Skip to content

Latest commit

 

History

History
377 lines (250 loc) · 10.2 KB

array-creation.rst

File metadata and controls

377 lines (250 loc) · 10.2 KB
Raw

Create Dask Arrays

We store and manipulate large arrays in a wide variety of ways. There are some standards like HDF5 and NetCDF but just as often people use custom storage solutions. This page talks about how to build dask graphs to interact with your array.

In principle we need functions that return NumPy arrays. These functions and their arrangement can be as simple or as complex as the situation dictates.

Simple case - Format Supports NumPy Slicing

Many storage formats have Python projects that expose storage using NumPy slicing syntax. These include HDF5, NetCDF, BColz, Zarr, GRIB, etc.. For example the HDF5 file format has the h5py Python project, which provides a Dataset object into which we can slice in NumPy fashion.

>>> import h5py
>>> f = h5py.File('myfile.hdf5') # HDF5 file
>>> d = f['/data/path']          # Pointer on on-disk array
>>> d.shape                      # d can be very large
(1000000, 1000000)

>>> x = d[:5, :5]                # We slice to get numpy arrays

It is common for Python wrappers of on-disk array formats to present a NumPy slicing syntax. The full dataset looks like a NumPy array with .shape and .dtype attributes even though the data hasn't yet been loaded in and still lives on disk. Slicing in to this array-like object fetches the appropriate data from disk and returns that region as an in-memory NumPy array.

For this common case dask.array presents the convenience function da.from_array

>>> import dask.array as da
>>> x = da.from_array(d, chunks=(1000, 1000))

Concatenation and Stacking

Often we store data in several different locations and want to stitch them together.

>>> filenames = sorted(glob('2015-*-*.hdf5')
>>> dsets = [h5py.File(fn)['/data'] for fn in filenames]
>>> arrays = [da.from_array(dset, chunks=(1000, 1000)) for dset in dsets]
>>> x = da.concatenate(arrays, axis=0)  # Concatenate arrays along first axis

For more information see :doc:`concatenation and stacking <array-stack>` docs.

Using dask.delayed

You can create a plan to arrange many numpy arrays into a grid with normal for loops using :doc:`dask.delayed<delayed-overview>` and then convert each of these Dask.delayed objects into a single-chunk Dask array with da.from_delayed. You can then arrange these single-chunk Dask arrays into a larger multiple-chunk Dask array using :doc:`concatenation and stacking <array-stack>`, as described above.

See :doc:`documentation on using dask.delayed with collections<delayed-collections>`.

From Dask.dataframe

You can create dask arrays from dask dataframes using the .values attribute or the .to_records() method.

>>> x = df.values
>>> x = df.to_records()

However these arrays do not have known chunk sizes (dask.dataframe does not track the number of rows in each partition) and so some operations like slicing will not operate correctly.

Interactions with NumPy arrays

Dask.array operations will automatically convert NumPy arrays into single-chunk dask arrays

>>> x = da.sum(np.ones(5))
>>> x.compute()
5

When NumPy and Dask arrays interact the result will be a Dask array. Automatic rechunking rules will generally slice the NumPy array into the appropriate Dask chunk shape

>>> x = da.ones(10, chunks=(5,))
>>> y = np.ones(10)
>>> z = x + y
>>> z
dask.array<add, shape=(10,), dtype=float64, chunksize=(5,)>

These interactions work not just for NumPy arrays, but for any object that has shape and dtype attributes and implements NumPy slicing syntax.

Chunks

We always specify a chunks argument to tell dask.array how to break up the underlying array into chunks. This strongly impacts performance. We can specify chunks in one of three ways

  • a blocksize like 1000
  • a blockshape like (1000, 1000)
  • explicit sizes of all blocks along all dimensions, like ((1000, 1000, 500), (400, 400))

Your chunks input will be normalized and stored in the third and most explicit form.

For performance, a good choice of chunks follows the following rules:

  1. A chunk should be small enough to fit comfortably in memory. We'll have many chunks in memory at once.
  2. A chunk must be large enough so that computations on that chunk take significantly longer than the 1ms overhead per task that dask scheduling incurs. A task should take longer than 100ms.
  3. Chunks should align with the computation that you want to do. For example if you plan to frequently slice along a particular dimension then it's more efficient if your chunks are aligned so that you have to touch fewer chunks. If you want to add two arrays then its convenient if those arrays have matching chunks patterns.

Unknown Chunks

Some arrays have unknown chunk sizes. These are designated using np.nan rather than an integer. These arrays support many but not all operations. In particular, operations like slicing are not possible and will result in an error.

