def corrupt(self):
"""
Return a (notw2, weight), a corrupt target word and its weight.
Note: This will return a different random value every call.
"""
from hyperparameters import HYPERPARAMETERS
import random
possible_targets = targetmap()[self.w1][self.l2]
assert len(possible_targets) > 1
assert self.w2 in possible_targets
notw2 = self.w2
cnt = 0
while self.w2 == notw2:
if HYPERPARAMETERS["NGRAM_FOR_TRAINING_NOISE"] == 0:
notw2 = random.choice(possible_targets)
pr = 1./len(possible_targets)
elif HYPERPARAMETERS["NGRAM_FOR_TRAINING_NOISE"] == 1:
assert 0
# import noise
# from common.myrandom import weighted_sample
# e[-1], pr = weighted_sample(noise.indexed_weights())
## from vocabulary import wordmap
## print wordmap.str(e[-1]), pr
else:
assert 0
cnt += 1
# Backoff to 0gram smoothing if we fail 10 times to get noise.
if cnt > 10: notw2 = random.choice(possible_targets)
if HYPERPARAMETERS["UNIFORM EXAMPLE WEIGHTS"]:
weight = 1.
else:
weight = 1./pr
return notw2, weight
def corrupt(self):
"""
Return a (notw2, weight), a corrupt target word and its weight.
Note: This will return a different random value every call.
"""
from hyperparameters import HYPERPARAMETERS
import random
possible_targets = targetmap()[self.w1][self.l2]
assert len(possible_targets) > 1
assert self.w2 in possible_targets
notw2 = self.w2
cnt = 0
while self.w2 == notw2:
if HYPERPARAMETERS["NGRAM_FOR_TRAINING_NOISE"] == 0:
notw2 = random.choice(possible_targets)
pr = 1. / len(possible_targets)
elif HYPERPARAMETERS["NGRAM_FOR_TRAINING_NOISE"] == 1:
assert 0
# import noise
# from common.myrandom import weighted_sample
# e[-1], pr = weighted_sample(noise.indexed_weights())
## from vocabulary import wordmap
## print wordmap.str(e[-1]), pr
else:
assert 0
cnt += 1
# Backoff to 0gram smoothing if we fail 10 times to get noise.
if cnt > 10: notw2 = random.choice(possible_targets)
if HYPERPARAMETERS["UNIFORM EXAMPLE WEIGHTS"]:
weight = 1.
else:
weight = 1. / pr
return notw2, weight
def get_training_biexample(l1, l2, f1, f2, falign):
"""
Generator of bilingual training examples from this bicorpus.
"""
import common.hyperparameters
HYPERPARAMETERS = common.hyperparameters.read("language-model")
WINDOW = HYPERPARAMETERS["WINDOW_SIZE"]
for ws1, ws2, links in bicorpus_sentences_and_alignments(l1, l2, f1, f2, falign):
for i1, i2 in links:
w1 = ws1[i1]
w2 = ws2[i2]
l2new = language(w2)
assert HYPERPARAMETERS["W2W SKIP TRANSLATIONS TO UNKNOWN WORD"]
# Skip translations to unknown words
if wordform(w2) == "*UNKNOWN*": continue
assert l2new == l2
# Skip translations from unknown words
if wordform(w1) == "*UNKNOWN*": continue
# If we are filtering examples by lemma
if not(HYPERPARAMETERS["W2W FOCUS LEMMAS"] is None or len (HYPERPARAMETERS["W2W FOCUS LEMMAS"]) == 0):
# print wordmap().str(w1), wordmap().str(w2)
assert language(w1) == "en"
# from lemmatizer import lemmatize
# if lemmatize(language(w1), wordform(w1)) not in HYPERPARAMETERS["W2W FOCUS LEMMAS"]:
# logging.debug("Focus word %s (lemma %s) not in our list of focus lemmas" % (`wordmap().str(w1)`, lemmatize(language(w1), wordform(w1))))
if wordform(w1) not in HYPERPARAMETERS["W2W FOCUS LEMMAS"]:
logging.debug("Focus word %s not in our list of focus lemmas" % (`wordmap().str(w1)`))
continue
if w1 not in targetmap():
logging.warning("No translations for word %s, skipping" % (`wordmap().str(w1)`))
continue
if l2new not in targetmap()[w1]:
logging.warning("Word %s has no translations for language %s, skipping" % (`wordmap().str(w1)`, l2new))
continue
if w2 not in targetmap()[w1][l2new]:
logging.error("Word %s cannot translate to word %s, skipping" % (`wordmap().str(w1)`, `wordmap().str(w2)`))
continue
if len(targetmap()[w1][l2new]) == 1:
logging.debug("Word %s has only one translation in language %s, skipping" % (`wordmap().str(w1)`, l2new))
