def broadcast_test(comm, value, kind, root):
if comm.rank == 0:
print("Broadcasting %s from root %d..." % (kind, root)),
got_value = None
got_value = mpi.broadcast(comm, value, root)
if comm.rank == 0:
print("OK.")
return
def broadcast_test(comm, value, kind, root):
if comm.rank == root:
print ("Broadcasting %s from root %d..." % (kind, root)),
got_value = mpi.broadcast(comm, value, root)
assert got_value == value
if comm.rank == root:
print "OK."
return
def do_training(indices, training_blob, heldout_blob, weights, weights_out, debiasing_weights):
"""
Helper/wrapper function for parallel perceptron training.
Runs one epoch of perceptron training and reports current accuracy on
training data and on heldout data.
"""
# Under de-biasing mode, we only allow features present in a given initial
# weight vector. These are features that have been "selected" under a previously
# run regularized training scheme.
valid_feature_names = None
if FLAGS.debiasing:
valid_feature_names = getFeatureNames(debiasing_weights)
for epoch in range(FLAGS.maxepochs):
# Randomize order of examples; broadcast this randomized order to all processes.
# The particular subset any perceptron process gets for this epoch is dependent
# upon this randomized ordering.
if myRank == 0 and FLAGS.shuffle:
random.shuffle(indices)
indices = mpi.broadcast(value=indices, root=0)
##################################################
# SEARCH: Find 1-best under current model
##################################################
# Run one epoch over training data
io_helper.write_master("===EPOCH %d TRAINING===\n" %(epoch))
newWeights_avg = perceptron_parallel(epoch, indices, training_blob, weights,
valid_feature_names)
####################################
# Dump weights for this iteration
####################################
if myRank == 0:
cPickle.dump(newWeights_avg, weights_out, protocol=cPickle.HIGHEST_PROTOCOL)
# Need to flush output somehow here. Does weights_out.flush() work?
weights_out.flush()
##################################################
# Try a corpus re-decode here with the new weights
# This returns a HELDOUT F-SCORE
##################################################
# Decode dev data with same new learned weight vector
if FLAGS.decodeheldout:
io_helper.write_master("===EPOCH %d DECODE HELDOUT===\n" %(epoch))
decode_parallel(newWeights_avg, indices_dev, heldout_blob, "dev")
if myRank == 0:
weights_out.close()
def do_training(indices, training_blob, heldout_blob, weights, weights_out, debiasing_weights):
"""
Helper/wrapper function for parallel perceptron training.
Runs one epoch of perceptron training and reports current accuracy on
training data and on heldout data.
"""
# Under de-biasing mode, we only allow features present in a given initial
# weight vector. These are features that have been "selected" under a previously
# run regularized training scheme.
valid_feature_names = None
if FLAGS.debiasing:
valid_feature_names = getFeatureNames(debiasing_weights)
for epoch in range(FLAGS.maxepochs):
# Randomize order of examples; broadcast this randomized order to all processes.
# The particular subset any perceptron process gets for this epoch is dependent
# upon this randomized ordering.
if myRank == 0 and FLAGS.shuffle:
random.shuffle(indices)
indices = mpi.broadcast(value=indices, root=0)
##################################################
# SEARCH: Find 1-best under current model
##################################################
# Run one epoch over training data
io_helper.write_master("===EPOCH %d TRAINING===\n" %(epoch))
newWeights_avg = perceptron_parallel(epoch, indices, training_blob, weights,
valid_feature_names)
####################################
# Dump weights for this iteration
####################################
if myRank == 0:
cPickle.dump(newWeights_avg, weights_out, protocol=cPickle.HIGHEST_PROTOCOL)
# Need to flush output somehow here. Does weights_out.flush() work?
weights_out.flush()
##################################################
# Try a corpus re-decode here with the new weights
# This returns a HELDOUT F-SCORE
##################################################
# Decode dev data with same new learned weight vector
if FLAGS.decodeheldout:
io_helper.write_master("===EPOCH %d DECODE HELDOUT===\n" %(epoch))
decode_parallel(newWeights_avg, indices_dev, heldout_blob, "dev")
if myRank == 0:
weights_out.close()
def perceptron_parallel(epoch, indices, blob, weights = None, valid_feature_names = None):
"""
Implements parallelized version of perceptron training for structured outputs
(Collins, 2002; McDonald, 2010).
