"""
redrock.external.desi
=====================
redrock wrapper tools for DESI
"""
import os
import sys
import re
import warnings
import traceback
import itertools
import argparse
import numpy as np
from astropy.io import fits
from astropy.table import Table, vstack
from desiutil.io import encode_table
from desiutil.depend import add_dependencies, setdep
from desispec.resolution import Resolution
from desispec.coaddition import coadd_fibermap
from desispec.specscore import compute_coadd_tsnr_scores
from desispec.maskbits import fibermask
from desispec.io.util import get_tempfilename
from desispec.io.fibermap import annotate_fibermap
from ..utils import elapsed, get_mp, distribute_work, getGPUCountMPI
from ..targets import (Spectrum, Target, DistTargets)
from ..templates import load_dist_templates
from ..results import write_zscan
from ..zfind import zfind
from ..zwarning import ZWarningMask
from .._version import __version__
from ..archetypes import All_archetypes, Archetype
[docs]def write_zbest(outfile, zbest, fibermap, exp_fibermap, tsnr2,
templates, archetypes=None,
spec_header=None):
"""Write zbest and fibermap Tables to outfile
Args:
outfile (str): output path.
zbest (Table): best fit table.
fibermap (Table): the coadded fibermap from the original inputs.
tsnr2 (Table): table of input coadded TSNR2 values
exp_fibermap (Table): the per-exposure fibermap from the orig inputs.
templates : list or dict of template objects
Options:
archetypes : list or dict of Archetype objects
spec_header (dict-like): header of HDU 0 of input spectra
Modifies input tables.meta['EXTNAME']
"""
header = _get_header(templates, archetypes, spec_header)
zbest.meta['EXTNAME'] = 'REDSHIFTS'
fibermap.meta['EXTNAME'] = 'FIBERMAP'
exp_fibermap.meta['EXTNAME'] = 'EXP_FIBERMAP'
tsnr2.meta['EXTNAME'] = 'TSNR2'
hx = fits.HDUList()
hx.append(fits.PrimaryHDU(header=header))
hx.append(fits.convenience.table_to_hdu(zbest))
# ADM annotate_fibermap adds units to fibermap HDUs.
hx.append(annotate_fibermap(fits.convenience.table_to_hdu(fibermap)))
hx.append(fits.convenience.table_to_hdu(exp_fibermap))
hx.append(fits.convenience.table_to_hdu(tsnr2))
outfile = os.path.expandvars(outfile)
outdir = os.path.dirname(os.path.abspath(outfile))
if not os.path.exists(outdir):
os.makedirs(outdir)
tempfile = outfile + '.tmp'
hx.writeto(tempfile, overwrite=True)
os.rename(tempfile, outfile)
return
[docs]def write_bestmodel(outfile, zbest, modeldict, wavedict, templates,
archetypes=None, spec_header=None):
"""
Writes best fit redrock model in the outfile
Args:
outfile (str): output file name (fits)
zbest (Table): best fit redshift table
modeldict (dict): models[ntargets, nwave] keyed by camera band
wavedict (dict): wavelength dictionary keyed by camera band
templates : list or dict of template objects
Options:
archetypes : list or dict of Archetype objects
spec_header (dict-like): header of HDU 0 of input spectra
The output file format mirrors the structure of an input coadd, with B/R/Z_MODEL
HDUs instead of B/R/Z_FLUX HDUs.
"""
header = _get_header(templates, archetypes, spec_header)
zbest.meta['EXTNAME'] = 'REDSHIFTS'
hx = fits.HDUList()
hx.append(fits.PrimaryHDU(header=header))
hx.append(fits.convenience.table_to_hdu(zbest))
for band in wavedict.keys():
BAND = band.upper()
hdu = fits.ImageHDU(name=f'{BAND}_WAVELENGTH')
hdu.data = wavedict[band]
hdu.header["BUNIT"] = "Angstrom"
hx.append(hdu)
hdu = fits.ImageHDU(name=f'{BAND}_MODEL')
hdu.data = modeldict[band].astype('float32')
hx.append(hdu)
outfile = os.path.expandvars(outfile)
outdir = os.path.dirname(os.path.abspath(outfile))
if not os.path.exists(outdir):
os.makedirs(outdir)
tempfile = get_tempfilename(outfile)
hx.writeto(tempfile, overwrite=True)
os.rename(tempfile, outfile)
return
[docs]class DistTargetsDESI(DistTargets):
"""Distributed targets for DESI.
DESI spectral data is grouped by sky location, but is just a random
collection of spectra for all targets. Read this into memory while
grouping by target ID, preserving order in which each target first appears.
We pass through the spectra files once to compute all the book-keeping
associated with regrouping the spectra by target. Then we pass through
again and actually read and distribute the data.
