def make_coinc_branch(dag, datafinds, seglistdict, time_slides, timing_params, psds_per_power, enable_clustering, tag, do_injections = False, verbose = False): # injection list if do_injections: assert len(time_slides) == 1 if verbose: print >>sys.stderr, "Building lalapps_binj jobs ..." binjnodes = power.make_binj_fragment(dag, seglistdict.extent_all(), time_slides.keys()[0], tag, 0.0, float(power.powerjob.get_opts()["low-freq-cutoff"]), float(power.powerjob.get_opts()["low-freq-cutoff"]) + float(power.powerjob.get_opts()["bandwidth"])) # add binj nodes as parents of the datafinds to force the binj's to # be run first. this ensures that once a datafind has run the # power jobs that follow it will immediately be able to run, which # helps depth-first dagman do smarter things. for node in datafinds: for binjnode in binjnodes: node.add_parent(binjnode) else: binjnodes = set() # single-instrument trigger generation trigger_nodes = power.make_single_instrument_stage(dag, datafinds, seglistdict, tag, timing_params, psds_per_power, binjnodes = binjnodes, verbose = verbose) if enable_clustering: if verbose: print >>sys.stderr, "building pre-lladd bucluster jobs ..." trigger_nodes = power.make_bucluster_fragment(dag, trigger_nodes, "PRELLADD_%s" % tag, verbose = verbose) # coincidence analysis coinc_nodes = set() binj_cache = set([cache_entry for node in binjnodes for cache_entry in node.get_output_cache()]) # otherwise too many copies of the offset vector will be fed into # burca assert len(binj_cache) < 2 for n, (time_slides_cache_entry, these_time_slides) in enumerate(time_slides.items()): if verbose: print >>sys.stderr, "%s %d/%d (%s):" % (tag, n + 1, len(time_slides), time_slides_cache_entry.path) tisi_cache = set([time_slides_cache_entry]) if do_injections: # lalapps_binj has already copied the time slide # document into its own output extra_input_cache = set() else: # ligolw_add needs to copy the time slide document # into is output extra_input_cache = tisi_cache nodes = set() for seg, parents, cache, clipseg in power.group_coinc_parents(trigger_nodes, these_time_slides, verbose = verbose): nodes |= power.make_lladd_fragment(dag, parents | binjnodes, "%s_%d" % (tag, n), segment = seg, input_cache = cache | binj_cache, extra_input_cache = extra_input_cache, remove_input = do_injections, preserve_cache = binj_cache | tisi_cache) if enable_clustering: if verbose: print >>sys.stderr, "building post-lladd bucluster jobs ..." nodes = power.make_bucluster_fragment(dag, nodes, "POSTLLADD_%s_%d" % (tag, n), verbose = verbose) if verbose: print >>sys.stderr, "building burca jobs ..." coinc_nodes |= power.make_burca_fragment(dag, nodes, "%s_%d" % (tag, n), verbose = verbose) if verbose: print >>sys.stderr, "done %s %d/%d" % (tag, n + 1, len(time_slides)) # injection identification if do_injections: if verbose: print >>sys.stderr, "building binjfind jobs ..." coinc_nodes = power.make_binjfind_fragment(dag, coinc_nodes, tag, verbose = verbose) # conversion to SQLite database files if verbose: print >>sys.stderr, "building sqlite jobs ..." coinc_nodes = power.make_sqlite_fragment(dag, coinc_nodes, tag, verbose = verbose) # done power.write_output_cache(coinc_nodes, "%s_%s_output.cache" % (os.path.splitext(dag.get_dag_file())[0], tag)) return coinc_nodes
def make_coinc_branch(dag, datafinds, seglists, time_slides, min_segment_length, pad, overlap, short_segment_duration, tag, vetoes_cache = set(), do_injections = False, injections_offset = 0.0, verbose = False): # # injection job # binjnodes = set() if do_injections: # don't know what to do with more than one list of offset # vectors assert len(time_slides) == 1 # get the largest injection offset's magnitude maxoffset = max(abs(offset) for offsetvectorlist in time_slides.values() for offsetvector in offsetvectorlist for offset in offsetvector.values()) # to save disk space and speed the dag along we don't # generate a