continue
# Extract the window of tokens around index i1. Pad with *LBOUNDARY* and *RBOUNDARY* as necessary.
min = i1 - (WINDOW-1)/2
max = i1 + (WINDOW-1)/2
lpad = 0
rpad = 0
if min < 0:
lpad = -min
min = 0
if max >= len(ws1):
rpad = max - (len(ws1)-1)
max = len(ws1)-1
assert lpad + (max - min + 1) + rpad == WINDOW
# print i1 - (WINDOW-1)/2, i1 + (WINDOW-1)/2
# print "min=%d, max=%d, lpad=%d, rpad=%d" % (min, max, lpad, rpad)
seq = [wordmap().id((None, "*LBOUNDARY*"))]*lpad + ws1[min:max+1] + [wordmap().id((None, "*RBOUNDARY*"))]*rpad
# print [wordmap.str(w) for w in seq]
assert len(seq) == WINDOW
# print ws1[i1 - (WINDOW-1)/2:i1 + (WINDOW-1)/2]
assert seq[(WINDOW-1)/2] == w1
yield BilingualExample(l1, seq, w1, w2)
print >> sys.stderr, "WEIRD WORD: %s" % word
word = string.lower(word)
assert len(vals[1:]) == HYPERPARAMETERS["EMBEDDING_SIZE"]
tot += 1
if tot % 10000 == 0:
print >> sys.stderr, "\tRead %d lines from %s" % (tot, HYPERPARAMETERS["W2W INITIAL EMBEDDINGS"])
if word in original_embeddings:
# print >> sys.stderr, "Skipping word %s (originally %s), we already have an embedding for it" % (word, vals[0])
continue
else:
original_embeddings[word] = numpy.array([float(v) for v in vals[1:]])
print >> sys.stderr, "...done reading embeddings from %s" % HYPERPARAMETERS["W2W INITIAL EMBEDDINGS"]
print >> sys.stderr, "Skipped %s words for which we had duplicate embeddings" % percent(tot-len(original_embeddings), tot)
print >> sys.stderr, stats()
reversemap = targetmap(name="reverse")
embeddings = numpy.zeros((wordmap().len, HYPERPARAMETERS["EMBEDDING_SIZE"]))
assert embeddings.shape == (wordmap().len, HYPERPARAMETERS["EMBEDDING_SIZE"])
ELANG = HYPERPARAMETERS["W2W INITIAL EMBEDDINGS LANGUAGE"]
for w in range(wordmap().len):
embedding = None
# If this word is in a different language than the embeddings.
if language(w) != HYPERPARAMETERS["W2W INITIAL EMBEDDINGS LANGUAGE"]:
if w not in reversemap:
print >> sys.stderr, "Word %s is not even in target map! Using *UNKNOWN*" % `wordmap().str(w)`
embedding = original_embeddings["*UNKNOWN*"]
elif ELANG not in reversemap[w]:
print >> sys.stderr, "Have no %s translations for word %s, only have %s, using *UNKNOWN*" % (ELANG, wordmap().str(w), reversemap[w].keys())
embedding = original_embeddings["*UNKNOWN*"]
if tot % 10000 == 0:
print >> sys.stderr, "\tRead %d lines from %s" % (
tot, HYPERPARAMETERS["W2W INITIAL EMBEDDINGS"])
if word in original_embeddings:
# print >> sys.stderr, "Skipping word %s (originally %s), we already have an embedding for it" % (word, vals[0])
continue
else:
original_embeddings[word] = numpy.array(
[float(v) for v in vals[1:]])
print >> sys.stderr, "...done reading embeddings from %s" % HYPERPARAMETERS[
"W2W INITIAL EMBEDDINGS"]
print >> sys.stderr, "Skipped %s words for which we had duplicate embeddings" % percent(
tot - len(original_embeddings), tot)
print >> sys.stderr, stats()
reversemap = targetmap(name="reverse")
embeddings = numpy.zeros(
(wordmap().len, HYPERPARAMETERS["EMBEDDING_SIZE"]))
assert embeddings.shape == (wordmap().len,
HYPERPARAMETERS["EMBEDDING_SIZE"])
ELANG = HYPERPARAMETERS["W2W INITIAL EMBEDDINGS LANGUAGE"]
for w in range(wordmap().len):