Args:
spectrafiles (str or list): a list of input files or pattern match
of files.
coadd (bool): if False, do not compute the coadds.
targetids (list): (optional) restrict the global set of target IDs
to this list.
first_target (int): (optional) integer offset of the first target to
consider in each file. Useful for debugging / testing.
n_target (int): (optional) number of targets to consider in each file.
Useful for debugging / testing.
comm (mpi4py.MPI.Comm): (optional) the MPI communicator.
cache_Rcsr: pre-calculate and cache sparse CSR format of resolution
matrix R
cosmics_nsig (float): cosmic rejection threshold used in coaddition
negflux_nsig (float): mask negative flux significance threshold
capacities (list): (optional) list of process capacities. If None,
use equal capacity per process. A process with higher capacity
can handle more work.
"""
### @profile
def __init__(self, spectrafiles, coadd=True, targetids=None,
first_target=None, n_target=None, comm=None, cache_Rcsr=False,
cosmics_nsig=0, negflux_nsig=5, capacities=None):
comm_size = 1
comm_rank = 0
if comm is not None:
comm_size = comm.size
comm_rank = comm.rank
# check the file list
if isinstance(spectrafiles, str):
import glob
spectrafiles = glob.glob(spectrafiles)
assert len(spectrafiles) > 0
self._spectrafiles = spectrafiles
self.cosmics_nsig = cosmics_nsig
# This is the mapping between specs to targets for each file
self._spec_to_target = {}
self._target_specs = {}
self._spec_keep = {}
self._spec_sliced = {}
# The bands for each file
self._bands = {}
self._wave = {}
# The full list of targets from all files
self._alltargetids = list()
# The fibermaps from all files
self._coadd_fmaps = {}
self._exp_fmaps = {}
self._tsnr2 = {} #- template signal-to-noise from SCORES
self.header0 = None #- header 0 of the first spectrafile
for sfile in spectrafiles:
hdus = None
nhdu = None
input_coadded = 'unknown'
coadd_fmap = None
exp_fmap = None
tsnr2 = None
if comm_rank == 0:
hdus = fits.open(sfile, memmap=False)
nhdu = len(hdus)
if self.header0 is None:
self.header0 = hdus[0].header.copy()
if 'EXP_FIBERMAP' in hdus:
input_coadded = True
coadd_fmap = encode_table(Table(hdus["FIBERMAP"].data,
copy=True).as_array())
exp_fmap = encode_table(Table(hdus["EXP_FIBERMAP"].data,
copy=True).as_array())
tsnr2 = encode_table(Table(hdus["SCORES"].data,
copy=True).as_array())
for col in tsnr2.colnames.copy():
if col == 'TARGETID' or col.startswith('TSNR2_'):
continue
else:
tsnr2.remove_column(col)
else:
input_coadded = False
tmpfmap = encode_table(Table(hdus["FIBERMAP"].data,
copy=True).as_array())
assert 'COADD_NUMEXP' not in tmpfmap.dtype.names
if np.all(tmpfmap['TILEID'] == tmpfmap['TILEID'][0]):
onetile = True
else:
onetile = False
coadd_fmap, exp_fmap = coadd_fibermap(tmpfmap, onetile=onetile)
scores = encode_table(Table(hdus["SCORES"].data,
copy=True).as_array())
tsnr2 = Table(compute_coadd_tsnr_scores(scores)[0])
#- we later rely upon exp_fmap having same order as the
#- uncoadded input fmap, so check that now
assert np.all(exp_fmap['TARGETID'] == tmpfmap['TARGETID'])
if comm is not None:
nhdu = comm.bcast(nhdu, root=0)
input_coadded = comm.bcast(input_coadded, root=0)
coadd_fmap = comm.bcast(coadd_fmap, root=0)
exp_fmap = comm.bcast(exp_fmap, root=0)
tsnr2 = comm.bcast(tsnr2, root=0)
self.header0 = comm.bcast(self.header0, root=0)
# Now every process has the fibermap and number of HDUs. Build the
# mapping between spectral rows and target IDs.
if targetids is None:
keep_targetids = coadd_fmap["TARGETID"]
else:
keep_targetids = targetids
# Select a subset of the target range from each file if desired.
if first_target is None:
first_target = 0
if first_target > len(keep_targetids):
raise RuntimeError("first_target value \"{}\" is beyond the "
"number of selected targets in the file".\
format(first_target))
if n_target is None:
nkeep = len(keep_targetids)
else:
nkeep = n_target
if first_target + nkeep > len(keep_targetids):
msg = "Requested first_target ({}) + nkeep ({})".format(
first_target, nkeep)
msg += " is larger than number of selected targets ({})".format(
len(keep_targetids))
raise RuntimeError(msg)
keep_targetids = keep_targetids[first_target:first_target+nkeep]
self._alltargetids.extend(keep_targetids)
