class jPosDepLearner:
def __init__(self, vocab, pos, rels, w2i, c2i, options):
self.model = ParameterCollection()
random.seed(1)
self.trainer = AdamTrainer(self.model)
#self.trainer = SimpleSGDTrainer(self.model)
self.activations = {
'tanh': tanh,
'sigmoid': logistic,
'relu': rectify,
'tanh3': (lambda x: tanh(cwise_multiply(cwise_multiply(x, x), x)))
}
self.activation = self.activations[options.activation]
self.blstmFlag = options.blstmFlag
self.labelsFlag = options.labelsFlag
self.costaugFlag = options.costaugFlag
self.bibiFlag = options.bibiFlag
self.ldims = options.lstm_dims
self.wdims = options.wembedding_dims
self.cdims = options.cembedding_dims
self.layers = options.lstm_layers
self.wordsCount = vocab
self.vocab = {word: ind + 3 for word, ind in w2i.iteritems()}
self.pos = {word: ind for ind, word in enumerate(pos)}
self.id2pos = {ind: word for ind, word in enumerate(pos)}
self.c2i = c2i
self.rels = {word: ind for ind, word in enumerate(rels)}
self.irels = rels
self.external_embedding, self.edim = None, 0
if options.external_embedding is not None:
external_embedding_fp = open(options.external_embedding, 'r')
external_embedding_fp.readline()
self.external_embedding = {
line.split(' ')[0]:
[float(f) for f in line.strip().split(' ')[1:]]
for line in external_embedding_fp
}
external_embedding_fp.close()
self.edim = len(self.external_embedding.values()[0])
self.noextrn = [0.0 for _ in xrange(self.edim)]
self.extrnd = {
word: i + 3
for i, word in enumerate(self.external_embedding)
}
self.elookup = self.model.add_lookup_parameters(
(len(self.external_embedding) + 3, self.edim))
for word, i in self.extrnd.iteritems():
self.elookup.init_row(i, self.external_embedding[word])
self.extrnd['*PAD*'] = 1
self.extrnd['*INITIAL*'] = 2
print 'Load external embedding. Vector dimensions', self.edim
if self.bibiFlag:
self.builders = [
VanillaLSTMBuilder(1, self.wdims + self.edim + self.cdims * 2,
self.ldims, self.model),
VanillaLSTMBuilder(1, self.wdims + self.edim + self.cdims * 2,
self.ldims, self.model)
]
self.bbuilders = [
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model),
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model)
]
elif self.layers > 0:
self.builders = [
VanillaLSTMBuilder(self.layers, self.wdims + self.edim,
self.ldims, self.model),
VanillaLSTMBuilder(self.layers, self.wdims + self.edim,
self.ldims, self.model)
]
else:
self.builders = [
SimpleRNNBuilder(1, self.wdims + self.edim + self.cdims * 2,
self.ldims, self.model),
SimpleRNNBuilder(1, self.wdims + self.edim + self.cdims * 2,
self.ldims, self.model)
]
self.ffSeqPredictor = FFSequencePredictor(
Layer(self.model, self.ldims * 2, len(self.pos), softmax))
self.hidden_units = options.hidden_units
self.hidden2_units = options.hidden2_units
self.vocab['*PAD*'] = 1
self.vocab['*INITIAL*'] = 2
self.wlookup = self.model.add_lookup_parameters(
(len(vocab) + 3, self.wdims))
self.clookup = self.model.add_lookup_parameters((len(c2i), self.cdims))
self.hidLayerFOH = self.model.add_parameters(
(self.hidden_units, self.ldims * 2))
self.hidLayerFOM = self.model.add_parameters(
(self.hidden_units, self.ldims * 2))
self.hidBias = self.model.add_parameters((self.hidden_units))
self.hid2Layer = self.model.add_parameters(
(self.hidden2_units, self.hidden_units))
self.hid2Bias = self.model.add_parameters((self.hidden2_units))
self.outLayer = self.model.add_parameters(
(1, self.hidden2_units
if self.hidden2_units > 0 else self.hidden_units))
if self.labelsFlag:
self.rhidLayerFOH = self.model.add_parameters(
(self.hidden_units, 2 * self.ldims))
self.rhidLayerFOM = self.model.add_parameters(
(self.hidden_units, 2 * self.ldims))
self.rhidBias = self.model.add_parameters((self.hidden_units))
self.rhid2Layer = self.model.add_parameters(
(self.hidden2_units, self.hidden_units))
self.rhid2Bias = self.model.add_parameters((self.hidden2_units))
self.routLayer = self.model.add_parameters(
(len(self.irels), self.hidden2_units
if self.hidden2_units > 0 else self.hidden_units))
self.routBias = self.model.add_parameters((len(self.irels)))
self.char_rnn = RNNSequencePredictor(
LSTMBuilder(1, self.cdims, self.cdims, self.model))
def __getExpr(self, sentence, i, j, train):
if sentence[i].headfov is None:
sentence[i].headfov = self.hidLayerFOH.expr() * concatenate(
[sentence[i].lstms[0], sentence[i].lstms[1]])
if sentence[j].modfov is None:
sentence[j].modfov = self.hidLayerFOM.expr() * concatenate(
[sentence[j].lstms[0], sentence[j].lstms[1]])
if self.hidden2_units > 0:
output = self.outLayer.expr() * self.activation(
self.hid2Bias.expr() + self.hid2Layer.expr() *
self.activation(sentence[i].headfov + sentence[j].modfov +
self.hidBias.expr())) # + self.outBias
else:
output = self.outLayer.expr() * self.activation(
sentence[i].headfov + sentence[j].modfov +
self.hidBias.expr()) # + self.outBias
return output
def __evaluate(self, sentence, train):
exprs = [[
self.__getExpr(sentence, i, j, train)
for j in xrange(len(sentence))
] for i in xrange(len(sentence))]
scores = np.array([[output.scalar_value() for output in exprsRow]
for exprsRow in exprs])
return scores, exprs
def pick_neg_log(self, pred, gold):
return -dynet.log(dynet.pick(pred, gold))
def __evaluateLabel(self, sentence, i, j):
if sentence[i].rheadfov is None:
sentence[i].rheadfov = self.rhidLayerFOH.expr() * concatenate(