embedding = None
# If this word is in a different language than the embeddings.
if language(w) != HYPERPARAMETERS["W2W INITIAL EMBEDDINGS LANGUAGE"]:
if w not in reversemap:
print >> sys.stderr, "Word %s is not even in target map! Using *UNKNOWN*" % ` wordmap(
).str(w) `
embedding = original_embeddings["*UNKNOWN*"]
def get_training_biexample(l1, l2, f1, f2, falign):
"""
Generator of bilingual training examples from this bicorpus.
"""
import common.hyperparameters
HYPERPARAMETERS = common.hyperparameters.read("language-model")
WINDOW = HYPERPARAMETERS["WINDOW_SIZE"]
for ws1, ws2, links in bicorpus_sentences_and_alignments(
l1, l2, f1, f2, falign):
for i1, i2 in links:
w1 = ws1[i1]
w2 = ws2[i2]
l2new = language(w2)
assert HYPERPARAMETERS["W2W SKIP TRANSLATIONS TO UNKNOWN WORD"]
# Skip translations to unknown words
if wordform(w2) == "*UNKNOWN*": continue
assert l2new == l2
# Skip translations from unknown words
if wordform(w1) == "*UNKNOWN*": continue
# If we are filtering examples by lemma
if not (HYPERPARAMETERS["W2W FOCUS LEMMAS"] is None
or len(HYPERPARAMETERS["W2W FOCUS LEMMAS"]) == 0):
# print wordmap().str(w1), wordmap().str(w2)
assert language(w1) == "en"
# from lemmatizer import lemmatize
# if lemmatize(language(w1), wordform(w1)) not in HYPERPARAMETERS["W2W FOCUS LEMMAS"]:
# logging.debug("Focus word %s (lemma %s) not in our list of focus lemmas" % (`wordmap().str(w1)`, lemmatize(language(w1), wordform(w1))))
if wordform(w1) not in HYPERPARAMETERS["W2W FOCUS LEMMAS"]:
logging.debug(
"Focus word %s not in our list of focus lemmas" %
( ` wordmap().str(w1) `))
continue
if w1 not in targetmap():
logging.warning("No translations for word %s, skipping" %
( ` wordmap().str(w1) `))
continue
if l2new not in targetmap()[w1]:
logging.warning(
"Word %s has no translations for language %s, skipping" %
( ` wordmap().str(w1) `, l2new))
continue
if w2 not in targetmap()[w1][l2new]:
logging.error("Word %s cannot translate to word %s, skipping" %
( ` wordmap().str(w1) `, ` wordmap().str(w2) `))
continue
if len(targetmap()[w1][l2new]) == 1:
logging.debug(
"Word %s has only one translation in language %s, skipping"
% ( ` wordmap().str(w1) `, l2new))
continue
# Extract the window of tokens around index i1. Pad with *LBOUNDARY* and *RBOUNDARY* as necessary.
min = i1 - (WINDOW - 1) / 2
max = i1 + (WINDOW - 1) / 2
lpad = 0
rpad = 0
if min < 0:
lpad = -min
min = 0
if max >= len(ws1):
rpad = max - (len(ws1) - 1)
max = len(ws1) - 1
assert lpad + (max - min + 1) + rpad == WINDOW
# print i1 - (WINDOW-1)/2, i1 + (WINDOW-1)/2
# print "min=%d, max=%d, lpad=%d, rpad=%d" % (min, max, lpad, rpad)
seq = [wordmap().id((None, "*LBOUNDARY*"))
] * lpad + ws1[min:max +
1] + [wordmap().id(
(None, "*RBOUNDARY*"))] * rpad
# print [wordmap.str(w) for w in seq]
assert len(seq) == WINDOW
# print ws1[i1 - (WINDOW-1)/2:i1 + (WINDOW-1)/2]
assert seq[(WINDOW - 1) / 2] == w1
yield BilingualExample(l1, seq, w1, w2)
