def write(self, filename="-", weights=None):
if weights is None:
weights = self.weights
if filename == "-":
outfile = sys.stdout
filename = "STDOUT" # careful overriding
else:
outfile = open(filename, "wt")
self.print_templates(outfile)
mytime = Mytime()
nonzero = 0
print >> logs, "sorting %d features..." % len(weights),
for i, f in enumerate(sorted(weights), 1):
if i == 1: # sorting done
print >> logs, "done in %.2lf seconds." % mytime.period()
print >> logs, "writing features to %s..." % filename
v = weights[f]
if math.fabs(v) > 1e-3:
print >> outfile, "%s\t%.5lf" % (f, v)
nonzero += 1
if self.unk > 0: # print known words
print >> outfile, " " + " ".join(sorted(
self.knowns)) # " " to mark
print >> logs, "%d nonzero feature instances written in %.2lf seconds." % \
(nonzero, mytime.period()) ## nonzero != i
def read_weights(self, filename, infertemplates=False):
'''instances are like "s0t-q0t=LRB-</s>=>LEFT 3.8234"'''
infile = self.read_templates(filename)
infertemplates = len(self.templates) < 1
if infertemplates:
print >> logs, "will infer templates from weights..."
mytime = Mytime()
i = 0
if infile is not None:
print >> logs, "reading feature weights from %s\t" % filename,
for i, line in enumerate(infile, 1):
if i % 200000 == 0:
print >> logs, "%d lines read..." % i,
if line[0] == " ":
# TODO: separate known words line (last line)
self.knowns = set(line.split())
print >> logs, "\n%d known words read." % len(self.knowns)
self.unk = 1 # in cae you forgot to say it; doesn't matter 1 or x
break
feat, weight = line.split()
self.weights[feat] = float(weight)
if infertemplates:
self.add_template(feat.split("=", 1)[0],
1) ## one occurrence
print >> logs, "\n%d feature instances (%d lines) read in %.2lf seconds." % \
(len(self.weights), i, mytime.period())
self.print_autoevals()
def write(self, filename="-", weights=None):
if weights is None:
weights = self.weights
if filename == "-":
outfile = sys.stdout
filename = "STDOUT" # careful overriding
else:
outfile = open(filename, "wt")
self.print_templates(outfile)
mytime = Mytime()
nonzero = 0
# my wvector
for action, feats in weights.iteritems():
for f in sorted(feats):
v = feats[f]
if math.fabs(float(v)) > 1e-3: # TODO
tid, feat = f.split("=", 1)
print >> outfile, "%s=%s=>%s\t%.5lf" % (
self.list_templates[int(tid)][0], feat, action,
float(v))
nonzero += 1
print >> logs, "%d nonzero feature instances written in %.2lf seconds." % \
(nonzero, mytime.period()) ## nonzero != i
def read_weights(self, filename, infertemplates=False):
'''instances are like "s0t-q0t=LRB-</s>=>LEFT 3.8234"'''
infile = self.read_templates(filename)
infertemplates = len(self.templates) <= 1
if infertemplates:
print >> logs, "will infer templates from weights..."
mytime = Mytime()
i = 0
if infile is not None:
print >> logs, "reading feature weights from %s\t" % filename,
for i, line in enumerate(infile, 1):
if i % 200000 == 0:
print >> logs, "%d lines read..." % i,
feat, weight = line.split()
weight = WVector.value_class(
float(weight)) # in case of mydouble
if FLAGS.use_template_id:
template, instance = feat.split("=", 1)
tid = self.templates[template]
feat = "%d=%s" % (tid, instance)
if Model.doublehash == 1:
if FLAGS.tuplefeats:
f, action = instance.rsplit("=>", 1)
#action = Model.mapnames[action]
fs = tuple(f.split("|"))
if FLAGS.integerize:
fs = tuple(map(Vocab.str2id, fs))
self.weights[action][(tid, ) + fs] = weight
else:
f, action = feat.rsplit("=>", 1)
#action = Model.mapnames[action]
self.weights[action][f] = weight
elif Model.doublehash == 2:
f, action = instance.rsplit("=>", 1)
action = Model.mapnames[action]
self.weights[action][tid][f] = weight
else:
self.weights[feat] = weight
if infertemplates:
self.add_template(feat.split("=", 1)[0],
1) ## one occurrence
print >> logs, "\n%d feature instances (%d lines) read in %.2lf seconds." % \
(len(self.weights), i, mytime.period())
self.print_autoevals()
yield raw_input()
except:
return
logs = sys.stderr
from collections import defaultdict
from svector import Vector
from model import Model
from deptree import DepTree, DepVal
from mytime import Mytime
mytime = Mytime()
import gflags as flags
FLAGS = flags.FLAGS
flags.DEFINE_integer("beam", 1, "beam width", short_name="b")
flags.DEFINE_integer(
"leftbeam", 1000, "leftptrs beam width"
) # number of left items (predictors to be combined w/ current)
flags.DEFINE_integer("kbest", 0, "kbest", short_name="k")
flags.DEFINE_boolean("forest", False, "dump the forest")
flags.DEFINE_boolean("earlystop", False,
"try early stop (compared with gold seq)")
flags.DEFINE_integer("debuglevel",