[sentence[i].lstms[0], sentence[i].lstms[1]])
if sentence[j].rmodfov is None:
sentence[j].rmodfov = self.rhidLayerFOM.expr() * concatenate(
[sentence[j].lstms[0], sentence[j].lstms[1]])
if self.hidden2_units > 0:
output = self.routLayer.expr() * self.activation(
self.rhid2Bias.expr() + self.rhid2Layer.expr() *
self.activation(sentence[i].rheadfov + sentence[j].rmodfov +
self.rhidBias.expr())) + self.routBias.expr()
else:
output = self.routLayer.expr() * self.activation(
sentence[i].rheadfov + sentence[j].rmodfov +
self.rhidBias.expr()) + self.routBias.expr()
return output.value(), output
def Save(self, filename):
self.model.save(filename)
def Load(self, filename):
self.model.populate(filename)
def Predict(self, conll_path):
with open(conll_path, 'r') as conllFP:
for iSentence, sentence in enumerate(read_conll(conllFP,
self.c2i)):
conll_sentence = [
entry for entry in sentence
if isinstance(entry, utils.ConllEntry)
]
for entry in conll_sentence:
wordvec = self.wlookup[int(self.vocab.get(
entry.norm, 0))] if self.wdims > 0 else None
evec = self.elookup[int(
self.extrnd.get(entry.form,
self.extrnd.get(entry.norm, 0))
)] if self.external_embedding is not None else None
last_state = self.char_rnn.predict_sequence(
[self.clookup[c] for c in entry.idChars])[-1]
rev_last_state = self.char_rnn.predict_sequence(
[self.clookup[c] for c in reversed(entry.idChars)])[-1]
entry.vec = concatenate(
filter(None,
[wordvec, evec, last_state, rev_last_state]))
entry.lstms = [entry.vec, entry.vec]
entry.headfov = None
entry.modfov = None
entry.rheadfov = None
entry.rmodfov = None
if self.blstmFlag:
lstm_forward = self.builders[0].initial_state()
lstm_backward = self.builders[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.lstms[1] = lstm_forward.output()
rentry.lstms[0] = lstm_backward.output()
if self.bibiFlag:
for entry in conll_sentence:
entry.vec = concatenate(entry.lstms)
blstm_forward = self.bbuilders[0].initial_state()
blstm_backward = self.bbuilders[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.vec)
blstm_backward = blstm_backward.add_input(
rentry.vec)
entry.lstms[1] = blstm_forward.output()
rentry.lstms[0] = blstm_backward.output()
scores, exprs = self.__evaluate(conll_sentence, True)
heads = decoder.parse_proj(scores)
#Multiple roots: heading to the previous "rooted" one
rootCount = 0
rootWid = -1
for index, head in enumerate(heads):
if head == 0:
rootCount += 1
if rootCount == 1:
rootWid = index
if rootCount > 1:
heads[index] = rootWid
rootWid = index
concat_layer = [
concatenate(entry.lstms) for entry in conll_sentence
]
outputFFlayer = self.ffSeqPredictor.predict_sequence(
concat_layer)
predicted_pos_indices = [
np.argmax(o.value()) for o in outputFFlayer
]
predicted_postags = [
self.id2pos[idx] for idx in predicted_pos_indices
]
for entry, head, pos in zip(conll_sentence, heads,
predicted_postags):
entry.pred_parent_id = head
entry.pred_relation = '_'
entry.pred_pos = pos
dump = False
if self.labelsFlag:
for modifier, head in enumerate(heads[1:]):
scores, exprs = self.__evaluateLabel(
conll_sentence, head, modifier + 1)
conll_sentence[modifier +
1].pred_relation = self.irels[max(
enumerate(scores),
key=itemgetter(1))[0]]
renew_cg()
if not dump:
yield sentence
def Train(self, conll_path):
errors = 0
batch = 0
eloss = 0.0
mloss = 0.0
eerrors = 0
etotal = 0
start = time.time()
with open(conll_path, 'r') as conllFP:
shuffledData = list(read_conll(conllFP, self.c2i))
random.shuffle(shuffledData)
errs = []
lerrs = []
posErrs = []
eeloss = 0.0
for iSentence, sentence in enumerate(shuffledData):
if iSentence % 500 == 0 and iSentence != 0:
print "Processing sentence number: %d" % iSentence, ", Loss: %.2f" % (
eloss / etotal), ", Time: %.2f" % (time.time() - start)
start = time.time()
eerrors = 0
eloss = 0.0
etotal = 0
lerrors = 0
ltotal = 0
conll_sentence = [
entry for entry in sentence
if isinstance(entry, utils.ConllEntry)
]
for entry in conll_sentence:
c = float(self.wordsCount.get(entry.norm, 0))
dropFlag = (random.random() < (c / (0.25 + c)))
wordvec = self.wlookup[
int(self.vocab.get(entry.norm, 0)
) if dropFlag else 0] if self.wdims > 0 else None
evec = None
if self.external_embedding is not None:
evec = self.elookup[self.extrnd.get(
entry.form, self.extrnd.get(entry.norm, 0)) if
(dropFlag or
(random.random() < 0.5)) else 0]
#entry.vec = concatenate(filter(None, [wordvec, evec]))
last_state = self.char_rnn.predict_sequence(
[self.clookup[c] for c in entry.idChars])[-1]
rev_last_state = self.char_rnn.predict_sequence(
[self.clookup[c] for c in reversed(entry.idChars)])[-1]
entry.vec = concatenate([
dynet.noise(fe, 0.2) for fe in filter(
None, [wordvec, evec, last_state, rev_last_state])
])
entry.lstms = [entry.vec, entry.vec]
entry.headfov = None
entry.modfov = None
entry.rheadfov = None
entry.rmodfov = None
if self.blstmFlag:
lstm_forward = self.builders[0].initial_state()
lstm_backward = self.builders[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.lstms[1] = lstm_forward.output()
rentry.lstms[0] = lstm_backward.output()
if self.bibiFlag:
for entry in conll_sentence:
entry.vec = concatenate(entry.lstms)
blstm_forward = self.bbuilders[0].initial_state()
blstm_backward = self.bbuilders[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.vec)
blstm_backward = blstm_backward.add_input(
