Support for dask arrays

It is possible to operate on dask arrays and spare the memory (or perhaps even time).

[1]:

# Necessary imports
import dask
import dask.multiprocessing
import physt
import numpy as np

import dask.array as da
from physt import h1, h2
%matplotlib inline

[2]:

# Create two arrays
np.random.seed(42)

SIZE = 2 ** 21
CHUNK = int(SIZE / 16)

million = np.random.rand(SIZE)#.astype(int)
million2 = (3 * million + np.random.normal(0., 0.3, SIZE))#.astype(int)

# Chunk them for dask
chunked = da.from_array(million, chunks=(CHUNK))
chunked2 = da.from_array(million2, chunks=(CHUNK))

Create histograms

h1, h2, … have their alternatives in physt.dask_compat. They should work similarly. Although, they are not complete and unexpected errors may occur.

[3]:

from physt.compat.dask import h1 as d1
from physt.compat.dask import h2 as d2

[4]:

# Use chunks to create a 1D histogram
ha = d1(chunked2, "fixed_width", bin_width=0.2)
check_ha = h1(million2, "fixed_width", bin_width=0.2)
ok = (ha == check_ha)
print("Check: ", ok)
ha.plot()
ha

Check:  True

[4]:

Histogram1D(bins=(28,), total=2097152, dtype=int64)
_images/dask_5_2.png 
[5]:

# Use chunks to create a 2D histogram
hb = d2(chunked, chunked2, "fixed_width", bin_width=.2, axis_names=["x", "y"])
check_hb = h2(million, million2, "fixed_width", bin_width=.2, axis_names=["x", "y"])
hb.plot(show_zero=False, cmap="rainbow")
ok = (hb == check_hb)
print("Check: ", ok)
hb

Check:  True

[5]:

Histogram2D(bins=(5, 28), total=2097152, dtype=int64)
_images/dask_6_2.png 
[6]:

# And another cross-check
hh = hb.projection("y")
hh.plot()
print("Check: ", np.array_equal(hh.frequencies, ha.frequencies))   # Just frequencies

Check:  True
_images/dask_7_1.png 
[8]:

# Use dask for normal arrays (will automatically split array to chunks)
d1(million2, "fixed_width", bin_width=0.2) == ha

[8]:

True

Some timings

Your results may vary substantially. These numbers are just for illustration, on 4-core (8-thread) machine. The real gain comes when we have data that don’t fit into memory.

Efficiency

[9]:

# Standard
%time h1(million2, "fixed_width", bin_width=0.2)

Wall time: 361 ms

[9]:

Histogram1D(bins=(28,), total=2097152, dtype=int64)

[10]:

# Same array, but using dask
%time d1(million2, "fixed_width", bin_width=0.2)

Wall time: 116 ms

[10]:

Histogram1D(bins=(28,), total=2097152, dtype=int64)

[11]:

# Most efficient: dask with already chunked data
%time d1(chunked2, "fixed_width", bin_width=0.2)

Wall time: 91.8 ms

[11]:

Histogram1D(bins=(28,), total=2097152, dtype=int64)

Different scheduling

[12]:

%time d1(chunked2, "fixed_width", bin_width=0.2)

Wall time: 76 ms

[12]:

Histogram1D(bins=(28,), total=2097152, dtype=int64)

[13]:

%%time
# Hyper-threading or not?
graph, name = d1(chunked2, "fixed_width", bin_width=0.2, compute=False)
dask.threaded.get(graph, name, num_workers=4)

Wall time: 114 ms

[13]:

Histogram1D(bins=(28,), total=2097152, dtype=int64)

[14]:

# Multiprocessing not so efficient for small arrays?
%time d1(chunked2, "fixed_width", bin_width=0.2, dask_method=dask.multiprocessing.get)

Wall time: 960 ms

[14]:

Histogram1D(bins=(28,), total=2097152, dtype=int64)