"""
# Which processor am I?
myRank = mpi.rank
# Setting for output of decoding Path
decodingPath = None
decodingPathFile = robustWrite("%s/%s%s" % (tmpdir, FLAGS.decoding_path_out,str(myRank)))
# Let processor 0 be the master.
masterRank = 0
# How many processors are there?
nProcs = mpi.size
##########################################
# Keep track of time to train this epoch
##########################################
startTime = time.time()
# Restart with weights from last epoch or 0.
# Will ignore any weights passed during function call.
weights_restart_filename = '%s/training-restart.%s' % (tmpdir, str(mpi.rank))
if os.path.isfile(weights_restart_filename):
weights_restart_file = open(weights_restart_filename, 'r')
weights = cPickle.load(weights_restart_file)
weights_restart_file.close()
else:
# If weights passed during function call is None start with empty.
if weights is None or len(weights) == 0:
weights = svector.Vector()
# Restart with previous running weight sum, also.
weights_sum_filename = '%s/training.%s' % (tmpdir, str(mpi.rank))
if os.path.isfile(weights_sum_filename):
weights_sum_file = open(weights_sum_filename, 'r')
weights_sum = cPickle.load(weights_sum_file)
weights_sum_file.close()
else:
weights_sum = svector.Vector()
numChanged = 0
done = False
for i, instanceID in enumerate(indices[:FLAGS.subset]):
if myRank == i % nProcs:
# Assign the current instances we will look at
f = blob['f_instances'][instanceID]
e = blob['e_instances'][instanceID]
etree = blob['etree_instances'][instanceID]
gold_str = blob['gold_instances'][instanceID]
inverse = None
if FLAGS.inverse_a is not None:
inverse = blob['inverse_instances'][instanceID]
a1 = None
if FLAGS.a1 is not None:
a1 = blob['a1_instances'][instanceID]
a2 = None
if FLAGS.a2 is not None:
a2 = blob['a2_instances'][instanceID]
ftree = None
if FLAGS.ftrees is not None:
ftree = blob['ftree_instances'][instanceID]
# Preprocess input data
# f, e are sequences of words
f = f.split() ; e = e.split()
# gold is a sequence of f-e link pairs
gold = Alignment.Alignment(gold_str, FLAGS.inverse)
# Initialize model for this instance
model = GridAlign.Model(f, e, etree, ftree, instanceID, weights, a1, a2,
inverse, LOCAL_FEATURES=blob['localFeatures'],
NONLOCAL_FEATURES=blob['nonlocalFeatures'],
FLAGS=FLAGS)
model.gold = gold
# Initialize model with data tables
model.pef = blob['pef']
model.pfe = blob['pfe']
# Load language model
model.lm = blob['lm']
# Align the current training instance
model.align()
if FLAGS.decoding_path_out is not None:
cPickle.dump(model.decodingPath, decodingPathFile, protocol=cPickle.HIGHEST_PROTOCOL)
######################################################################
# Weight updating
######################################################################
LEARNING_RATE = FLAGS.learningrate
# Set the oracle item
oracle = None
if FLAGS.oracle in ['gold','hope']:
oracle = model.oracle
else:
sys.stderr.write("ERROR: Unknown oracle class: %s\n" %(FLAGS.oracle))
# Set the hypothesis item
hyp = None
if FLAGS.hyp in ['1best', 'fear']:
hyp = model.hyp
else:
sys.stderr.write("ERROR: Unknown hyp class: %s\n" %(FLAGS.hyp))
# Debiasing
if FLAGS.debiasing:
validate_features(oracle.scoreVector, valid_feature_names)
validate_features(hyp.scoreVector, valid_feature_names)
deltas = None
if set(hyp.links) != set(oracle.links):