rentry.vec)
entry.lstms[1] = blstm_forward.output()
rentry.lstms[0] = blstm_backward.output()
scores, exprs = self.__evaluate(conll_sentence, True)
gold = [entry.parent_id for entry in conll_sentence]
heads = decoder.parse_proj(scores,
gold if self.costaugFlag else None)
if self.labelsFlag:
for modifier, head in enumerate(gold[1:]):
rscores, rexprs = self.__evaluateLabel(
conll_sentence, head, modifier + 1)
goldLabelInd = self.rels[conll_sentence[modifier +
1].relation]
wrongLabelInd = max(((l, scr)
for l, scr in enumerate(rscores)
if l != goldLabelInd),
key=itemgetter(1))[0]
if rscores[goldLabelInd] < rscores[wrongLabelInd] + 1:
lerrs.append(rexprs[wrongLabelInd] -
rexprs[goldLabelInd])
e = sum([1 for h, g in zip(heads[1:], gold[1:]) if h != g])
eerrors += e
if e > 0:
loss = [(exprs[h][i] - exprs[g][i])
for i, (h, g) in enumerate(zip(heads, gold))
if h != g] # * (1.0/float(e))
eloss += (e)
mloss += (e)
errs.extend(loss)
etotal += len(conll_sentence)
concat_layer = [
concatenate(entry.lstms) for entry in conll_sentence
]
concat_layer = [dynet.noise(fe, 0.2) for fe in concat_layer]
outputFFlayer = self.ffSeqPredictor.predict_sequence(
concat_layer)
posIDs = [self.pos.get(entry.pos) for entry in conll_sentence]
for pred, gold in zip(outputFFlayer, posIDs):
posErrs.append(self.pick_neg_log(pred, gold))
if iSentence % 1 == 0 or len(errs) > 0 or len(
lerrs) > 0 or len(posErrs) > 0:
eeloss = 0.0
if len(errs) > 0 or len(lerrs) > 0 or len(posErrs) > 0:
eerrs = (esum(errs + lerrs + posErrs)
) #* (1.0/(float(len(errs))))
eerrs.scalar_value()
eerrs.backward()
self.trainer.update()
errs = []
lerrs = []
posErrs = []
renew_cg()
if len(errs) > 0:
eerrs = (esum(errs + lerrs + posErrs)) #* (1.0/(float(len(errs))))
eerrs.scalar_value()
eerrs.backward()
self.trainer.update()
errs = []
lerrs = []
posErrs = []
eeloss = 0.0
renew_cg()
self.trainer.update()
print "Loss: %.2f" % (mloss / iSentence)
class jPosDepLearner:
def __init__(self, vocab, pos, rels, w2i, c2i, options):
self.model = ParameterCollection()
random.seed(1)
self.trainer = AdamTrainer(self.model)
#if options.learning_rate is not None:
# self.trainer = AdamTrainer(self.model, alpha=options.learning_rate)
# print("Adam initial learning rate:", options.learning_rate)
self.activations = {'tanh': tanh, 'sigmoid': logistic, 'relu': rectify,
'tanh3': (lambda x: tanh(cwise_multiply(cwise_multiply(x, x), x)))}
self.activation = self.activations[options.activation]
self.blstmFlag = options.blstmFlag
self.labelsFlag = options.labelsFlag
self.costaugFlag = options.costaugFlag
self.bibiFlag = options.bibiFlag
self.ldims = options.lstm_dims
self.wdims = options.wembedding_dims
self.cdims = options.cembedding_dims
self.layers = options.lstm_layers
self.wordsCount = vocab
self.vocab = {word: ind + 3 for word, ind in w2i.iteritems()}
self.pos = {word: ind for ind, word in enumerate(pos)}
self.id2pos = {ind: word for ind, word in enumerate(pos)}
self.c2i = c2i
self.rels = {word: ind for ind, word in enumerate(rels)}
self.irels = rels
self.pdims = options.pembedding_dims
self.vocab['*PAD*'] = 1
self.vocab['*INITIAL*'] = 2
self.wlookup = self.model.add_lookup_parameters((len(vocab) + 3, self.wdims))
self.clookup = self.model.add_lookup_parameters((len(c2i), self.cdims))
self.plookup = self.model.add_lookup_parameters((len(pos), self.pdims))
if options.external_embedding is not None:
ext_embeddings, ext_emb_dim = load_embeddings_file(options.external_embedding, lower=True)
assert (ext_emb_dim == self.wdims)
print("Initializing word embeddings by pre-trained vectors")
count = 0
for word in self.vocab:
_word = unicode(word, "utf-8")
if _word in ext_embeddings:
count += 1
self.wlookup.init_row(self.vocab[word], ext_embeddings[_word])
print("Vocab size: %d; #words having pretrained vectors: %d" % (len(self.vocab), count))
self.pos_builders = [VanillaLSTMBuilder(1, self.wdims + self.cdims * 2, self.ldims, self.model),
VanillaLSTMBuilder(1, self.wdims + self.cdims * 2, self.ldims, self.model)]
self.pos_bbuilders = [VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model),
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model)]
if self.bibiFlag:
self.builders = [VanillaLSTMBuilder(1, self.wdims + self.cdims * 2 + self.pdims, self.ldims, self.model),
VanillaLSTMBuilder(1, self.wdims + self.cdims * 2 + self.pdims, self.ldims, self.model)]
self.bbuilders = [VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model),
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model)]
elif self.layers > 0:
self.builders = [VanillaLSTMBuilder(self.layers, self.wdims + self.cdims * 2 + self.pdims, self.ldims, self.model),
VanillaLSTMBuilder(self.layers, self.wdims + self.cdims * 2 + self.pdims, self.ldims, self.model)]
else:
self.builders = [SimpleRNNBuilder(1, self.wdims + self.cdims * 2, self.ldims, self.model),
SimpleRNNBuilder(1, self.wdims + self.cdims * 2, self.ldims, self.model)]
self.ffSeqPredictor = FFSequencePredictor(Layer(self.model, self.ldims * 2, len(self.pos), softmax))
self.hidden_units = options.hidden_units
self.hidBias = self.model.add_parameters((self.ldims * 8))
self.hidLayer = self.model.add_parameters((self.hidden_units, self.ldims * 8))
self.hid2Bias = self.model.add_parameters((self.hidden_units))
self.outLayer = self.model.add_parameters((1, self.hidden_units if self.hidden_units > 0 else self.ldims * 8))