single injection list for the entire analysis # run, instead a separate list is constructed for each # block of data to be analyzed. we need to be careful that # two nearby injection lists don't contain injections for # the same time, so we protract the segments by the time # step and coalesce so that only gaps between segments # larger than twice the time step result in separate files # being generated. we could allow smaller gaps to survive, # but this way we don't have to worry about it. # injections_offset is a number between 0 and 1 in units of # the period between injections for seg in seglists.union(seglists).protract(power.binjjob.time_step + maxoffset).coalesce().contract(power.binjjob.time_step + maxoffset): binjnodes |= power.make_binj_fragment(dag, seg.protract(maxoffset), time_slides.keys()[0], tag, offset = injections_offset) # artificial parent-child relationship to induce dagman to # submit binj jobs as the corresponding datafinds complete # instead of submiting all of one kind before any of the next. # makes dag run faster because it allows string search jobs to # start moving onto the cluster without waiting for all the # datafinds and/or all the binjs to complete for datafindnode in datafinds: seg = segments.segment(datafindnode.get_start(), datafindnode.get_end()) for binjnode in binjnodes: if seg.intersects(power.cache_span(binjnode.get_output_cache())): binjnode.add_parent(datafindnode) # # trigger generator jobs # # set max job length to ~3600 s (will be clipped to an allowed # size) trigger_nodes = cosmicstring.make_single_instrument_stage(dag, datafinds, seglists, tag, min_segment_length, pad, overlap, short_segment_duration, max_job_length = 3600, binjnodes = binjnodes, verbose = verbose) # # coincidence analysis # coinc_nodes = [] for n, (time_slides_cache_entry, these_time_slides) in enumerate(time_slides.items()): if verbose: print >>sys.stderr, "%s %d/%d (%s):" % (tag, n + 1, len(time_slides), time_slides_cache_entry.path) coinc_nodes.append(set()) # # lalapps_cafe & ligolw_add # tisi_cache = set([time_slides_cache_entry]) lladd_nodes = set() for seg, parents, cache, clipseg in power.group_coinc_parents(trigger_nodes, these_time_slides, extentlimit = 50000000.0 / (len(these_time_slides) or 1), verbose = verbose): binj_cache = set(cache_entry for node in binjnodes for cache_entry in node.get_output_cache() if cache_entry.segment.intersects(seg)) # otherwise too many copies of the offset vector # will be fed into burca assert len(binj_cache) < 2 if do_injections: # lalapps_binj has already copied the time # slide document into its own output extra_input_cache = vetoes_cache else: # ligolw_add needs to copy the time slide # document into its output extra_input_cache = tisi_cache | vetoes_cache these_lladd_nodes = power.make_lladd_fragment(dag, parents | binjnodes, "%s_%d" % (tag, n), segment = seg, input_cache = cache | binj_cache, extra_input_cache = extra_input_cache, remove_input = do_injections and clipseg is not None, preserve_cache = binj_cache | tisi_cache | vetoes_cache) if clipseg is not None: # # this is a fragment of a too-large burca # job, construct it specially and add the # command-line option needed to clip the # output # assert len(these_lladd_nodes) == 1 coinc_nodes[-1] |= power.make_burca_fragment(dag, these_lladd_nodes, "%s_%d" % (tag, n), coincidence_segments = segments.segmentlist([clipseg]), verbose = verbose) else: # # this is not a fragment of a too-large # burca job, add it to the pool of files to # be processed by the burcas that don't # require special clipping command line # options # lladd_nodes |= these_lladd_nodes # # lalapps_burca pool. these are the burca jobs that don't # require special clipping command line options, and so can # bulk-process many files with each job # if verbose: print >>sys.stderr, "building burca jobs ..." coinc_nodes[-1] |= power.make_burca_fragment(dag, lladd_nodes, "%s_%d" % (tag, n), verbose = verbose) if verbose: print >>sys.stderr, "done %s %d/%d" % (tag, n + 1, len(time_slides)) # # lalapps_binjfind # if do_injections: if verbose: print >>sys.stderr, "building binjfind jobs ..." coinc_nodes = [power.make_binjfind_fragment(dag, these_coinc_nodes, "%s_%d" % (tag, n), verbose = verbose) for n, these_coinc_nodes in enumerate(coinc_nodes)] # # ligolw_sqlite and lalapps_run_sqlite # if verbose: print >>sys.stderr, "building sqlite jobs ..." coinc_nodes = [power.make_sqlite_fragment(dag, these_coinc_nodes, "%s_%d" % (tag, n), verbose = verbose) for n, these_coinc_nodes in enumerate(coinc_nodes)] coinc_nodes = [cosmicstring.make_run_sqlite_fragment(dag, these_coinc_nodes, "%s_%d" % (tag, n), clipsegments_sql_filename) for n, these_coinc_nodes in enumerate(coinc_nodes)] # # lalapps_string_meas_likelihood # if verbose: print >>sys.stderr, "building lalapps_string_meas_likelihood jobs ..." likelihood_nodes = [cosmicstring.make_meas_likelihood_fragment(dag, these_coinc_nodes, "%s_%d" % (tag, n)) for n, these_coinc_nodes in enumerate(coinc_nodes)] # # write output cache # if verbose: print >>sys.stderr, "writing output cache ..." for n, (these_coinc_nodes, these_likelihood_nodes) in enumerate(zip(coinc_nodes, likelihood_nodes)): power.write_output_cache(these_coinc_nodes | these_likelihood_nodes, "%s_%s_output.cache" % (os.path.splitext(dag.get_dag_file())[0], "%s_%d" % (tag, n))) # # done # return coinc_nodes, likelihood_nodes
def make_coinc_branch(dag,
datafinds,
seglistdict,
time_slides,
timing_params,
psds_per_power,
enable_clustering,
tag,
do_injections=False,
verbose=False):
# injection list
if do_injections:
assert len(time_slides) == 1
if verbose:
print >> sys.stderr, "Building lalapps_binj jobs ..."
binjnodes = power.make_binj_fragment(
dag, seglistdict.extent_all(),
time_slides.keys()[0], tag, 0.0,
float(power.powerjob.get_opts()["low-freq-cutoff"]),
float(power.powerjob.get_opts()["low-freq-cutoff"]) +
float(power.powerjob.get_opts()["bandwidth"]))
# add binj nodes as parents of the datafinds to force the binj's to
# be run first. this ensures that once a datafind has run the
# power jobs that follow it will immediately be able to run, which
# helps depth-first dagman do smarter things.
for node in datafinds:
for binjnode in binjnodes:
node.add_parent(binjnode)
else:
binjnodes = set()
# single-instrument trigger generation
trigger_nodes = power.make_single_instrument_stage(dag,
datafinds,
seglistdict,
tag,
timing_params,
psds_per_power,
binjnodes=binjnodes,
verbose=verbose)
if enable_clustering:
if verbose:
print >> sys.stderr, "building pre-lladd bucluster jobs ..."
trigger_nodes = power.make_bucluster_fragment(dag,
trigger_nodes,
"PRELLADD_%s" % tag,
verbose=verbose)
# coincidence analysis
coinc_nodes = set()
binj_cache = set([
cache_entry for node in binjnodes
for cache_entry in node.get_output_cache()
])
# otherwise too many copies of the offset vector will be fed into
# burca
assert len(binj_cache) < 2
for n, (time_slides_cache_entry,
these_time_slides) in enumerate(time_slides.items()):
if verbose:
print >> sys.stderr, "%s %d/%d (%s):" % (
tag, n + 1, len(time_slides), time_slides_cache_entry.path)
tisi_cache = set([time_slides_cache_entry])
if do_injections:
# lalapps_binj has already copied the time slide
# document into its own output
extra_input_cache = set()
else:
# ligolw_add needs to copy the time slide document
# into is output
extra_input_cache = tisi_cache
nodes = set()
for seg, parents, cache, clipseg in power.group_coinc_parents(
trigger_nodes, these_time_slides, verbose=verbose):
nodes |= power.make_lladd_fragment(
dag,
parents | binjnodes,
"%s_%d" % (tag, n),
segment=seg,
input_cache=cache | binj_cache,
extra_input_cache=extra_input_cache,
remove_input=do_injections,
preserve_cache=binj_cache | tisi_cache)
if enable_clustering:
if verbose:
print >> sys.stderr, "building post-lladd bucluster jobs ..."
