avoids code duplication and hooks data loading and saving to the powerful |project| transcoding framework that is explained next.

Reading and writing image and video data

|project| provides support to load and save data from many different file types including Matlab .mat files, various image file types and video data. File types and specific serialization and de-serialization is switched automatically using filename extensions. Knowing this, saving an array in a different format is just a matter of choosing the right extension. This is illustrated in the following example, where an image generated randomly using the method NumPy :py:meth:`numpy.random.random_integers()`, is saved in JPEG format. The image must be of type uint8 or uint16.

As for reading the video files, although it is possible to read a video using the :py:meth:``, you should use the methods of the class :py:class:`` to read frame by frame and avoid overloading your machine's memory. In the following example you can see how to create a video and save it using the class :py:class:`` and load it again using the class :py:class:``. The created video will have 30 frames generated randomly. Due to FFMPEG constrains, the width and height of the video need to be multiples of two.

The loaded image files are 3D arrays (for RGB format) or 2D arrays (for greyscale) of type uint8 or uint16, while the loaded videos are sequences of frames, usually 4D arrays of type uint8. All the extensions and formats for images and videos supported in your version of |project| can be listed using the |project|'s utility

|project| supports a number of binary formats in a manner similar to the cases shown above. Writing binary files is achieved using the :py:class:`` with the right file extension passed as an argument, just as was shown in the example above. These additional formats are:

  • Matlab (.mat), Matlab arrays, supports all integer, float and complex varieties [matlab.array.binary];
  • bob3 (.bindata), supports single or double precision float numbers, only 1-D [bob3.array.binary];
  • bob beta (.bin), supports all element types in |project| and any dimensionality [bob.array.binary] (deprecated);
  • bob alpha (.tensor) [tensor.array.binary] (deprecated);

Loading and saving Matlab data

An alternative for saving data in .mat files using :py:meth:``, would be to save them as a `HDF5`_ file which then can be easily read in Matlab. Similarly, instead of having to read .mat files using :py:meth:``, you can save your Matlab data in `HDF5`_ format, which then can be easily read from |project|. Detailed instructions about how to save and load data from Matlab to and from `HDF5`_ files can be found here.

Loading and saving audio files

|project| does not yet support audio files (no wav codec). However, it is possible to use the `SciPy`_ module :py:mod:`` to do the job. For instance, to read a wave file, just use the :py:func:`` function.

>>> import
>>> filename = '/home/user/sample.wav'
>>> samplerate, data =
>>> print type(data)
<type 'numpy.ndarray'>
>>> print data.shape
(132474, 2)

In the above example, the stereo audio signal is represented as a 2D NumPy :py:class:`numpy.ndarray`. The first dimension corresponds to the time index (132474 frames) and the second dimesnion correpsonds to one of the audio channel (2 channels, stereo). The values in the array correpsond to the wave magnitudes.

To save a NumPy :py:class:`numpy.ndarray` into a wave file, the :py:func:`` could be used, which also requires the framerate to be specified.

Jump to Line
Something went wrong with that request. Please try again.