numChanged += 1
###############################################################
# WEIGHT UPDATES
################################################################
deltas = oracle.scoreVector - hyp.scoreVector
weights = weights + LEARNING_RATE*deltas
# Even if we didnt update, the current weight vector should count towards the sum!
weights_sum += weights
# L1 Projection step
# if w in [-tau, tau], w -> 0
# else, move w closer to 0 by tau.
if FLAGS.tau is not None:
for index, w in weights_sum.iteritems():
if w == 0:
del weights_sum[index]
continue
if index[-3:] == '_nb':
continue
if w > 0 and w <= FLAGS.tau and not FLAGS.negreg:
del weights_sum[index]
elif w < 0 and w >= (FLAGS.tau * -1):
del weights_sum[index]
elif w > 0 and w > FLAGS.tau and not FLAGS.negreg:
weights_sum[index] -= FLAGS.tau
elif w < 0 and w < (FLAGS.tau * -1):
weights_sum[index] += FLAGS.tau
# Set uniq pickled output file for this process
# Holds sum of weights over each iteration for this process
output_filename = "%s/training.%s" %(tmpdir, str(mpi.rank))
output_file = open(output_filename,'w')
# Dump all weights used during this node's run; to be averaged by master along with others
cPickle.dump(weights_sum, output_file, protocol=cPickle.HIGHEST_PROTOCOL)
output_file.close()
# Remeber just the last weights used for this process; start here next epoch.
output_filename_last_weights = "%s/training-restart.%s" %(tmpdir, str(mpi.rank))
output_file_last_weights = open(output_filename_last_weights,'w')
cPickle.dump(weights, output_file_last_weights, protocol=cPickle.HIGHEST_PROTOCOL)
output_file_last_weights.close()
decodingPathFile.close()
#############################################
# Gather "done" messages from workers
#############################################
# Synchronize
done = mpi.gather(value=True,root=0)
#####################################################################################
# Compute f-measure over all alignments
#####################################################################################
masterWeights = svector.Vector()
if myRank == masterRank:
decodePathFiles = {}
# Read pickled output
for rank in range(nProcs):
input_filename = tmpdir+'/training.'+str(rank)
input_file = open(input_filename,'r')
masterWeights += cPickle.load(input_file)
input_file.close()
decodePathFiles[rank] = robustRead("%s/%s%s" % (tmpdir, FLAGS.decoding_path_out, str(rank)) )
sys.stderr.write("Done reading data.\n")
sys.stderr.write("len(masterWeights)= %d\n"%(len(masterWeights)))
sys.stderr.flush()
######################################################
# AVERAGED WEIGHTS
######################################################
sys.stderr.write("[%d] Averaging weights.\n" %(mpi.rank))
sys.stderr.flush()
masterWeights = masterWeights / (len(indices) * (epoch+1))
# Dump master weights to file
# There is only one weight vector in this file at a time.
mw = robustWrite(tmpdir+'/weights')
cPickle.dump(masterWeights,mw,protocol=cPickle.HIGHEST_PROTOCOL)
mw.close()
# Write decoding path
decodingPathList = []
if FLAGS.decoding_path_out is not None:
path_out = robustWrite(FLAGS.decoding_path_out, encoding="utf-8")
for i, instanceID in enumerate(indices[:FLAGS.subset]):
node = i % nProcs
chosenTree = cPickle.load(decodePathFiles[node])
heappush(decodingPathList, (instanceID, chosenTree))
orderedList = [heappop(decodingPathList)[1] for _ in xrange(len(decodingPathList))]
path_out.write(u"\n".join(orderedList))
path_out.close()
# CLEAN UP
for i in range(nProcs):
decodePathFiles[i].close()