nodes = power.make_bucluster_fragment(dag,
nodes,
"POSTLLADD_%s_%d" % (tag, n),
verbose=verbose)
if verbose:
print >> sys.stderr, "building burca jobs ..."
coinc_nodes |= power.make_burca_fragment(dag,
nodes,
"%s_%d" % (tag, n),
verbose=verbose)
if verbose:
print >> sys.stderr, "done %s %d/%d" % (tag, n + 1,
len(time_slides))
# injection identification
if do_injections:
if verbose:
print >> sys.stderr, "building binjfind jobs ..."
coinc_nodes = power.make_binjfind_fragment(dag,
coinc_nodes,
tag,
verbose=verbose)
# conversion to SQLite database files
if verbose:
print >> sys.stderr, "building sqlite jobs ..."
coinc_nodes = power.make_sqlite_fragment(dag,
coinc_nodes,
tag,
verbose=verbose)
# done
power.write_output_cache(
coinc_nodes,
"%s_%s_output.cache" % (os.path.splitext(dag.get_dag_file())[0], tag))
return coinc_nodes
def make_coinc_branch(dag,
datafinds,
seglists,
time_slides,
min_segment_length,
pad,
overlap,
short_segment_duration,
tag,
vetoes_cache=set(),
do_injections=False,
injections_offset=0.0,
verbose=False):
#
# injection job
#
binjnodes = set()
if do_injections:
# don't know what to do with more than one list of offset
# vectors
assert len(time_slides) == 1
# get the largest injection offset's magnitude
maxoffset = max(
abs(offset) for offsetvectorlist in time_slides.values()
for offsetvector in offsetvectorlist
for offset in offsetvector.values())
# to save disk space and speed the dag along we don't
# generate a single injection list for the entire analysis
# run, instead a separate list is constructed for each
# block of data to be analyzed. we need to be careful that
# two nearby injection lists don't contain injections for
# the same time, so we protract the segments by the time
# step and coalesce so that only gaps between segments
# larger than twice the time step result in separate files
# being generated. we could allow smaller gaps to survive,
# but this way we don't have to worry about it.
# injections_offset is a number between 0 and 1 in units of
# the period between injections
for seg in seglists.union(seglists).protract(
power.binjjob.time_step +
maxoffset).coalesce().contract(power.binjjob.time_step +
maxoffset):
binjnodes |= power.make_binj_fragment(dag,
seg.protract(maxoffset),
time_slides.keys()[0],
tag,
offset=injections_offset)