>>> x.shape
(np.nan, np.nan)

>>> x[0]
ValueError: Array chunk sizes unknown

Chunks Examples

We show of how different inputs for chunks= cut up the following array:

1 2 3 4 5 6
7 8 9 0 1 2
3 4 5 6 7 8
9 0 1 2 3 4
5 6 7 8 9 0
1 2 3 4 5 6

We show how different chunks= arguments split the array into different blocks

chunks=3: Symmetric blocks of size 3:

1 2 3  4 5 6
7 8 9  0 1 2
3 4 5  6 7 8

9 0 1  2 3 4
5 6 7  8 9 0
1 2 3  4 5 6

chunks=2: Symmetric blocks of size 2:

1 2  3 4  5 6
7 8  9 0  1 2

3 4  5 6  7 8
9 0  1 2  3 4

5 6  7 8  9 0
1 2  3 4  5 6

chunks=(3, 2): Asymmetric but repeated blocks of size (3, 2):

1 2  3 4  5 6
7 8  9 0  1 2
3 4  5 6  7 8

9 0  1 2  3 4
5 6  7 8  9 0
1 2  3 4  5 6

chunks=(1, 6): Asymmetric but repeated blocks of size (1, 6):

1 2 3 4 5 6

7 8 9 0 1 2

3 4 5 6 7 8

9 0 1 2 3 4

5 6 7 8 9 0

1 2 3 4 5 6

chunks=((2, 4), (3, 3)): Asymmetric and non-repeated blocks:

1 2 3  4 5 6
7 8 9  0 1 2

3 4 5  6 7 8
9 0 1  2 3 4
5 6 7  8 9 0
1 2 3  4 5 6

chunks=((2, 2, 1, 1), (3, 2, 1)): Asymmetric and non-repeated blocks:

1 2 3  4 5  6
7 8 9  0 1  2

3 4 5  6 7  8
9 0 1  2 3  4

5 6 7  8 9  0

1 2 3  4 5  6

Discussion

The latter examples are rarely provided by users on original data but arise from complex slicing and broadcasting operations. Generally people use the simplest form until they need more complex forms. The choice of chunks should align with the computations you want to do.

For example, if you plan to take out thin slices along the first dimension then you might want to make that dimension skinnier than the others. If you plan to do linear algebra then you might want more symmetric blocks.

Store Dask Arrays

In Memory

If you have a small amount of data, you can call np.array or .compute() on your Dask array to turn in to a normal NumPy array:

>>> x = da.arange(6, chunks=3)
>>> y = x**2
>>> np.array(y)
array([0, 1, 4, 9, 16, 25])

>>> y.compute()
array([0, 1, 4, 9, 16, 25])

HDF5

Use the to_hdf5 function to store data into HDF5 using h5py:

>>> da.to_hdf5('myfile.hdf5', '/y', y)  # doctest: +SKIP

Store several arrays in one computation with the function da.to_hdf5 by passing in a dict:

>>> da.to_hdf5('myfile.hdf5', {'/x': x, '/y': y})  # doctest: +SKIP

Other On-Disk Storage

Alternatively, you can store dask arrays in any object that supports numpy-style slice assignment like h5py.Dataset, or bcolz.carray:

>>> import bcolz  # doctest: +SKIP
>>> out = bcolz.zeros(shape=y.shape, rootdir='myfile.bcolz')  # doctest: +SKIP
>>> da.store(y, out)  # doctest: +SKIP

You can store several arrays in one computation by passing lists of sources and destinations:

>>> da.store([array1, array2], [output1, output2])  # doctest: +SKIP

Plugins

We can run arbitrary user-defined functions on dask.arrays whenever they are constructed. This allows us to build a variety of custom behaviors that improve debugging, user warning, etc.. You can register a list of functions to run on all dask.arrays to the global array_plugins= value:

>>> def f(x):
...     print(x.nbytes)

>>> with dask.set_options(array_plugins=[f]):
...     x = da.ones((10, 1), chunks=(5, 1))
...     y = x.dot(x.T)
80
80
800
800

If the plugin function returns None then the input Dask.array will be returned without change. If the plugin function returns something else then that value will be the result of the constructor.

Examples

Automatically compute

We may wish to turn some Dask.array code into normal NumPy code. This is useful for example to track down errors immediately that would otherwise be hidden by Dask's lazy semantics.

>>> with dask.set_options(array_plugins=[lambda x: x.compute()]):
...     x = da.arange(5, chunks=2)

>>> x  # this was automatically converted into a numpy array
array([0, 1, 2, 3, 4])

Warn on large chunks

We may wish to warn users if they are creating chunks that are too large

def warn_on_large_chunks(x):
    shapes = list(itertools.product(*x.chunks))
    nbytes = [x.dtype.itemsize * np.prod(shape) for shape in shapes]
    if any(nb > 1e9 for nb in nbytes):
        warnings.warn("Array contains very large chunks")

with dask.set_options(array_plugins=[warn_on_large_chunks]):
    ...

Combine

You can also combine these plugins into a list. They will run one after the other, chaining results through them.

with dask.set_options(array_plugins=[warn_on_large_chunks, lambda x: x.compute()]):
    ...