if self.labelsFlag:
self.rhidBias = self.model.add_parameters((self.ldims * 8))
self.rhidLayer = self.model.add_parameters((self.hidden_units, self.ldims * 8))
self.rhid2Bias = self.model.add_parameters((self.hidden_units))
self.routLayer = self.model.add_parameters(
(len(self.irels), self.hidden_units if self.hidden_units > 0 else self.ldims * 8))
self.routBias = self.model.add_parameters((len(self.irels)))
self.ffRelPredictor = FFSequencePredictor(
Layer(self.model, self.hidden_units if self.hidden_units > 0 else self.ldims * 8, len(self.irels),
softmax))
self.char_rnn = RNNSequencePredictor(LSTMBuilder(1, self.cdims, self.cdims, self.model))
def __getExpr(self, sentence, i, j):
if sentence[i].headfov is None:
sentence[i].headfov = concatenate([sentence[i].lstms[0], sentence[i].lstms[1]])
if sentence[j].modfov is None:
sentence[j].modfov = concatenate([sentence[j].lstms[0], sentence[j].lstms[1]])
_inputVector = concatenate(
[sentence[i].headfov, sentence[j].modfov, dynet.abs(sentence[i].headfov - sentence[j].modfov),
dynet.cmult(sentence[i].headfov, sentence[j].modfov)])
if self.hidden_units > 0:
output = self.outLayer.expr() * self.activation(
self.hid2Bias.expr() + self.hidLayer.expr() * self.activation(
_inputVector + self.hidBias.expr()))
else:
output = self.outLayer.expr() * self.activation(_inputVector + self.hidBias.expr())
return output
def __evaluate(self, sentence):
exprs = [[self.__getExpr(sentence, i, j) for j in xrange(len(sentence))] for i in xrange(len(sentence))]
scores = np.array([[output.scalar_value() for output in exprsRow] for exprsRow in exprs])
return scores, exprs
def pick_neg_log(self, pred, gold):
return -dynet.log(dynet.pick(pred, gold))
def __getRelVector(self, sentence, i, j):
if sentence[i].rheadfov is None:
sentence[i].rheadfov = concatenate([sentence[i].lstms[0], sentence[i].lstms[1]])
if sentence[j].rmodfov is None:
sentence[j].rmodfov = concatenate([sentence[j].lstms[0], sentence[j].lstms[1]])
_outputVector = concatenate(
[sentence[i].rheadfov, sentence[j].rmodfov, abs(sentence[i].rheadfov - sentence[j].rmodfov),
cmult(sentence[i].rheadfov, sentence[j].rmodfov)])
if self.hidden_units > 0:
return self.rhid2Bias.expr() + self.rhidLayer.expr() * self.activation(
_outputVector + self.rhidBias.expr())
else:
return _outputVector
def Save(self, filename):
self.model.save(filename)
def Load(self, filename):
self.model.populate(filename)
def PredictFromText(self, text_file,output_file):
#pdb.set_trace()
#with open(output_file,'w',buffering=20*(1024**2)) as ofp:
# with open(text_file, 'r',buffering=20*(1024**2)) as t_fp:
with open(output_file,'w') as ofp:
with open(text_file, 'r') as t_fp:
try:
for line in t_fp:
itf = tempfile.NamedTemporaryFile(delete=False)
otf = tempfile.NamedTemporaryFile(delete=False)
with open(itf.name,'w'):
itf.write(line)
itf.flush()
conllConvertToFile(itf.name,otf.name)
conll_gen = self.Predict(otf.name)
for sentence in conll_gen:
for entry in sentence[1:]:
fields = str(entry).split('\t')
if (len(fields) > 1):
ofp.write(fields[0] + '\t' + fields[1] + '\t' + fields[3] + '\t' + fields[6] + '\t' + fields[7] + '\n')
else:
ofp.write(str(entry) + '\n')
ofp.write('\n')
os.remove(itf.name)
os.remove(otf.name)
except RuntimeError as re:
print("Unexpected error:", sys.exc_info()[0])
traceback.print_exc(file=sys.stdout)
#instead of pass clear
sys.exc_clear()
def Predict(self, conll_path):
with open(conll_path, 'r') as conllFP:
for iSentence, sentence in enumerate(read_conll_predict(conllFP, self.c2i, self.wordsCount)):
conll_sentence = [entry for entry in sentence if isinstance(entry, utils.ConllEntry)]
for entry in conll_sentence:
wordvec = self.wlookup[int(self.vocab.get(entry.norm, 0))] if self.wdims > 0 else None
last_state = self.char_rnn.predict_sequence([self.clookup[c] for c in entry.idChars])[-1]
rev_last_state = self.char_rnn.predict_sequence([self.clookup[c] for c in reversed(entry.idChars)])[
-1]
entry.vec = concatenate(filter(None, [wordvec, last_state, rev_last_state]))
entry.pos_lstms = [entry.vec, entry.vec]
entry.headfov = None
entry.modfov = None
entry.rheadfov = None
entry.rmodfov = None
#Predicted pos tags
lstm_forward = self.pos_builders[0].initial_state()
lstm_backward = self.pos_builders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.pos_lstms[1] = lstm_forward.output()
rentry.pos_lstms[0] = lstm_backward.output()
for entry in conll_sentence:
entry.pos_vec = concatenate(entry.pos_lstms)
blstm_forward = self.pos_bbuilders[0].initial_state()
blstm_backward = self.pos_bbuilders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.pos_vec)
blstm_backward = blstm_backward.add_input(rentry.pos_vec)
entry.pos_lstms[1] = blstm_forward.output()
rentry.pos_lstms[0] = blstm_backward.output()
concat_layer = [concatenate(entry.pos_lstms) for entry in conll_sentence]
outputFFlayer = self.ffSeqPredictor.predict_sequence(concat_layer)
predicted_pos_indices = [np.argmax(o.value()) for o in outputFFlayer]
predicted_postags = [self.id2pos[idx] for idx in predicted_pos_indices]
# Add predicted pos tags for parsing prediction
for entry, posid in zip(conll_sentence, predicted_pos_indices):
entry.vec = concatenate([entry.vec, self.plookup[posid]])
entry.lstms = [entry.vec, entry.vec]
if self.blstmFlag:
lstm_forward = self.builders[0].initial_state()
lstm_backward = self.builders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.lstms[1] = lstm_forward.output()
rentry.lstms[0] = lstm_backward.output()
if self.bibiFlag:
for entry in conll_sentence:
entry.vec = concatenate(entry.lstms)
blstm_forward = self.bbuilders[0].initial_state()
blstm_backward = self.bbuilders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.vec)
blstm_backward = blstm_backward.add_input(rentry.vec)
entry.lstms[1] = blstm_forward.output()
rentry.lstms[0] = blstm_backward.output()
scores, exprs = self.__evaluate(conll_sentence)
heads = decoder.parse_proj(scores)
# Multiple roots: heading to the previous "rooted" one
rootCount = 0
rootWid = -1
for index, head in enumerate(heads):
if head == 0:
rootCount += 1
if rootCount == 1:
rootWid = index
if rootCount > 1:
heads[index] = rootWid
rootWid = index
for entry, head, pos in zip(conll_sentence, heads, predicted_postags):
entry.pred_parent_id = head
entry.pred_relation = '_'
entry.pred_pos = pos
dump = False
if self.labelsFlag:
concat_layer = [self.__getRelVector(conll_sentence, head, modifier + 1) for modifier, head in
enumerate(heads[1:])]
outputFFlayer = self.ffRelPredictor.predict_sequence(concat_layer)
predicted_rel_indices = [np.argmax(o.value()) for o in outputFFlayer]
predicted_rels = [self.irels[idx] for idx in predicted_rel_indices]
for modifier, head in enumerate(heads[1:]):
conll_sentence[modifier + 1].pred_relation = predicted_rels[modifier]
renew_cg()
if not dump:
yield sentence
def Train(self, conll_path):
eloss = 0.0
mloss = 0.0
eerrors = 0
etotal = 0
start = time.time()
with open(conll_path, 'r') as conllFP:
shuffledData = list(read_conll(conllFP, self.c2i))
random.shuffle(shuffledData)
errs = []
lerrs = []
posErrs = []
for iSentence, sentence in enumerate(shuffledData):
if iSentence % 500 == 0 and iSentence != 0:
print "Processing sentence number: %d" % iSentence, ", Loss: %.4f" % (
eloss / etotal), ", Time: %.2f" % (time.time() - start)
start = time.time()
eerrors = 0
eloss = 0.0
etotal = 0
conll_sentence = [entry for entry in sentence if isinstance(entry, utils.ConllEntry)]
for entry in conll_sentence:
c = float(self.wordsCount.get(entry.norm, 0))
dropFlag = (random.random() < (c / (0.25 + c)))
wordvec = self.wlookup[
int(self.vocab.get(entry.norm, 0)) if dropFlag else 0] if self.wdims > 0 else None
last_state = self.char_rnn.predict_sequence([self.clookup[c] for c in entry.idChars])[-1]
rev_last_state = self.char_rnn.predict_sequence([self.clookup[c] for c in reversed(entry.idChars)])[
-1]
entry.vec = dynet.dropout(concatenate(filter(None, [wordvec, last_state, rev_last_state])), 0.33)
entry.pos_lstms = [entry.vec, entry.vec]
entry.headfov = None
entry.modfov = None
entry.rheadfov = None
entry.rmodfov = None
#POS tagging loss
lstm_forward = self.pos_builders[0].initial_state()
lstm_backward = self.pos_builders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.pos_lstms[1] = lstm_forward.output()
rentry.pos_lstms[0] = lstm_backward.output()
for entry in conll_sentence:
entry.pos_vec = concatenate(entry.pos_lstms)
blstm_forward = self.pos_bbuilders[0].initial_state()
blstm_backward = self.pos_bbuilders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.pos_vec)
blstm_backward = blstm_backward.add_input(rentry.pos_vec)
entry.pos_lstms[1] = blstm_forward.output()
rentry.pos_lstms[0] = blstm_backward.output()
concat_layer = [dynet.dropout(concatenate(entry.pos_lstms), 0.33) for entry in conll_sentence]
outputFFlayer = self.ffSeqPredictor.predict_sequence(concat_layer)
posIDs = [self.pos.get(entry.pos) for entry in conll_sentence]
for pred, gold in zip(outputFFlayer, posIDs):
posErrs.append(self.pick_neg_log(pred, gold))
# Add predicted pos tags
for entry, poses in zip(conll_sentence, outputFFlayer):
entry.vec = concatenate([entry.vec, dynet.dropout(self.plookup[np.argmax(poses.value())], 0.33)])
entry.lstms = [entry.vec, entry.vec]
#Parsing losses
if self.blstmFlag:
lstm_forward = self.builders[0].initial_state()
lstm_backward = self.builders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.lstms[1] = lstm_forward.output()
rentry.lstms[0] = lstm_backward.output()
if self.bibiFlag:
for entry in conll_sentence:
entry.vec = concatenate(entry.lstms)
blstm_forward = self.bbuilders[0].initial_state()
blstm_backward = self.bbuilders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.vec)
blstm_backward = blstm_backward.add_input(rentry.vec)
entry.lstms[1] = blstm_forward.output()
rentry.lstms[0] = blstm_backward.output()
scores, exprs = self.__evaluate(conll_sentence)
gold = [entry.parent_id for entry in conll_sentence]
heads = decoder.parse_proj(scores, gold if self.costaugFlag else None)
if self.labelsFlag:
concat_layer = [dynet.dropout(self.__getRelVector(conll_sentence, head, modifier + 1), 0.33) for
modifier, head in enumerate(gold[1:])]
outputFFlayer = self.ffRelPredictor.predict_sequence(concat_layer)
relIDs = [self.rels[conll_sentence[modifier + 1].relation] for modifier, _ in enumerate(gold[1:])]
for pred, goldid in zip(outputFFlayer, relIDs):
lerrs.append(self.pick_neg_log(pred, goldid))
e = sum([1 for h, g in zip(heads[1:], gold[1:]) if h != g])
eerrors += e
if e > 0:
loss = [(exprs[h][i] - exprs[g][i]) for i, (h, g) in enumerate(zip(heads, gold)) if h != g] # * (1.0/float(e))
