Working with big data

Warning

All the features described in this chapter are in beta state.

Although most of them work as described, their operation may not always be optimal, well-documented and/or consistent with their in-memory counterparts.

Therefore, although efforts will be taken to minimise major disruptions, the syntax and features described here may change in patch and minor HyperSpy releases. If you experience issues with HyperSpy’s lazy features please report them to the developers.

New in version 1.2.

HyperSpy makes it possible to analyse data larger than the available memory by providing “lazy” versions of most of its signals and functions. In most cases the syntax remains the same. This chapter describes how to work with data larger than memory using the LazySignal class and its derivatives.

Creating Lazy Signals

Lazy Signals from external data

If the data is large and not loaded by HyperSpy (for example a hdf5.Dataset or similar), first wrap it in dask.array.Array as shown here and then pass it as normal and call as_lazy():

>>> import h5py
>>> f = h5py.File("myfile.hdf5")           # Load the file
>>> data = f['/data/path']                 # Get the data
>>> import dask.array as da                # Import dask to wrap
>>> chunks = (1000,100)                    # Chunk as appropriate
>>> x = da.from_array(data, chunks=chunks) # Wrap the data in dask
>>> s = hs.signals.Signal1D(x).as_lazy() # Create the lazy signal

Loading lazily

To load the data lazily, pass the keyword lazy=True. As an example, loading a 34.9 GB .blo file on a regular laptop might look like:

>>> s = hs.load("shish26.02-6.blo", lazy=True)
>>> s
<LazySignal2D, title: , dimensions: (400, 333|512, 512)>
>>> s.data
dask.array<array-e..., shape=(333, 400, 512, 512), dtype=uint8, chunksize=(20, 12, 512, 512)>
>>> print(s.data.dtype, s.data.nbytes / 1e9)
uint8 34.9175808
>>> s.change_dtype("float") # To be able to perform decomposition, etc.
>>> print(s.data.dtype, s.data.nbytes / 1e9)
float64 279.3406464

Loading the dataset in the original unsigned integer format would require around 35GB of memory. To store it in a floating-point format one would need almost 280GB of memory. However, with the lazy processing both of these steps are near-instantaneous and require very little computational resources.

New in version 1.4: close_file()

Currently when loading an hdf5 file lazily the file remains open at least while the signal exists. In order to close it explicitly, use the close_file() method. Alternatively, you could close it on calling compute() by passing the keyword argument close_file=True e.g.:

>>> s = hs.load("file.hspy", lazy=True)
>>> ssum = s.sum(axis=0)
>>> ssum.compute(close_file=True) # closes the file.hspy file

Lazy stacking

Occasionally the full dataset consists of many smaller files. To combine them into a one large LazySignal, we can stack them lazily (both when loading or afterwards):

>>> siglist = hs.load("*.hdf5")
>>> s = hs.stack(siglist, lazy=True)
>>> # Or load lazily and stack afterwards:
>>> siglist = hs.load("*.hdf5", lazy=True)
>>> s = hs.stack(siglist) # no need to pass 'lazy', as signals already lazy
>>> # Or do everything in one go:
>>> s = hs.load("*.hdf5", lazy=True, stack=True)

Casting signals as lazy

To convert a regular HyperSpy signal to a lazy one such that any future operations are only performed lazily, use the as_lazy() method:

>>> s = hs.signals.Signal1D(np.arange(150.).reshape((3, 50)))
>>> s
<Signal1D, title: , dimensions: (3|50)>
>>> sl = s.as_lazy()
>>> sl
<LazySignal1D, title: , dimensions: (3|50)>

Machine learning

Decomposition algorithms for machine learning often perform large matrix manipulations, requiring significantly more memory than the data size. To perform decomposition operation lazily, HyperSpy provides access to several “online” algorithms as well as dask’s lazy SVD algorithm. Online algorithms perform the decomposition by operating serially on chunks of data, enabling the lazy decomposition of large datasets. In line with the standard HyperSpy signals, lazy decomposition() offers the following online algorithms:

Available lazy decomposition algorithms in HyperSpy

Algorithm	Method
“SVD” (default)	dask.array.linalg.svd()
“PCA”	sklearn.decomposition.IncrementalPCA
“ORPCA”	ORPCA
“ORNMF”	ORNMF

See also

decomposition() for more details on decomposition with non-lazy signals.