# This is the spectral row to target mapping using the original
# global indices (before slicing).
if input_coadded:
input_targetids = coadd_fmap['TARGETID']
else:
input_targetids = exp_fmap['TARGETID']
self._spec_to_target[sfile] = [ x if y in keep_targetids else -1 \
for x, y in enumerate(input_targetids) ]
# The reduced set of spectral rows.
self._spec_keep[sfile] = [ x for x in self._spec_to_target[sfile] \
if x >= 0 ]
# The mapping between original spectral indices and the sliced ones
self._spec_sliced[sfile] = { x : y for y, x in \
enumerate(self._spec_keep[sfile]) }
# Slice the fibermap to keep just the requested targets
keep_coadd = np.isin(coadd_fmap['TARGETID'], keep_targetids)
self._coadd_fmaps[sfile] = coadd_fmap[keep_coadd]
self._tsnr2[sfile] = tsnr2[keep_coadd]
keep_exp = np.isin(exp_fmap['TARGETID'], keep_targetids)
self._exp_fmaps[sfile] = exp_fmap[keep_exp]
if input_coadded:
input_targetids = input_targetids[keep_coadd]
else:
input_targetids = input_targetids[keep_exp]
# For each target, store the sliced row index of all spectra,
# so that we can do a fast lookup later.
self._target_specs[sfile] = {}
for id in keep_targetids:
self._target_specs[sfile][id] = [ x for x, y in \
enumerate(input_targetids) if y == id ]
# We need some more metadata information for each file-
# specifically, the bands that are used and their wavelength grids.
# That information will allow us to pre-allocate our local target
# list and then fill that with one pass through all HDUs in the
# files.
self._bands[sfile] = []
self._wave[sfile] = dict()
if comm_rank == 0:
for h in range(nhdu):
name = None
if "EXTNAME" not in hdus[h].header:
continue
name = hdus[h].header["EXTNAME"]
mat = re.match(r"(.*)_(.*)", name)
if mat is None:
continue
band = mat.group(1).lower()
htype = mat.group(2)
if htype == "WAVELENGTH":
if band not in self._bands[sfile]:
self._bands[sfile].append(band)
self._wave[sfile][band] = \
hdus[h].data.astype(np.float64).copy()
if comm is not None:
self._bands[sfile] = comm.bcast(self._bands[sfile], root=0)
self._wave[sfile] = comm.bcast(self._wave[sfile], root=0)
if comm_rank == 0:
hdus.close()
# _alltargetids can have repeats from multiple files. Trim to
# unique set while retaining order in which they appeared
sortedidx = np.unique(self._alltargetids, return_index=True)[1]
ii = np.argsort(sortedidx)
unique_targetids = np.asarray(self._alltargetids)[sortedidx[ii]]
self._alltargetids = unique_targetids
self._keep_targets = unique_targetids.copy()
# Now we have the metadata for all targets in all files. Distribute
# the targets among process weighted by the amount of work to do for
# each target. This weight is either "1" if we are going to use coadds
# or the number of spectra if we are using all the data.
tweights = None
if not coadd:
tweights = dict()
for t in self._keep_targets:
tweights[t] = 0
for sfile in spectrafiles:
if t in self._target_specs[sfile]:
tweights[t] += len(self._target_specs[sfile][t])
self.capacities = capacities
if self.capacities is None:
self.is_lopsided = False
self._proc_targets = distribute_work(comm_size,
self._keep_targets, weights=tweights)
else:
self.is_lopsided = True
self._proc_targets = distribute_work(comm_size,
self._keep_targets, weights=tweights, capacities=self.capacities)
self._my_targets = self._proc_targets[comm_rank]
# Reverse mapping- target ID to index in our list
self._my_target_indx = {y : x for x, y in enumerate(self._my_targets)}
# Now every process has its local target IDs assigned. Pre-create our
# local target list with empty spectral data (except for wavelengths)
self._my_data = list()
for t in self._my_targets:
speclist = list()
for sfile in spectrafiles:
for b in self._bands[sfile]:
if t in self._target_specs[sfile]:
nspec = len(self._target_specs[sfile][t])
for s in range(nspec):
sindx = self._target_specs[sfile][t][s]
speclist.append(Spectrum(wave=self._wave[sfile][b],
flux=None, ivar=None, R=None, band=b))
self._my_data.append(Target(t, speclist, coadd=False))