######################################################################
# All processes read and load new averaged weights
######################################################################
# But make sure worker nodes don't attempt to read from the weights
# file before the root node has written it.
# Sync-up with a blocking broadcast call
ready = mpi.broadcast(value=True, root=0)
mw = robustRead(tmpdir+'/weights')
masterWeights = cPickle.load(mw)
mw.close()
######################################################################
# Print report for this iteration
######################################################################
elapsedTime = time.time() - startTime
if myRank == masterRank:
# masterRank is printing elapsed time.
# May differ at each node.
sys.stderr.write("Time: %0.2f\n" %(elapsedTime))
sys.stderr.write("[%d] Finished training.\n" %(mpi.rank))
return masterWeights
a2_dev_instances = []
gold_dev_instances = []
inverse_dev_instances = []
tmpdir = None
if mpi.rank == 0:
base_tempdir = None
if FLAGS.tempdir is not None:
base_tempdir = FLAGS.tempdir
else:
base_tempdir = tempfile.gettempdir()
if base_tempdir is None:
base_tempdir = "."
tmpdir = tempfile.mkdtemp(prefix='align-'+str(os.getpid())+'-',
dir=base_tempdir)
tmpdir = mpi.broadcast(value=tmpdir, root=0)
################################################
# Load training examples
################################################
count = 1
etree_file_handle = readDependencyForestFile(FLAGS.etrees)
for f, e, etree in izip(file_handles['f'], file_handles['e'], etree_file_handle):
f_instances.append(f.strip())
e_instances.append(e.strip())
etree_instances.append(etree.strip())
if FLAGS.train and FLAGS.partial!=-1 and count>= FLAGS.partial:
break
count += 1
indices = range(len(e_instances))
all_nweights = mpi.gather(mpi.world, nweights, parallel.master)
if parallel.rank == parallel.master:
sumweights = sum(all_outweights, svector.Vector())
outweights = sumweights / float(sum(all_nweights))
log.write("summed feature weights: %s n=%d\n" %
(sumweights * watch_features, sum(all_nweights)))
log.write("averaged feature weights: %s\n" %
(outweights * watch_features))
if opts.outweightfilename:
if not opts.parallel or parallel.rank == parallel.master:
outweightfile.write("%s\n" % outweights)
outweightfile.flush()
if opts.parallel:
outweights = mpi.broadcast(mpi.world, outweights, parallel.master)
# Process heldout data
if not opts.parallel or parallel.rank != parallel.master:
saveweights = thedecoder.weights
thedecoder.weights = outweights
if opts.parallel:
outsents = parallel.pmap(process_heldout,
heldoutsents,
tag=0,
verbose=1)
else:
outsents = (process_heldout(sent) for sent in heldoutsents)
if opts.parallel:
all_outweights = mpi.gather(mpi.world, outweights, parallel.master)
all_nweights = mpi.gather(mpi.world, nweights, parallel.master)
if parallel.rank == parallel.master:
sumweights = sum(all_outweights, svector.Vector())
outweights = sumweights / float(sum(all_nweights))
log.write("summed feature weights: %s n=%d\n" % (sumweights * watch_features, sum(all_nweights)))
log.write("averaged feature weights: %s\n" % (outweights * watch_features))
if opts.outweightfilename:
if not opts.parallel or parallel.rank == parallel.master:
outweightfile.write("%s\n" % outweights)
outweightfile.flush()
if opts.parallel:
outweights = mpi.broadcast(mpi.world, outweights, parallel.master)
# Process heldout data
if not opts.parallel or parallel.rank != parallel.master:
saveweights = thedecoder.weights
thedecoder.weights = outweights
if opts.parallel:
outsents = parallel.pmap(process_heldout, heldoutsents, tag=0, verbose=1)
else:
outsents = (process_heldout(sent) for sent in heldoutsents)
if not opts.parallel or parallel.rank == parallel.master:
heldout_score_comps = svector.Vector()
for outsent in outsents:
def perceptron_parallel(epoch, indices, blob, weights = None, valid_feature_names = None):
"""
Implements parallelized version of perceptron training for structured outputs
(Collins, 2002; McDonald, 2010).