eloss += (e)
mloss += (e)
errs.extend(loss)
etotal += len(conll_sentence)
if iSentence % 1 == 0:
if len(errs) > 0 or len(lerrs) > 0 or len(posErrs) > 0:
eerrs = (esum(errs + lerrs + posErrs))
eerrs.scalar_value()
eerrs.backward()
self.trainer.update()
errs = []
lerrs = []
posErrs = []
renew_cg()
print "Loss: %.4f" % (mloss / iSentence)
class jPosDepLearner:
def __init__(self, vocab, pos, rels, w2i, c2i, caps, options):
self.model = ParameterCollection()
random.seed(1)
self.trainer = AdamTrainer(self.model)
#if options.learning_rate is not None:
# self.trainer = AdamTrainer(self.model, alpha=options.learning_rate)
# print("Adam initial learning rate:", options.learning_rate)
self.activations = {
'tanh': tanh,
'sigmoid': logistic,
'relu': rectify,
'tanh3': (lambda x: tanh(cwise_multiply(cwise_multiply(x, x), x)))
}
self.activation = self.activations[options.activation]
self.blstmFlag = options.blstmFlag
self.labelsFlag = options.labelsFlag
self.costaugFlag = options.costaugFlag
self.bibiFlag = options.bibiFlag
self.depFlag = options.depFlag
self.sNerFlag = options.sNerFlag
self.ldims = options.lstm_dims
self.wdims = options.wembedding_dims
self.cdims = options.cembedding_dims
self.reldims = options.relembedding_dims
self.layers = options.lstm_layers
self.wordsCount = vocab
self.vocab = {word: ind + 3 for word, ind in w2i.iteritems()}
self.pos = {word: ind for ind, word in enumerate(pos)}
self.id2pos = {ind: word for ind, word in enumerate(pos)}
self.c2i = c2i
self.caps = {word: ind for ind, word in enumerate(caps)}
self.rels = {word: ind for ind, word in enumerate(rels)}
self.irels = rels
self.pdims = options.pembedding_dims
self.vocab['*PAD*'] = 1
self.vocab['*INITIAL*'] = 2
self.wlookup = self.model.add_lookup_parameters(
(len(vocab) + 3, self.wdims))
self.clookup = self.model.add_lookup_parameters((len(c2i), self.cdims))
self.plookup = self.model.add_lookup_parameters((len(pos), self.pdims))
self.caps_lookup = self.model.add_lookup_parameters(
(len(caps), self.cdims))
transition_array = np.random.rand(len(pos) + 2, len(pos) + 2)
#cap_array=np.random.rand(len(caps),len(pos))
def normalizeprobs(arr):
return np.array([np.divide(arr1, sum(arr1)) for arr1 in arr])
self.nertrans_lookup = self.model.add_lookup_parameters(
(len(pos) + 2, len(pos) + 2))
#self.caplookup = self.model.lookup_parameters_from_numpy(normalizeprobs(cap_array))
if options.external_embedding is not None:
ext_embeddings, ext_emb_dim = load_embeddings_file(
options.external_embedding, lower=True)
assert (ext_emb_dim == self.wdims)
print("Initializing word embeddings by pre-trained vectors")
count = 0
for word in self.vocab:
_word = unicode(word, "utf-8")
if _word in ext_embeddings:
count += 1
self.wlookup.init_row(self.vocab[word],
ext_embeddings[_word])
print("Vocab size: %d; #words having pretrained vectors: %d" %
(len(self.vocab), count))
self.ffSeqPredictor = FFSequencePredictor(
Layer(self.model, self.ldims * 2, len(self.pos), "idt"))
self.hidden_units = options.hidden_units
if not self.depFlag:
self.pos_builders = [
VanillaLSTMBuilder(1, self.wdims + self.cdims * 3, self.ldims,
self.model),
VanillaLSTMBuilder(1, self.wdims + self.cdims * 3, self.ldims,
self.model)
]
self.pos_bbuilders = [
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model),
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model)
]
self.ffSeqPredictor = FFSequencePredictor(
Layer(self.model, self.ldims * 2, len(self.pos), softmax))
self.hidden_units = options.hidden_units
if self.depFlag:
if self.bibiFlag:
self.builders = [
VanillaLSTMBuilder(1, self.wdims + self.cdims * 3,
self.ldims, self.model),
VanillaLSTMBuilder(1, self.wdims + self.cdims * 3,
self.ldims, self.model)
]
self.bbuilders = [
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims,
self.model),
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims,
self.model)
]
elif self.layers > 0:
self.builders = [
VanillaLSTMBuilder(self.layers,
self.wdims + self.cdims * 3, self.ldims,
self.model),
VanillaLSTMBuilder(self.layers,
self.wdims + self.cdims * 3, self.ldims,
self.model)
]
else:
self.builders = [
SimpleRNNBuilder(1, self.wdims + self.cdims * 3,
self.ldims, self.model),
SimpleRNNBuilder(1, self.wdims + self.cdims * 3,
self.ldims, self.model)
]
self.hidBias = self.model.add_parameters((self.ldims * 8))
self.hidLayer = self.model.add_parameters(
(self.hidden_units, self.ldims * 8))
self.hid2Bias = self.model.add_parameters((self.hidden_units))
self.outLayer = self.model.add_parameters(
(1,
self.hidden_units if self.hidden_units > 0 else self.ldims *
8))
if self.labelsFlag:
self.rhidBias = self.model.add_parameters((self.ldims * 8))
self.rhidLayer = self.model.add_parameters(
(self.hidden_units, self.ldims * 8))
self.rhid2Bias = self.model.add_parameters((self.hidden_units))
self.routLayer = self.model.add_parameters(
(len(self.irels), self.hidden_units
if self.hidden_units > 0 else self.ldims * 8))
self.routBias = self.model.add_parameters((len(self.irels)))
self.ffRelPredictor = FFSequencePredictor(
Layer(
self.model, self.hidden_units
if self.hidden_units > 0 else self.ldims * 8,
len(self.irels), softmax))
if self.sNerFlag:
self.sner_builders = [
VanillaLSTMBuilder(
1, self.wdims + self.cdims * 3 + self.reldims,
self.ldims, self.model),
VanillaLSTMBuilder(