# Iterate over the data and broadcast. Every process selects the rows
# of each table that contain pieces of local target data and copies it
# into place.
# these are for tracking offsets within the spectra for each target.
tspec_flux = { x : 0 for x in self._my_targets }
tspec_ivar = tspec_flux.copy()
tspec_mask = tspec_flux.copy()
tspec_res = tspec_flux.copy()
for sfile in spectrafiles:
rows = self._spec_keep[sfile]
if len(rows) == 0:
continue
hdus = None
if comm_rank == 0:
hdus = fits.open(sfile, memmap=False)
for b in self._bands[sfile]:
extname = "{}_{}".format(b.upper(), "FLUX")
hdata = None
badflux = None
if comm_rank == 0:
hdata = hdus[extname].data[rows]
# check for NaN and Inf here (should never happen of course)
badflux = np.isnan(hdata) | np.isinf(hdata) | np.isneginf(hdata)
hdata[badflux] = 0.0
if comm is not None:
hdata = comm.bcast(hdata, root=0)
badflux = comm.bcast(badflux, root=0)
toff = 0
for t in self._my_targets:
if t in self._target_specs[sfile]:
for trow in self._target_specs[sfile][t]:
self._my_data[toff].spectra[tspec_flux[t]].flux = \
hdata[trow].astype(np.float64).copy()
tspec_flux[t] += 1
toff += 1
extname = "{}_{}".format(b.upper(), "IVAR")
hdata = None
if comm_rank == 0:
hdata = hdus[extname].data[rows]
# check for NaN and Inf here (should never happen of course)
bad = np.isnan(hdata) | np.isinf(hdata) | np.isneginf(hdata)
hdata[bad] = 0.0
hdata[badflux] = 0.0 # also set ivar=0 to bad flux
if comm is not None:
hdata = comm.bcast(hdata, root=0)
toff = 0
for t in self._my_targets:
if t in self._target_specs[sfile]:
for trow in self._target_specs[sfile][t]:
self._my_data[toff].spectra[tspec_ivar[t]].ivar = \
hdata[trow].astype(np.float64).copy()
# mask significantly negative data (set ivar to 0)
sigma = self._my_data[toff].spectra[tspec_ivar[t]].flux * \
np.sqrt(self._my_data[toff].spectra[tspec_ivar[t]].ivar)
bignegflux = (sigma<-abs(negflux_nsig))
self._my_data[toff].spectra[tspec_ivar[t]].ivar[bignegflux] = 0.0
tspec_ivar[t] += 1
toff += 1
extname = "{}_{}".format(b.upper(), "MASK")
hdata = None
if comm_rank == 0:
if extname in hdus:
hdata = hdus[extname].data[rows]
if comm is not None:
hdata = comm.bcast(hdata, root=0)
if hdata is not None:
toff = 0
for t in self._my_targets:
if t in self._target_specs[sfile]:
for trow in self._target_specs[sfile][t]:
self._my_data[toff].spectra[tspec_mask[t]]\
.ivar *= (hdata[trow] == 0)
tspec_mask[t] += 1
toff += 1
extname = "{}_{}".format(b.upper(), "RESOLUTION")
hdata = None
if comm_rank == 0:
hdata = hdus[extname].data[rows]
if comm is not None:
hdata = comm.bcast(hdata, root=0)
toff = 0
for t in self._my_targets:
if t in self._target_specs[sfile]:
for trow in self._target_specs[sfile][t]:
dia = Resolution(hdata[trow].astype(np.float64))
self._my_data[toff].spectra[tspec_res[t]].R = dia
#- Coadds replace Rcsr so only compute if not coadding
if not coadd and cache_Rcsr:
self._my_data[toff].spectra[tspec_res[t]].Rcsr = dia.tocsr()
tspec_res[t] += 1
toff += 1
del hdata
if comm_rank == 0:
hdus.close()
# Compute the coadds now if we are going to use those
if coadd:
for t in self._my_data:
t.compute_coadd(cache_Rcsr,cosmics_nsig=self.cosmics_nsig)
self.fibermap = Table(np.hstack([ self._coadd_fmaps[x] \
for x in self._spectrafiles ]))
self.exp_fibermap = Table(np.hstack([ self._exp_fmaps[x] \
for x in self._spectrafiles ]))
self.tsnr2 = Table(np.hstack([ self._tsnr2[x] \
for x in self._spectrafiles ]))
super(DistTargetsDESI, self).__init__(self._keep_targets, comm=comm)
def _local_target_ids(self):
return self._my_targets
def _local_data(self):
return self._my_data
[docs]def rrdesi(options=None, comm=None):
"""Estimate redshifts for DESI targets.
This loads distributed DESI targets from one or more spectra grouping
files and computes the redshifts. The outputs are written to a redrock
scan file and a DESI redshift catalog.
Args:
options (list): optional list of commandline options to parse.
comm (mpi4py.Comm): MPI communicator to use.
