It is possible to operate on dask arrays and spare the memory (or perhaps even time).
# 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
# 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))
h1
, h2
, ... have their alternatives in physt.dask_compat
. They should work similarly. Although, they are not complete and unexpected errors may occur.
from physt.compat.dask import h1 as d1
from physt.compat.dask import h2 as d2
# Use chunks to create a 1D histogram
ha = d1(chunked2, "fixed_width", 0.2)
check_ha = h1(million2, "fixed_width", 0.2)
ok = (ha == check_ha)
print("Check: ", ok)
ha.plot()
ha
Check: True
Histogram1D(bins=(28,), total=2097152, dtype=int64)
# Use chunks to create a 2D histogram
hb = d2(chunked, chunked2, "fixed_width", .2, axis_names=["x", "y"])
check_hb = h2(million, million2, "fixed_width", .2, axis_names=["x", "y"])
hb.plot(show_zero=False, cmap="rainbow")
ok = (hb == check_hb)
print("Check: ", ok)
hb
Check: True
Histogram2D(bins=(5, 28), total=2097152, dtype=int64)
# And another cross-check
hh = hb.projection("y")
hh.plot()
print("Check: ", np.array_equal(hh.frequencies, ha.frequencies)) # Just frequencies
Check: True
# Use dask for normal arrays (will automatically split array to chunks)
d1(million2, "fixed_width", 0.2) == ha
True
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.
# Standard
%time h1(million2, "fixed_width", 0.2)
CPU times: user 485 ms, sys: 391 ms, total: 876 ms Wall time: 268 ms
Histogram1D(bins=(28,), total=2097152, dtype=int64)
# Same array, but using dask
%time d1(million2, "fixed_width", 0.2)
CPU times: user 389 ms, sys: 10 ms, total: 399 ms Wall time: 95.4 ms
Histogram1D(bins=(28,), total=2097152, dtype=int64)
# Most efficient: dask with already chunked data
%time d1(chunked2, "fixed_width", 0.2)
CPU times: user 412 ms, sys: 9.86 ms, total: 422 ms Wall time: 73.1 ms
Histogram1D(bins=(28,), total=2097152, dtype=int64)
%time d1(chunked2, "fixed_width", 0.2)
CPU times: user 383 ms, sys: 7.54 ms, total: 391 ms Wall time: 74.3 ms
Histogram1D(bins=(28,), total=2097152, dtype=int64)
%%time
# Hyper-threading or not?
graph, name = d1(chunked2, "fixed_width", 0.2, compute=False)
dask.threaded.get(graph, name, num_workers=4)
CPU times: user 381 ms, sys: 13.4 ms, total: 394 ms Wall time: 128 ms
Histogram1D(bins=(28,), total=2097152, dtype=int64)
# Multiprocessing not so efficient for small arrays?
%time d1(chunked2, "fixed_width", 0.2, dask_method=dask.multiprocessing.get)
CPU times: user 262 ms, sys: 360 ms, total: 622 ms Wall time: 714 ms
Histogram1D(bins=(28,), total=2097152, dtype=int64)