1, self.wdims + self.cdims * 3 + self.reldims,
self.ldims, self.model)
]
self.sner_bbuilders = [
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims,
self.model),
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims,
self.model)
]
##relation embeddings
self.rellookup = self.model.add_lookup_parameters(
(len(self.rels), self.reldims))
self.char_rnn = RNNSequencePredictor(
LSTMBuilder(1, self.cdims, self.cdims, self.model))
def __getExpr(self, sentence, i, j):
if sentence[i].headfov is None:
sentence[i].headfov = concatenate(
[sentence[i].lstms[0], sentence[i].lstms[1]])
if sentence[j].modfov is None:
sentence[j].modfov = concatenate(
[sentence[j].lstms[0], sentence[j].lstms[1]])
_inputVector = concatenate([
sentence[i].headfov, sentence[j].modfov,
dynet.abs(sentence[i].headfov - sentence[j].modfov),
dynet.cmult(sentence[i].headfov, sentence[j].modfov)
])
if self.hidden_units > 0:
output = self.outLayer.expr() * self.activation(
self.hid2Bias.expr() + self.hidLayer.expr() *
self.activation(_inputVector + self.hidBias.expr()))
else:
output = self.outLayer.expr() * self.activation(
_inputVector + self.hidBias.expr())
return output
def __evaluate(self, sentence):
exprs = [[
self.__getExpr(sentence, i, j) for j in xrange(len(sentence))
] for i in xrange(len(sentence))]
scores = np.array([[output.scalar_value() for output in exprsRow]
for exprsRow in exprs])
return scores, exprs
def pick_neg_log(self, pred, gold):
return -dynet.log(dynet.pick(pred, gold))
def pick_neg_log_2(self, pred_param, gold):
gold_arr = inputVector(
[1 if gold == i else 0 for i in range(len(self.pos) + 2)])
x = scalarInput(1)
pred_arr = softmax(pred_param * x)
return -dynet.log(transpose(pred_arr) * gold_arr)
def pick_gold_score(self, preds, golds):
score = 0
prev_tag = len(self.pos)
for pred, gold in zip(preds, golds):
score += dynet.pick(pred, gold) + dynet.pick(
self.nertrans_lookup[gold], prev_tag)
prev_tag = gold
score += dynet.pick(self.nertrans_lookup[len(self.pos) + 1], prev_tag)
return score
def pick_crf_score(self, preds, golds):
return dynet.exp(
self.pick_gold_tag_score(preds, golds) +
self.pick_gold_trans_score(golds))
def forward_score(self, preds):
def log_sum_exp(tag_score_arr):
argmax = np.argmax(tag_score_arr.value())
max_score = tag_score_arr[argmax]
score = max_score
max_arr = dynet.concatenate(
[max_score for i in range(len(self.pos) + 2)])
score += dynet.log(
dynet.sum_dim(dynet.exp(tag_score_arr - max_arr), [0]))
return score
score = 0
len1 = len(self.pos) + 2
for_score = [-1e10 for i in range(len1)]
for_score[-2] = 0
#print(len(preds))
for i, pred in enumerate(preds):
tag_scores = [dynet.scalarInput(-1e10) for j in range(len1)]
for i, score in enumerate(pred):
tag_score = dynet.concatenate([
score + dynet.pick(self.nertrans_lookup[i], prev_tag) +
for_score[prev_tag] for prev_tag in range(len1)
])
log_1 = log_sum_exp(tag_score)
tag_scores[i] = log_1
#print("tag score: %f"%log_1.value())
for_score = tag_scores
#print(dynet.concatenate(for_score).value())
term_exp = dynet.concatenate([
score + tr
for score, tr in zip(for_score, self.nertrans_lookup[len(self.pos)
+ 1])
])
term_score = log_sum_exp(term_exp)
#print("score : %f"%term_score.value())
return term_score
def nextPerm(self, perm1, taglen):
a = []
for ind, x in enumerate(reversed(perm1)):
if x < taglen - 1:
for i in range(len(perm1) - ind - 1):
a.append(perm1[i])
a.append(x + 1)
for i in range(ind):
a.append(0)
return a
return -1
## takes toooo long
def forward_score2(self, taglen, senlen, preds):
score = 0
perm1 = [0 for i in range(senlen)]
while perm1 != -1:
score += self.pick_crf_score(preds, perm1)
perm1 = self.nextPerm(perm1, taglen)
return score
def __getRelVector(self, sentence, i, j):
if sentence[i].rheadfov is None:
sentence[i].rheadfov = concatenate(
[sentence[i].lstms[0], sentence[i].lstms[1]])
if sentence[j].rmodfov is None:
sentence[j].rmodfov = concatenate(
[sentence[j].lstms[0], sentence[j].lstms[1]])
_outputVector = concatenate([
sentence[i].rheadfov, sentence[j].rmodfov,
abs(sentence[i].rheadfov - sentence[j].rmodfov),
cmult(sentence[i].rheadfov, sentence[j].rmodfov)
])
if self.hidden_units > 0:
return self.rhid2Bias.expr() + self.rhidLayer.expr(
) * self.activation(_outputVector + self.rhidBias.expr())
else:
return _outputVector
def Save(self, filename):
self.model.save(filename)
def Load(self, filename):
self.model.populate(filename)
def Predict(self, conll_path, dep_epoch=1, ner_epoch=1):
with open(conll_path, 'r') as conllFP:
if ner_epoch == 0:
read_conll_nerdep = read_conll_predict(conllFP, self.c2i,
self.wordsCount)
else:
read_conll_nerdep = read_conll_predict_ner(
conllFP, self.c2i, self.wordsCount)
for iSentence, sentence in enumerate(read_conll_nerdep):
conll_sentence = [
entry for entry in sentence
if isinstance(entry, utils.ConllEntry)
]
for entry in conll_sentence:
capvec = self.caps_lookup[entry.capInfo]
wordvec = self.wlookup[int(self.vocab.get(
entry.norm, 0))] if self.wdims > 0 else None
last_state = self.char_rnn.predict_sequence(
[self.clookup[c] for c in entry.idChars])[-1]
rev_last_state = self.char_rnn.predict_sequence(
[self.clookup[c] for c in reversed(entry.idChars)])[-1]
entry.vec = concatenate(
filter(None,
[wordvec, last_state, rev_last_state, capvec]))
entry.vec2 = concatenate(
filter(None,