"""
global_start = elapsed(None, "", comm=comm)
parser = argparse.ArgumentParser(description="Estimate redshifts from"
" DESI target spectra.")
parser.add_argument("-t", "--templates", type=str, default=None,
required=False, help="template file or directory")
parser.add_argument("--archetypes", type=str, default=None,
required=False,
help="archetype file or directory for final redshift comparison")
parser.add_argument("--archetype-legendre-degree", type=int, default=2,
required=False, help="if archetypes are provided legendre polynomials upto deg_legendre-1 will be used (default is 2)")
parser.add_argument("--archetype-nnearest", type=int, default=None,
required=False, help="number of nearest archetypes (in chi2 space) to be used in archetype modeling, must be greater than 1 (default is None)")
parser.add_argument("--archetype-legendre-percamera", default=True, action="store_true",
required=False, help="If True, in archetype mode the fitting will be done for each camera/band")
parser.add_argument("--archetype-legendre-prior", type=float, default=0.1,
required=False, help="sigma to add as prior in solving linear equation, 1/sig**2 will be added, default is 0.1, Note: provide anything <=0 if you do not want to add any prior")
parser.add_argument("--archetypes-no-legendre", default=False, action="store_true",
required=False, help="Use this flag with archetypes if want to TURN OFF all archetype related default values")
parser.add_argument("--zminfit-npoints", type=int, default=15,
required=False, help="number of finer redshift to be used around best fit redshifts (default is 15)")
parser.add_argument("-d", "--details", type=str, default=None,
required=False, help="output file for full redrock fit details")
parser.add_argument("-o", "--outfile", type=str, default=None,
required=False, help="output FITS file with best redshift per target")
parser.add_argument("--model", type=str, default=None,
required=False, help="output FITS file containing spectral model")
parser.add_argument("--targetids", type=str, default=None,
required=False, help="comma-separated list of target IDs")
parser.add_argument("--mintarget", type=int, default=None,
required=False, help="first target to process in each file")
parser.add_argument("--priors", type=str, default=None,
required=False, help="optional redshift prior file")
parser.add_argument("--chi2-scan", type=str, default=None,
required=False, help="Load the chi2-scan from the input file")
parser.add_argument("--zscan-galaxy", type=str, default='-0.005,1.7,3e-4',
required=False, help="Redshift scan parameters for galaxies: zmin,zmax,dz")
parser.add_argument("--zscan-qso", type=str, default='0.05,6.0,5e-4',
required=False, help="Redshift scan parameters for QSO: zmin,zmax,dz")
parser.add_argument("--zscan-star", type=str, default='-0.002,0.00201,4e-5',
required=False, help="Redshift scan parameters for stars: zmin,zmax,dz")
parser.add_argument("-n", "--ntargets", type=int,
required=False, help="the number of targets to process in each file")
parser.add_argument("--nminima", type=int, default=None,
required=False, help="the number of redshift minima to search")
parser.add_argument("--allspec", default=False, action="store_true",
required=False, help="use individual spectra instead of coadd")
parser.add_argument("--ncpu", type=int, default=None,
required=False, help="DEPRECATED: the number of multiprocessing"
" processes; use --mp instead")
parser.add_argument("--mp", type=int, default=0,
required=False, help="if not using MPI, the number of multiprocessing"
" processes to use (defaults to half of the hardware threads)")
parser.add_argument("--no-skymask", default=False, action="store_true",
required=False, help="Do not do extra masking of sky lines")
parser.add_argument("--no-mpi-abort", default=False, action="store_true",
required=False, help="Do not call MPI Abort upon failure of a single rank")
parser.add_argument("--debug", default=False, action="store_true",
required=False, help="debug with ipython (only if communicator has a "
"single process)")
parser.add_argument("--cosmics-nsig", type=float, default=0,
required=False, help="n sigma cosmic ray threshold in coaddition")
parser.add_argument("--negflux-nsig", type=float, default=5,
required=False, help="n sigma negative flux mask threshold")
parser.add_argument("-i", "--infiles", nargs='+', required=True,
help="Input spectra, coadd, or cframe files")
parser.add_argument("--gpu", action="store_true",
required=False, help="use GPUs")
parser.add_argument("--max-gpuprocs", type=int, default=None,
required=False, help="limit number of MPI processes using GPUs")
args = None
if options is None:
args = parser.parse_args()
else:
args = parser.parse_args(options)
if args.ncpu is not None:
print('WARNING: --ncpu is deprecated; use --mp instead')
args.mp = args.ncpu
comm_size = 1
comm_rank = 0
if comm is not None:
comm_size = comm.size
comm_rank = comm.rank
# Check arguments- all processes have this, so just check on the first
# process
# if user sets --nminima, use that
if args.nminima is not None:
nminima = args.nminima
# otherwise use different defaults for archetypes vs. non-archetypes
elif args.archetypes is None:
nminima = 3 # default for non-archetypes
else:
nminima = 9 # default for archetypes
if comm_rank == 0:
if args.debug and comm_size != 1:
print("--debug can only be used if the communicator has one "
" process")
sys.stdout.flush()
if comm is not None:
comm.Abort()
if (args.details is None) and (args.outfile is None):
parser.print_help()
print("ERROR: --details or --outfile required")
sys.stdout.flush()
if comm is not None:
comm.Abort()
else:
sys.exit(1)
if len(args.infiles) == 0:
print("ERROR: must provide input files")
sys.stdout.flush()
if comm is not None:
comm.Abort()
else:
sys.exit(1)
if (args.targetids is not None) and ((args.mintarget is not None) \
or (args.ntargets is not None)):
print("ERROR: cannot select targets by both ID and range")
sys.stdout.flush()
if comm is not None:
comm.Abort()
else:
sys.exit(1)
if args.gpu:
try:
import cupy
gpu_ok = cupy.is_available()
except ImportError:
gpu_ok = False
if not gpu_ok:
print("ERROR: cupy or GPU not available")
sys.stdout.flush()
if comm is not None:
comm.Abort()
else:
sys.exit(1)
if args.archetypes is not None:
print('\n===== Archetype argument is provided, doing necessary checks=======\n')
if os.path.exists(args.archetypes) and os.access(args.archetypes, os.R_OK):
if os.path.isdir(args.archetypes):
if os.listdir(args.archetypes):
print('Archetype directory exists and readable and it is not empty..')