"""
# Which processor am I?
myRank = mpi.rank
# Let processor 0 be the master.
masterRank = 0
# How many processors are there?
nProcs = mpi.size
##########################################
# Keep track of time to train this epoch
##########################################
startTime = time.time()
# Restart with weights from last epoch or 0.
# Will ignore any weights passed during function call.
weights_restart_filename = '%s/training-restart.%s' % (tmpdir, str(mpi.rank))
if os.path.isfile(weights_restart_filename):
weights_restart_file = open(weights_restart_filename, 'r')
weights = cPickle.load(weights_restart_file)
weights_restart_file.close()
else:
# If weights passed during function call is None start with empty.
if weights is None or len(weights) == 0:
weights = svector.Vector()
# Restart with previous running weight sum, also.
weights_sum_filename = '%s/training.%s' % (tmpdir, str(mpi.rank))
if os.path.isfile(weights_sum_filename):
weights_sum_file = open(weights_sum_filename, 'r')
weights_sum = cPickle.load(weights_sum_file)
weights_sum_file.close()
else:
weights_sum = svector.Vector()
numChanged = 0
done = False
for i, instanceID in enumerate(indices[:FLAGS.subset]):
if myRank == i % nProcs:
# Assign the current instances we will look at
f = blob['f_instances'][instanceID]
e = blob['e_instances'][instanceID]
etree = blob['etree_instances'][instanceID]
gold_str = blob['gold_instances'][instanceID]
inverse = None
if FLAGS.inverse is not None:
inverse = blob['inverse_instances'][instanceID]
a1 = None
if FLAGS.a1 is not None:
a1 = blob['a1_instances'][instanceID]
a2 = None
if FLAGS.a2 is not None:
a2 = blob['a2_instances'][instanceID]
ftree = None
if FLAGS.ftrees is not None:
ftree = blob['ftree_instances'][instanceID]
# Preprocess input data
# f, e are sequences of words
f = f.split() ; e = e.split()
# gold is a sequence of f-e link pairs
gold = Alignment.Alignment(gold_str)
# Initialize model for this instance
model = GridAlign.Model(f, e, etree, ftree, instanceID, weights, a1, a2,
inverse, LOCAL_FEATURES=blob['localFeatures'],
NONLOCAL_FEATURES=blob['nonlocalFeatures'],
FLAGS=FLAGS)
model.gold = gold
# Initialize model with data tables
model.pef = blob['pef']
model.pfe = blob['pfe']
# Align the current training instance
model.align()
######################################################################
# Weight updating
######################################################################
LEARNING_RATE = FLAGS.learningrate
# Set the oracle item
oracle = None
if FLAGS.oracle == 'gold':
oracle = model.oracle
elif FLAGS.oracle == 'hope':
oracle = model.hope
else:
sys.stderr.write("ERROR: Unknown oracle class: %s\n" %(FLAGS.oracle))
# Set the hypothesis item
hyp = None
if FLAGS.hyp == '1best':
hyp = model.modelBest
elif FLAGS.hyp == 'fear':
hyp = model.fear
else:
sys.stderr.write("ERROR: Unknown hyp class: %s\n" %(FLAGS.hyp))
# Debiasing
if FLAGS.debiasing:
validate_features(oracle.scoreVector, valid_feature_names)
validate_features(hyp.scoreVector, valid_feature_names)
deltas = None
if set(hyp.links) != set(oracle.links):