0,
def train(self):
start_mem = memory()
starttime = time.time()
print >> logs, "starting perceptron at", time.ctime()
best_prec = 0
for it in xrange(1, self.iter + 1):
print >> logs, "iteration %d starts..............%s" % (
it, time.ctime())
curr_mem = memory()
iterstarttime = time.time()
self.decoder.num_edges = 0
num_updates, early_updates, num_steps = self.one_pass_on_train()
iterendtime = time.time()
print >> logs, "memory usage at iter %d: extra %s, total %s" % (
it, human(memory(curr_mem)), human(memory(start_mem)))
if FLAGS.debuglevel >= 1:
print >> logs, "weights=", self.weights
curr_mem = memory()
print >> logs, "iteration %d training finished at %s. now evaluating on dev..." % (
it, time.ctime())
## avgweights = self.avg_weights() if self.avg else self.weights
avgtime = 0
timer = Mytime()
if self.avg:
## print >> logs, " w=", self.weights
## print >> logs, " ".join(map(str, [x.get_step() for x in self.weights.values()]))
self.weights.set_avg(self.c)
avgtime += timer.gap()
if FLAGS.debuglevel >= 1:
print >> logs, "avgweights=", self.weights
prec = self.eval_on_dev()
print >> logs, "eval on dev took %.1f seconds." % timer.gap()
print >> logs, "at iteration {0}, updates= {1} (early {4}), dev= {2}{7}, |w|= {3}, time= {5:.3f}h acctime= {6:.3f}h, root={10:.1%}"\
.format(it, num_updates, prec, len(self.weights), early_updates,
(time.time() - iterstarttime)/3600,
(time.time() - starttime)/3600.,
"+" if prec > best_prec else "",
num_steps, self.decoder.num_edges,
prec.root())
logs.flush()
if prec > best_prec:
best_prec = prec
best_it = it
best_wlen = len(self.weights)
best_time = time.time() - starttime
print >> logs, "new high at iteration {0}: {1}. Dumping Weights...".format(
it, prec)
if not FLAGS.dump_last:
self.dump(self.weights)
else:
self.bestweights = self.weights.deepcopy()
if self.avg:
timer = Mytime()
self.weights.reset_avg(self.c) # restore weights
t = timer.gap()
print >> logs, "avg weights (set/reset) took %.1f+%.1f=%.1f seconds." % (
avgtime, t, avgtime + t)
## self.decoder.model.weights = self.weights # restore non-avg
## del avgweights
gc.collect()
if FLAGS.mydouble:
from mydouble import counts
print >> logs, "mydouble usage and freed: %d %d" % counts()
print >> logs, "peaked at iteration {0}: {1} ({3:.1f}h), |bestw|= {2}.".format(
best_it, best_prec, best_wlen, best_time / 3600)
print >> logs, best_prec.details()
print >> logs, "perceptron training of %d iterations finished on %s (took %.2f hours)" % \
(it, time.ctime(), (time.time() - starttime)/3600.)
if FLAGS.dump_last:
self.dump(self.bestweights)
def read_weights_and_insert_different_noise(self,
filename,
noise_info,
infertemplates=False):
# ADDED CODE HERE
'''instances are like "s0t-q0t=LRB-</s>=>LEFT 3.8234"'''
"""
noise_info = {
'method': FLAGS.noise_method,
'mu': FLAGS.mu,
'sigma': FLAGS.sigma,
'noise_file_path': FLAGS.noise_file_path,
}
"""
method = noise_info['method']
mu = noise_info['mu']
sigma = noise_info['sigma']
noise_file_path = noise_info['noise_file_path']
noises_vector = numpy.load(noise_file_path)
infile = self.read_templates(filename)
infertemplates = len(self.templates) <= 1
if infertemplates:
print >> logs, "will infer templates from weights..."
mytime = Mytime()
i = 0
if infile is not None:
print >> logs, "reading feature weights from %s\t" % filename,
# for i, line in enumerate(infile, 1):
for i, (line, noises_vector_coeff) in enumerate(
zip(infile, noises_vector), 1):
if i % 200000 == 0:
print >> logs, "%d lines read..." % i,
feat, weight = line.split()
weight = float(
weight
) # WVector.value_class(float(weight)) # in case of mydouble
noise_ = float(
noises_vector_coeff) * numpy.random.randn() + float(mu)
if method == 'a':
weight = weight + noise_
elif method == 'm':
weight = weight * noise_
if FLAGS.use_template_id:
template, instance = feat.split("=", 1)
tid = self.templates[template]
feat = "%d=%s" % (tid, instance)
if Model.doublehash == 1:
if FLAGS.tuplefeats:
f, action = instance.rsplit("=>", 1)
# action = Model.mapnames[action]
fs = tuple(f.split("|"))
self.weights[action][(tid, ) + fs] = weight
else:
f, action = feat.rsplit("=>", 1)
# action = Model.mapnames[action]
self.weights[action][f] = weight
elif Model.doublehash == 2:
f, action = instance.rsplit("=>", 1)
action = Model.mapnames[action]
self.weights[action][tid][f] = weight
else:
self.weights[feat] = weight
if infertemplates:
self.add_template(feat.split("=", 1)[0],
1) ## one occurrence
print >> logs, "\n%d feature instances (%d lines) read in %.2lf seconds." % \
(len(self.weights), i, mytime.period())
self.print_autoevals()