print('Archetype will be applied to all spectype')
else:
print('ERROR: Archetype directory is empty')
sys.stdout.flush()
if comm is not None:
comm.Abort()
else:
sys.exit(1)
if os.path.isfile(args.archetypes):
print('Archetype is a file and it exists and readable')
print('Archetype will only be applied to SPECTYPE %s'%(os.path.basename(args.archetypes).split('-')[1].split('.')[0].upper()))
else:
print("ERROR: can't find archetypes_dir or it is unreadable")
sys.stdout.flush()
if comm is not None:
comm.Abort()
else:
sys.exit(1)
#- All ranks read the archetypes file(s)
archetypes = None
if args.archetypes is not None:
archetypes = All_archetypes(archetypes_dir=args.archetypes, verbose=(comm_rank==0)).archetypes
targetids = None
if args.targetids is not None:
targetids = [ int(x) for x in args.targetids.split(",") ]
n_target = None
if args.ntargets is not None:
n_target = args.ntargets
first_target = None
if args.mintarget is not None:
first_target = args.mintarget
elif n_target is not None:
first_target = 0
# Multiprocessing processes to use if MPI is disabled.
mpprocs = 0
if comm is None:
mpprocs = get_mp(args.mp)
print("Running with {} processes".format(mpprocs))
if "OMP_NUM_THREADS" in os.environ:
nthread = int(os.environ["OMP_NUM_THREADS"])
if nthread != 1:
print("WARNING: {} multiprocesses running, each with "
"{} threads ({} total)".format(mpprocs, nthread,
mpprocs*nthread))
print("WARNING: Please ensure this is <= the number of "
"physical cores on the system")
else:
print("WARNING: using multiprocessing, but the OMP_NUM_THREADS")
print("WARNING: environment variable is not set- your system may")
print("WARNING: be oversubscribed.")
sys.stdout.flush()
elif comm_rank == 0:
print("Running with {} processes".format(comm_size))
sys.stdout.flush()
# GPU configuration
if args.gpu:
# Determine which processes will use a GPU
if args.max_gpuprocs is not None:
max_gpuprocs = args.max_gpuprocs
else:
#Check actual number of GPUs available
if (comm is not None):
#Use custom method that checks PCI ids for MPI
max_gpuprocs = getGPUCountMPI(comm)
else:
#cupy getDeviceCount works for non MPI
import cupy
max_gpuprocs = cupy.cuda.runtime.getDeviceCount()
use_gpu = comm_rank < max_gpuprocs
# Determine cpu/gpu process capacities for target distribution
if comm is not None:
gpu_proc_flags = comm.allgather(use_gpu)
else:
gpu_proc_flags = [use_gpu, ]
if (mpprocs > 1):
#Force mpprocs == 1 for multiprocessing mode with GPU
print("WARNING: using GPU mode without MPI requires --mp 1")
print("WARNING: Overriding {} multiprocesses to force this.".format(mpprocs))
print("WARNING: Running with 1 process.")
mpprocs = 1
ngpu_procs = sum(gpu_proc_flags)
ncpu_procs = comm_size - ngpu_procs
if ngpu_procs > 0 and ncpu_procs > 0:
# On Perlmutter, 1:15 seems like a good ratio
#capacities = [1 if is_gpu_proc else 1.0/15 for is_gpu_proc in gpu_proc_flags]
#With new GPU implementation of zscan, use 1:10000 so that only GPU-enabled
#procs get allocated targets
capacities = [1 if is_gpu_proc else 1.0/10000 for is_gpu_proc in gpu_proc_flags]
else:
capacities = None
# Redistribute templates after rebinning when using GPUs
redistribute_templates = True
else:
use_gpu = False
capacities = None
redistribute_templates = False
try:
# Load and distribute the targets
if comm_rank == 0:
print("Loading targets...")