numChanged += 1
###############################################################
# WEIGHT UPDATES
################################################################
deltas = oracle.scoreVector - hyp.scoreVector
weights = weights + LEARNING_RATE*deltas
# Even if we didnt update, the current weight vector should count towards the sum!
weights_sum += weights
# L1 Projection step
# if w in [-tau, tau], w -> 0
# else, move w closer to 0 by tau.
if FLAGS.tau is not None:
for index, w in weights_sum.iteritems():
if w == 0:
del weights_sum[index]
continue
if index[-3:] == '_nb':
continue
if w > 0 and w <= FLAGS.tau and not FLAGS.negreg:
del weights_sum[index]
elif w < 0 and w >= (FLAGS.tau * -1):
del weights_sum[index]
elif w > 0 and w > FLAGS.tau and not FLAGS.negreg:
weights_sum[index] -= FLAGS.tau
elif w < 0 and w < (FLAGS.tau * -1):
weights_sum[index] += FLAGS.tau
# Set uniq pickled output file for this process
# Holds sum of weights over each iteration for this process
output_filename = "%s/training.%s" %(tmpdir, str(mpi.rank))
output_file = open(output_filename,'w')
# Dump all weights used during this node's run; to be averaged by master along with others
cPickle.dump(weights_sum, output_file, protocol=cPickle.HIGHEST_PROTOCOL)
output_file.close()
# Remeber just the last weights used for this process; start here next epoch.
output_filename_last_weights = "%s/training-restart.%s" %(tmpdir, str(mpi.rank))
output_file_last_weights = open(output_filename_last_weights,'w')
cPickle.dump(weights, output_file_last_weights, protocol=cPickle.HIGHEST_PROTOCOL)
output_file_last_weights.close()
#############################################
# Gather "done" messages from workers
#############################################
# Synchronize
done = mpi.gather(value=True,root=0)
#####################################################################################
# Compute f-measure over all alignments
#####################################################################################
masterWeights = svector.Vector()
if myRank == masterRank:
# Read pickled output
for rank in range(nProcs):
input_filename = tmpdir+'/training.'+str(rank)
input_file = open(input_filename,'r')
masterWeights += cPickle.load(input_file)
input_file.close()
sys.stderr.write("Done reading data.\n")
sys.stderr.write("len(masterWeights)= %d\n"%(len(masterWeights)))
sys.stderr.flush()
######################################################
# AVERAGED WEIGHTS
######################################################
sys.stderr.write("[%d] Averaging weights.\n" %(mpi.rank))
sys.stderr.flush()
masterWeights = masterWeights / (len(indices) * (epoch+1))
# Dump master weights to file
# There is only one weight vector in this file at a time.
mw = robustWrite(tmpdir+'/weights')
cPickle.dump(masterWeights,mw,protocol=cPickle.HIGHEST_PROTOCOL)
mw.close()
######################################################################
# All processes read and load new averaged weights
######################################################################
# But make sure worker nodes don't attempt to read from the weights
# file before the root node has written it.
# Sync-up with a blocking broadcast call
ready = mpi.broadcast(value=True, root=0)
mw = robustRead(tmpdir+'/weights')
masterWeights = cPickle.load(mw)
mw.close()
######################################################################
# Print report for this iteration
######################################################################
elapsedTime = time.time() - startTime
if myRank == masterRank:
# masterRank is printing elapsed time.
# May differ at each node.
sys.stderr.write("Time: %0.2f\n" %(elapsedTime))
sys.stderr.write("[%d] Finished training.\n" %(mpi.rank))
return masterWeights
a2_dev_instances = []
gold_dev_instances = []
inverse_dev_instances = []
tmpdir = None
if mpi.rank == 0:
base_tempdir = None
if FLAGS.tempdir is not None:
base_tempdir = FLAGS.tempdir
else:
base_tempdir = tempfile.gettempdir()
if base_tempdir is None:
base_tempdir = "."
tmpdir = tempfile.mkdtemp(prefix='align-'+str(os.getpid())+'-',
dir=base_tempdir)
tmpdir = mpi.broadcast(value=tmpdir, root=0)
################################################
# Load training examples
################################################
count = 1
etree_file_handle = readDependencyFile(FLAGS.etrees)
for f, e, etree in izip(file_handles['f'], file_handles['e'], etree_file_handle):
f_instances.append(f.strip())
e_instances.append(e.strip())
etree_instances.append(etree.strip())
if FLAGS.train and FLAGS.partial!=-1 and count>= FLAGS.partial:
break
count += 1
indices = range(len(e_instances))