# artificial parent-child relationship to induce dagman to
# submit binj jobs as the corresponding datafinds complete
# instead of submiting all of one kind before any of the next.
# makes dag run faster because it allows string search jobs to
# start moving onto the cluster without waiting for all the
# datafinds and/or all the binjs to complete
for datafindnode in datafinds:
seg = segments.segment(datafindnode.get_start(),
datafindnode.get_end())
for binjnode in binjnodes:
if seg.intersects(power.cache_span(
binjnode.get_output_cache())):
binjnode.add_parent(datafindnode)
#
# trigger generator jobs
#
# set max job length to ~3600 s (will be clipped to an allowed
# size)
trigger_nodes = cosmicstring.make_single_instrument_stage(
dag,
datafinds,
seglists,
tag,
min_segment_length,
pad,
overlap,
short_segment_duration,
max_job_length=3600,
binjnodes=binjnodes,
verbose=verbose)
#
# coincidence analysis
#
coinc_nodes = []
for n, (time_slides_cache_entry,
these_time_slides) in enumerate(time_slides.items()):
if verbose:
print("%s %d/%d (%s):" %
(tag, n + 1, len(time_slides), time_slides_cache_entry.path),
file=sys.stderr)
coinc_nodes.append(set())
#
# lalapps_cafe & ligolw_add
#
tisi_cache = set([time_slides_cache_entry])
lladd_nodes = set()
for segnum, (seg, parents, cache, clipseg) in enumerate(
power.group_coinc_parents(trigger_nodes,
these_time_slides,
extentlimit=150000000.0 /
(len(these_time_slides) or 1),
verbose=verbose)):
binj_cache = set(cache_entry for node in binjnodes
for cache_entry in node.get_output_cache()
if cache_entry.segment.intersects(seg))
# otherwise too many copies of the offset vector
# will be fed into burca
assert len(binj_cache) < 2
if do_injections:
# lalapps_binj has already copied the time
# slide document into its own output
extra_input_cache = vetoes_cache
else:
# ligolw_add needs to copy the time slide
# document into its output
extra_input_cache = tisi_cache | vetoes_cache
these_lladd_nodes = power.make_lladd_fragment(
dag,
parents | binjnodes,
"%s_%d_%x" % (tag, n, segnum),
segment=seg,
input_cache=cache | binj_cache | segments_cache,
extra_input_cache=extra_input_cache,
remove_input=do_injections and clipseg is not None,
preserve_cache=binj_cache | segments_cache | tisi_cache
| vetoes_cache)
if clipseg is not None:
#
# this is a fragment of a too-large burca
# job, construct it specially and add the
# command-line option needed to clip the
# output
#
assert len(these_lladd_nodes) == 1
coinc_nodes[-1] |= power.make_burca_fragment(
dag,
these_lladd_nodes,
"%s_%d" % (tag, n),
coincidence_segments=segments.segmentlist([clipseg]),
verbose=verbose)
else:
#
# this is not a fragment of a too-large
# burca job, add it to the pool of files to
# be processed by the burcas that don't
# require special clipping command line
# options
#
lladd_nodes |= these_lladd_nodes
#
# lalapps_burca pool. these are the burca jobs that don't
# require special clipping command line options, and so can
# bulk-process many files with each job
#
if verbose:
print("building burca jobs ...", file=sys.stderr)
coinc_nodes[-1] |= power.make_burca_fragment(dag,
lladd_nodes,
"%s_%d" % (tag, n),
verbose=verbose)
if verbose:
print("done %s %d/%d" % (tag, n + 1, len(time_slides)),
file=sys.stderr)
#
# lalapps_binjfind
#
if do_injections:
if verbose:
print("building binjfind jobs ...", file=sys.stderr)
coinc_nodes = [
power.make_binjfind_fragment(dag,
these_coinc_nodes,
"%s_%d" % (tag, n),
verbose=verbose)
for n, these_coinc_nodes in enumerate(coinc_nodes)
]
#
# ligolw_sqlite and lalapps_run_sqlite
#
if verbose:
print("building sqlite jobs ...", file=sys.stderr)
coinc_nodes = [
power.make_sqlite_fragment(dag,
these_coinc_nodes,
"%s_%d" % (tag, n),
verbose=verbose)
for n, these_coinc_nodes in enumerate(coinc_nodes)
]
coinc_nodes = [
cosmicstring.make_run_sqlite_fragment(dag, these_coinc_nodes,
"%s_%d" % (tag, n),
clipsegments_sql_filename)
for n, these_coinc_nodes in enumerate(coinc_nodes)
]
#
# lalapps_string_meas_likelihood
#
if verbose:
print("building lalapps_string_meas_likelihood jobs ...",
file=sys.stderr)
likelihood_nodes = [
cosmicstring.make_meas_likelihood_fragment(dag, these_coinc_nodes,
"%s_%d" % (tag, n))
for n, these_coinc_nodes in enumerate(coinc_nodes)
]
#
# write output cache
#
if verbose:
print("writing output cache ...", file=sys.stderr)
for n, (these_coinc_nodes, these_likelihood_nodes) in enumerate(
zip(coinc_nodes, likelihood_nodes)):
power.write_output_cache(
these_coinc_nodes | these_likelihood_nodes, "%s_%s_output.cache" %
(os.path.splitext(dag.get_dag_file())[0], "%s_%d" % (tag, n)))
#
# done
#
return coinc_nodes, likelihood_nodes