sys.stdout.flush()
start = elapsed(None, "", comm=comm)
# Load the targets. If comm is None, then the target data will be
# stored in shared memory.
targets = DistTargetsDESI(args.infiles, coadd=(not args.allspec),
targetids=targetids, first_target=first_target, n_target=n_target,
comm=comm, cache_Rcsr=True, cosmics_nsig=args.cosmics_nsig,
negflux_nsig=abs(args.negflux_nsig),
capacities=capacities)
#- Mask some problematic sky lines
if not args.no_skymask:
for t in targets.local():
for s in t.spectra:
ii = (5572. <= s.wave) & (s.wave <= 5582.)
ii |= (9792. <= s.wave) & (s.wave <= 9795.)
s.ivar[ii] = 0.0
# Get the dictionary of wavelength grids (with keys as wavehashes)
dwave = targets.wavegrids()
ncamera = len(list(dwave.keys())) # number of cameras for given instrument
if args.archetypes_no_legendre or args.archetypes is None:
if comm_rank == 0:
print('no archetypes or --archetypes-no-legendre; will turn off all the Legendre related arguments')
archetype_legendre_prior = None
archetype_legendre_degree =0
archetype_legendre_percamera = False
else:
if comm_rank == 0 and args.archetypes is not None:
print('Will be using default archetype values.')
print('Number of minimum redshift for which archetype redshift fitting will be done = %d'%(nminima))
if ncamera>=1:
archetype_legendre_percamera = True
if comm_rank == 0 and args.archetypes is not None:
print('Number of cameras = %d, percamera fitting will be done'%(ncamera))
else:
archetype_legendre_percamera = False
if comm_rank == 0 and args.archetypes is not None:
print('No per camera fitting will be done.')
if args.archetype_legendre_degree>0:
if comm_rank == 0 and args.archetypes is not None:
print('legendre polynomials of degrees %s will be added to Archetypes'%([i for i in range(args.archetype_legendre_degree)]))
else:
if comm_rank == 0 and args.archetypes is not None:
print('No Legendre polynomial will be added to archetypes.')
if args.archetype_legendre_prior<=0:
archetype_legendre_prior = None
printstr = 'no'
else:
archetype_legendre_prior = args.archetype_legendre_prior
printstr = 'a'
if comm_rank == 0 and args.archetypes is not None:
if archetype_legendre_prior is None:
print('A zero or negative prior is provided so setting archetype_legendre_prior = None')
print('archetype_legendre_prior = %s has been provided, so %s prior will be added while solving for the coefficients'%(archetype_legendre_prior, printstr))
if args.archetype_nnearest is not None:
if comm_rank == 0 and args.archetypes is not None:
print('nearest neighbour = %d is provided, will do the final fitting for N-best nearest archetypes in chi2 space..'%(args.archetype_nnearest))
stop = elapsed(start, "Read and distribution of {} targets"\
.format(len(targets.all_target_ids)), comm=comm)
# Read the template data
# Pass both use_gpu (this proc) and args.gpu (if any proc is using GPU)
dtemplates = load_dist_templates(dwave, templates=args.templates,
zscan_galaxy=args.zscan_galaxy,
zscan_qso=args.zscan_qso,
zscan_star=args.zscan_star,
comm=comm, mp_procs=mpprocs,
redistribute=redistribute_templates,
use_gpu=use_gpu, gpu_mode=args.gpu)
# Compute the redshifts, including both the coarse scan and the
# refinement. This function only returns data on the rank 0 process.
start = elapsed(None, "", comm=comm)
scandata, zfit = zfind(targets, dtemplates, mpprocs,
nminima=nminima, archetypes=archetypes,
priors=args.priors, chi2_scan=args.chi2_scan, use_gpu=use_gpu, zminfit_npoints=args.zminfit_npoints, per_camera=archetype_legendre_percamera, deg_legendre=args.archetype_legendre_degree, n_nearest=args.archetype_nnearest, prior_sigma=archetype_legendre_prior, ncamera=ncamera)
stop = elapsed(start, "Computing redshifts", comm=comm)
zbest = None
# Set some DESI-specific ZWARN bits from input fibermap
if comm_rank == 0:
fiberstatus = targets.fibermap['COADD_FIBERSTATUS']
poorpos = (fiberstatus & fibermask.POORPOSITION) != 0
badpos = (fiberstatus & fibermask.BADPOSITION) != 0
broken = (fiberstatus & fibermask.BROKENFIBER) != 0
unassigned = (fiberstatus & fibermask.UNASSIGNED) != 0
bad = targets.fibermap['OBJTYPE'] == 'BAD'
sky = targets.fibermap['OBJTYPE'] == 'SKY'
badcoverage = np.zeros(len(fiberstatus), dtype=bool)
for key in ('BADCOLUMN', 'BADAMPB', 'BADAMPR', 'BADAMPZ'):
if key in fibermask.names():
badcoverage |= (fiberstatus & fibermask.mask(key)) != 0
targetids = targets.fibermap['TARGETID']
ii = np.isin(zfit['targetid'], targetids[poorpos])
zfit['zwarn'][ii] |= ZWarningMask.POORDATA
ii = np.isin(zfit['targetid'], targetids[badpos | broken | unassigned | bad])
zfit['zwarn'][ii] |= ZWarningMask.NODATA
# ADM set NODATA cases to z=0, GALAXY with zero coefficients.
zfit['spectype'][ii] = 'GALAXY'
zfit['subtype'][ii] = ''
zfit['coeff'][ii] = 0.