Navigator plot

The default signal navigator is the sum of the signal across all signal dimensions and all but 1 or 2 navigation dimensions. If the dataset is large, this can take a significant amount of time to perform with every plot. By default, a navigator is computed with minimally required approach to obtain a good signal-to-noise ratio image: the sum is taken on a single chunk of the signal space, in order to avoid to compute the navigator for the whole dataset. In the following example, the signal space is divided in 25 chunks (5 along on each axis), and therefore computing the navigation will only be perfomed over a small subset of the whole dataset by taking the sum on only 1 chunk out of 25:

>>> import dask.array as da
>>> import hyperspy.api as hs
>>> data = da.random.random((100, 100, 1000, 1000), chunks=('auto', 'auto', 200, 200))
>>> s = hs.signals.Signal2D(data).as_lazy()
>>> s.plot()

In the example above, the calculation of the navigation is fast but the actual visualisation of the dataset is slow, each for each navigation index change, 25 chunks of the dataset needs to be fetched from the harddrive. In the following example, the signal space contains a single chunk (instead of 25, in the previous example) and the calculating the navigator will then be slower (~20x) because the whole dataset will need to processed, however in this case, the visualisation will be faster, because only a single chunk will fetched from the harddrive when changing navigation indices:

>>> data = da.random.random((100, 100, 1000, 1000), chunks=('auto', 'auto', 1000, 1000))
>>> s = hs.signals.Signal2D(data).as_lazy()
>>> s.plot()

This approach depends heavily on the chunking of the data and may not be always suitable. The compute_navigator() can be used to calculate the navigator efficient and store the navigator, so that it can be used when plotting and saved for the later loading of the dataset. The compute_navigator() has optional argument to specify the index where the sum needs to be calculated and how to rechunk the dataset when calculating the navigator. This allows to efficiently calculate the navigator without changing the actual chunking of the dataset, since the rechunking only takes during the computation of the navigator:

>>> data = da.random.random((100, 100, 1000, 1000), chunks=('auto', 'auto', 100, 100))
>>> s = hs.signals.Signal2D(data).as_lazy()
>>> s.compute_navigator(chunks_number=5)
>>> s.plot()

>>> data = da.random.random((100, 100, 2000, 400), chunks=('auto', 'auto', 100, 100))
>>> s = hs.signals.Signal2D(data).as_lazy()
>>> s
<LazySignal2D, title: , dimensions: (100, 100|400, 2000)>
>>> s.compute_navigator(chunks_number=(2, 10))
>>> s.plot()
>>> s.navigator.original_metadata
└── sum_from = [slice(200, 400, None), slice(1000, 1200, None)]

The index can also be specified following the HyperSpy indexing signal1D syntax for float and interger.

>>> data = da.random.random((100, 100, 2000, 400), chunks=('auto', 'auto', 100, 100))
>>> s = hs.signals.Signal2D(data).as_lazy()
>>> s
<LazySignal2D, title: , dimensions: (100, 100|400, 2000)>
>>> s.compute_navigator(index=0, chunks_number=(2, 10))
>>> s.navigator.original_metadata
└── sum_from = [slice(0, 200, None), slice(0, 200, None)]

An alternative is to calculate the navigator separately and store it in the signal using the navigator setter.

>>> data = da.random.random((100, 100, 1000, 1000), chunks=('auto', 'auto', 100, 100))
>>> s = hs.signals.Signal2D(data).as_lazy()
>>> s
<LazySignal2D, title: , dimensions: (100, 100|1000, 1000)>
>>> # for fastest results, just pick one signal space pixel
>>> nav = s.isig[500, 500]
>>> # Alternatively, sum as per default behaviour of non-lazy signal
>>> nav = s.sum(s.axes_manager.signal_axes)
>>> nav
<LazySignal2D, title: , dimensions: (|100, 100)>
>>> # Compute the result
>>> nav.compute()
[########################################] | 100% Completed | 13.1s
>>> s.navigator = nav
>>> s.plot()

Alternatively, it is possible to not have a navigator, and use sliders instead:

>>> s
<LazySignal2D, title: , dimensions: (200, 200|512, 512)>
>>> s.plot(navigator='slider')

Practical tips

Despite the limitations detailed below, most HyperSpy operations can be performed lazily. Important points are:

Chunking

Data saved in the HDF5 format is typically divided into smaller chunks which can be loaded separately into memory, allowing lazy loading. Chunk size can dramatically affect the speed of various HyperSpy algorithms, so chunk size is worth careful consideration when saving a signal. HyperSpy’s default chunking sizes are probably not optimal for a given data analysis technique. For more comprehensible documentation on chunking, see the dask array chunks and best practices docs. The chunks saved into HDF5 will match the dask array chunks in s.data.chunks when lazy loading. Chunk shape should follow the axes order of the numpy shape (s.data.shape), not the hyperspy shape. The following example shows how to chunk one of the two navigation dimensions into smaller chunks:

>>> import dask.array as da
>>> data = da.random.random((10,200,300))
>>> data.chunksize
(10, 200, 300)

>>> s = hs.signals.Signal1D(data)
>>> s # Note the reversed order of navigation dimensions
<Signal1D, title: , dimensions: (200, 10|300)>

>>> s.save('chunked_signal.hspy', chunks=(10, 100, 300)) # Chunking first hyperspy dimension (second array dimension)
>>> s2 = hs.load('chunked_signal.hspy', lazy=True)
>>> s2.data.chunksize
(10, 100, 300)

New in version 1.3.2.

