Add batch sizing for less memory

This commit is contained in:
rstiskalek 2023-04-01 07:11:39 +01:00
parent 63ab3548b4
commit 070b9b6c1b

View file

@ -18,6 +18,7 @@ kNN-CDF calculation
from gc import collect
import numpy
from scipy.interpolate import interp1d
from scipy.stats import binned_statistic
from tqdm import tqdm
@ -125,7 +126,8 @@ class kNN_CDF:
return cdf
def __call__(self, *knns, nneighbours, Rmax, nsamples, rmin, rmax, neval,
verbose=True, random_state=42, dtype=numpy.float32):
batch_size=None, verbose=True, random_state=42,
left_nan=True, right_nan=True, dtype=numpy.float32):
"""
Calculate the CDF for a set of kNNs of CSiBORG halo catalogues.
@ -146,10 +148,20 @@ class kNN_CDF:
Maximum distance to evaluate the CDF.
neval : int
Number of points to evaluate the CDF.
batch_size : int, optional
Number of random points to sample in each batch. By default equal
to `nsamples`, however recommeded to be smaller to avoid requesting
too much memory,
verbose : bool, optional
Verbosity flag.
random_state : int, optional
Random state for the random number generator.
left_nan : bool, optional
Whether to set values where the CDF is 0 to `numpy.nan`. By
default `True`.
right_nan : bool, optional
Whether to set values where the CDF is 1 to `numpy.nan` after its
first occurence to 1. By default `True`.
dtype : numpy dtype, optional
Calculation data type. By default `numpy.float32`.
@ -160,22 +172,40 @@ class kNN_CDF:
cdfs : 2 or 3-dimensional array
CDFs evaluated at `rs`.
"""
rand = self.rvs_in_sphere(nsamples, Rmax, random_state=random_state)
batch_size = nsamples if batch_size is None else batch_size
assert nsamples >= batch_size
nbatches = nsamples // batch_size # Number of batches
cdfs = [None] * len(knns)
# Preallocate the bins and the CDF array
bins = numpy.logspace(numpy.log10(rmin), numpy.log10(rmax), neval)
cdfs = numpy.zeros((len(knns), nneighbours, neval - 1), dtype=dtype)
for i, knn in enumerate(tqdm(knns) if verbose else knns):
dist, _indxs = knn.kneighbors(rand, nneighbours)
dist = dist.astype(dtype)
del _indxs
collect()
# Loop over batches. This is to avoid generating large mocks
# requiring a lot of memory. Add counts to the CDF array
for j in range(nbatches):
rand = self.rvs_in_sphere(batch_size, Rmax,
random_state=random_state + j)
dist, _indxs = knn.kneighbors(rand, nneighbours)
for k in range(nneighbours): # Count for each neighbour
_counts, __, __ = binned_statistic(
dist[:, k], dist[:, k], bins=bins, statistic="count",
range=(rmin, rmax))
cdfs[i, k, :] += _counts
rs = (bins[1:] + bins[:-1]) / 2 # Bin centers
cdfs = numpy.cumsum(cdfs, axis=-1) # Cumulative sum, i.e. the CDF
for i in range(len(knns)):
for k in range(nneighbours):
cdfs[i, k, :] /= cdfs[i, k, -1]
# Set to NaN values after the first point where the CDF is 1
if right_nan:
ns = numpy.where(cdfs[i, k, :] == 1.)[0]
if ns.size > 1:
cdfs[i, k, ns[1]:] = numpy.nan
cdf = [None] * nneighbours
for j in range(nneighbours):
rs, cdf[j] = self.cdf_from_samples(
dist[:, j], rmin=rmin, rmax=rmax, neval=neval)
cdfs[i] = cdf
# Set to NaN values where the CDF is 0
if left_nan:
cdfs[cdfs == 0] = numpy.nan
cdfs = numpy.asanyarray(cdfs)
cdfs = cdfs[0, ...] if len(knns) == 1 else cdfs
return rs, cdfs