zfit['fitmethod'][ii] = 'NONE'
ii = np.isin(zfit['targetid'], targetids[broken])
zfit['zwarn'][ii] |= ZWarningMask.UNPLUGGED
ii = np.isin(zfit['targetid'], targetids[sky])
zfit['zwarn'][ii] |= ZWarningMask.SKY
ii = np.isin(zfit['targetid'], targetids[badcoverage])
zfit['zwarn'][ii] |= ZWarningMask.LITTLE_COVERAGE
# gather templates and archetypes for final I/O header metadata
templates = dict()
for dt in dtemplates:
spectype = dt.template.template_type
subtype = dt.template.sub_type
templates[(spectype,subtype)] = dt.template
archetypes = None
if not args.archetypes is None:
archetypes = All_archetypes(archetypes_dir=args.archetypes,
verbose=(comm_rank==0)).archetypes
# Write the outputs
if args.details is not None:
start = elapsed(None, "", comm=comm)
if comm_rank == 0:
write_zscan(args.details, scandata, zfit, clobber=True)
stop = elapsed(start, "Writing zscan data took", comm=comm)
if args.outfile:
start = elapsed(None, "", comm=comm)
if comm_rank == 0:
zbest = zfit[zfit['znum'] == 0]
# Remove extra columns not needed for zbest
zbest.remove_columns(['zz', 'zzchi2', 'znum'])
# Change to upper case like SDSS / DESI
for colname in zbest.colnames:
if colname.islower():
zbest.rename_column(colname, colname.upper())
# Allow 4 char for ARCH (vs. PCA/NMF) even if archetypes aren't used
zbest['FITMETHOD'] = zbest['FITMETHOD'].astype('S4')
write_zbest(args.outfile, zbest,
targets.fibermap, targets.exp_fibermap,
targets.tsnr2,
templates, archetypes=archetypes,
spec_header=targets.header0)
if comm is not None:
zbest = comm.bcast(zbest, root=0)
stop = elapsed(start, f"Writing {args.outfile} took", comm=comm)
#- Evaluate models
if args.model is not None:
#- Evaluate best-fit models for local targets,
#- then collect into all_models on rank 0
local_models = targets.eval_models(zbest, templates, archetypes)
local_targetids = targets.local_target_ids()
if targets.comm is not None:
all_models = targets.comm.gather(local_models, root=0)
all_targetids = targets.comm.gather(local_targetids, root=0)
else:
all_models = [local_models,]
all_targetids = [local_targetids,]
stop = elapsed(start, f"Evaluating best fit models took", comm=comm)
if comm_rank == 0:
#- collapse list of lists of dictionaries -> single list of dictionaries
all_models = list(itertools.chain.from_iterable(all_models))
#- find sort orrder to match zbest
all_targetids = np.concatenate(all_targetids)
xsorted = np.argsort(all_targetids)
ypos = np.searchsorted(all_targetids[xsorted], zbest['TARGETID'])
sort_indices = xsorted[ypos]
#- double check sorting
all_targetids = np.array(all_targetids)[sort_indices]
np.testing.assert_array_equal(zbest['TARGETID'].data, all_targetids)
#- convert list[targets] of dict[bands] into dict[band] of sorted array[targets]
#- for DESI, the wavegrid keys = wavehashes = camera bands b/r/z
all_models_dict = dict()
wave_dict = targets.wavegrids()
for band in wave_dict:
all_models_dict[band] = np.vstack([m[band] for m in all_models])[sort_indices]
write_bestmodel(args.model, zbest, all_models_dict, wave_dict,
templates, archetypes, spec_header=targets.header0)
stop = elapsed(start, f"Writing {args.model} took", comm=comm)
except Exception as err:
exc_type, exc_value, exc_traceback = sys.exc_info()
lines = traceback.format_exception(exc_type, exc_value, exc_traceback)
lines = [ "Proc {}: {}".format(comm_rank, x) for x in lines ]
print("--- Process {} raised an exception ---".format(comm_rank))
print("".join(lines))
sys.stdout.flush()
if comm is None or args.no_mpi_abort:
raise err
else:
comm.Abort()
global_stop = elapsed(global_start, "Total run time", comm=comm)
if args.debug:
import IPython
IPython.embed()
return