By default, HyperSpy tries to optimize the chunking for most operations. However, it is sometimes possible to manually set a more optimal chunking manually. Therefore, many operations take a rechunk or optimize keyword argument to disable automatic rechunking.

Computing lazy signals

Upon saving lazy signals, the result of computations is stored on disk.

In order to store the lazy signal in memory (i.e. make it a normal HyperSpy signal) it has a compute() method:

>>> s
<LazySignal2D, title: , dimensions: (|512, 512)>
>>> s.compute()
[########################################] | 100% Completed |  0.1s
>>> s
<Signal2D, title: , dimensions: (|512, 512)>

Lazy operations that affect the axes

When using lazy signals the computation of the data is delayed until requested. However, the changes to the axes properties are performed when running a given function that modfies them i.e. they are not performed lazily. This can lead to hard to debug issues when the result of a given function that is computed lazily depends on the value of the axes parameters that may have changed before the computation is requested. Therefore, in order to avoid such issues, it is reccomended to explicitly compute the result of all functions that are affected by the axes parameters. This is the reason why e.g. the result of shift1D() is not lazy.

Limitations

Most operations can be performed lazily. However, lazy operations come with a few limitations and constraints that we detail below.

Immutable signals

An important limitation when using LazySignal is the inability to modify existing data (immutability). This is a logical consequence of the DAG (tree structure, explained in Behind the scenes –technical details), where a complete history of the processing has to be stored to traverse later.

In fact, lazy evaluation removes the need for such operation, since only additional tree branches are added, requiring very little resources. In practical terms the following fails with lazy signals:

>>> s = hs.signals.BaseSignal([0]).as_lazy()
>>> s += 1
Traceback (most recent call last):
  File "<ipython-input-6-1bd1db4187be>", line 1, in <module>
    s += 1
  File "<string>", line 2, in __iadd__
  File "/home/fjd29/Python/hyperspy3/hyperspy/signal.py", line 1591, in _binary_operator_ruler
    getattr(self.data, op_name)(other)
AttributeError: 'Array' object has no attribute '__iadd__'

However, when operating lazily there is no clear benefit to using in-place operations. So, the operation above could be rewritten as follows:

>>> s = hs.signals.BaseSignal([0]).as_lazy()
>>> s = s + 1

Or even better:

>>> s = hs.signals.BaseSignal([0]).as_lazy()
>>> s1 = s + 1

Other minor differences

• Histograms for a LazySignal do not support knuth and blocks binning algorithms.

• CircleROI sets the elements outside the ROI to np.nan instead of using a masked array, because dask does not support masking. As a convenience, nansum, nanmean and other nan* signal methods were added to mimic the workflow as closely as possible.

Behind the scenes –technical details

Standard HyperSpy signals load the data into memory for fast access and processing. While this behaviour gives good performance in terms of speed, it obviously requires at least as much computer memory as the dataset, and often twice that to store the results of subsequent computations. This can become a significant problem when processing very large datasets on consumer-oriented hardware.

HyperSpy offers a solution for this problem by including LazySignal and its derivatives. The main idea of these classes is to perform any operation (as the name suggests) lazily (delaying the execution until the result is requested (e.g. saved, plotted)) and in a blocked fashion. This is achieved by building a “history tree” (formally called a Directed Acyclic Graph (DAG)) of the computations, where the original data is at the root, and any further operations branch from it. Only when a certain branch result is requested, the way to the root is found and evaluated in the correct sequence on the correct blocks.

The “magic” is performed by (for the sake of simplicity) storing the data not as numpy.ndarray, but dask.array.Array (see the dask documentation). dask offers a couple of advantages:

• Arbitrary-sized data processing is possible. By only loading a couple of chunks at a time, theoretically any signal can be processed, albeit slower. In practice, this may be limited: (i) some operations may require certain chunking pattern, which may still saturate memory; (ii) many chunks should fit into the computer memory comfortably at the same time.

• Loading only the required data. If a certain part (chunk) of the data is not required for the final result, it will not be loaded at all, saving time and resources.

• Able to extend to a distributed computing environment (clusters). :py:dask.distributed (see the dask documentation) offers a straightforward way to expand the effective memory for computations to that of a cluster, which allows performing the operations significantly faster than on a single machine.