[wordvec, last_state, rev_last_state, capvec]))
entry.pos_lstms = [entry.vec, entry.vec]
entry.headfov = None
entry.modfov = None
entry.rheadfov = None
entry.rmodfov = None
if not self.depFlag:
#Predicted pos tags
lstm_forward = self.pos_builders[0].initial_state()
lstm_backward = self.pos_builders[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.pos_lstms[1] = lstm_forward.output()
rentry.pos_lstms[0] = lstm_backward.output()
for entry in conll_sentence:
entry.pos_vec = concatenate(entry.pos_lstms)
blstm_forward = self.pos_bbuilders[0].initial_state()
blstm_backward = self.pos_bbuilders[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.pos_vec)
blstm_backward = blstm_backward.add_input(
rentry.pos_vec)
entry.pos_lstms[1] = blstm_forward.output()
rentry.pos_lstms[0] = blstm_backward.output()
concat_layer = [
concatenate(entry.pos_lstms)
for entry in conll_sentence
]
#cap_info_sentence=[self.caplookup[entry.capInfo] for entry in conll_sentence]
outputFFlayer = self.ffSeqPredictor.predict_sequence(
concat_layer)
best_parentids, bestscores = self.ffSeqPredictor.viterbi_sequence(
outputFFlayer, self.nertrans_lookup)
predicted_pos_indices = [
np.argmax(o.value()) for o in outputFFlayer
]
root_predicted_postags = ["O"]
predicted_postags = [
self.id2pos[idx] for idx in best_parentids
]
for pos in predicted_postags:
root_predicted_postags.append(pos)
if iSentence < 5:
for word, tag in zip(conll_sentence,
root_predicted_postags):
print("word : {} gold : {} pred : {}".format(
word.form, word.pos, tag))
for entry, pos in zip(conll_sentence,
root_predicted_postags):
entry.pred_pos = pos
dump = False
if self.depFlag:
# Add predicted pos tags for parsing prediction
#for entry, posid in zip(conll_sentence, viterbi_pred_tagids):
# entry.vec = concatenate([entry.vec, self.plookup[posid]])
# entry.lstms = [entry.vec, entry.vec]
for entry in conll_sentence:
entry.lstms = [entry.vec, entry.vec]
if self.blstmFlag:
lstm_forward = self.builders[0].initial_state()
lstm_backward = self.builders[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.lstms[1] = lstm_forward.output()
rentry.lstms[0] = lstm_backward.output()
if self.bibiFlag:
for entry in conll_sentence:
entry.vec = concatenate(entry.lstms)
blstm_forward = self.bbuilders[0].initial_state()
blstm_backward = self.bbuilders[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(
entry.vec)
blstm_backward = blstm_backward.add_input(
rentry.vec)
entry.lstms[1] = blstm_forward.output()
rentry.lstms[0] = blstm_backward.output()
scores, exprs = self.__evaluate(conll_sentence)
heads = decoder.parse_proj(scores)
# Multiple roots: heading to the previous "rooted" one
rootCount = 0
rootWid = -1
for index, head in enumerate(heads):
if head == 0:
rootCount += 1
if rootCount == 1:
rootWid = index
if rootCount > 1:
heads[index] = rootWid
rootWid = index
for entry, head in zip(conll_sentence, heads):
entry.pred_parent_id = head
entry.pred_relation = '_'
#entry.pred_pos = pos
if self.labelsFlag:
concat_layer = [
self.__getRelVector(conll_sentence, head,
modifier + 1)
for modifier, head in enumerate(heads[1:])
]
outputFFlayer = self.ffRelPredictor.predict_sequence(
concat_layer)
predicted_rel_indices = [
np.argmax(o.value()) for o in outputFFlayer
]
predicted_rels = [
self.irels[idx] for idx in predicted_rel_indices
]
for modifier, head in enumerate(heads[1:]):
conll_sentence[
modifier +
1].pred_relation = predicted_rels[modifier]
if self.sNerFlag and ner_epoch == 1:
conll_sentence[0].vec = concatenate([
conll_sentence[0].vec2,
self.rellookup[self.rels["rroot"]]
])
for entry, pred in zip(conll_sentence[1:],
predicted_rel_indices):
relvec = self.rellookup[pred]
# for entry, posid in zip(conll_sentence, viterbi_pred_tagids):
entry.vec = concatenate([entry.vec2, relvec])
for entry in conll_sentence:
entry.ner2_lstms = [entry.vec, entry.vec]
slstm_forward = self.sner_builders[0].initial_state()
slstm_backward = self.sner_builders[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
lstm_forward = slstm_forward.add_input(entry.vec)
lstm_backward = slstm_backward.add_input(
rentry.vec)
entry.ner2_lstms[1] = lstm_forward.output()
rentry.ner2_lstms[0] = lstm_backward.output()
for entry in conll_sentence:
entry.ner2_vec = concatenate(entry.ner2_lstms)
sblstm_forward = self.sner_bbuilders[0].initial_state()
sblstm_backward = self.sner_bbuilders[1].initial_state(
)
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
blstm_forward = sblstm_forward.add_input(
entry.ner2_vec)
blstm_backward = sblstm_backward.add_input(
rentry.ner2_vec)
entry.ner2_lstms[1] = blstm_forward.output()
rentry.ner2_lstms[0] = blstm_backward.output()
concat_layer = [
dynet.dropout(concatenate(entry.ner2_lstms), 0.33)
for entry in conll_sentence
]
outputFFlayer = self.ffSeqPredictor.predict_sequence(
concat_layer)
best_parentids, bestscores = self.ffSeqPredictor.viterbi_sequence(
outputFFlayer, self.nertrans_lookup)
predicted_pos_indices = [
np.argmax(o.value()) for o in outputFFlayer
]
root_predicted_postags = ["O"]
predicted_postags = [
self.id2pos[idx] for idx in best_parentids
]
for pos in predicted_postags:
root_predicted_postags.append(pos)
if iSentence < 1:
for word, tag in zip(conll_sentence,
root_predicted_postags):
print("word : {} gold : {} pred : {}".format(
word.form, word.pos, tag))
for entry, pos in zip(conll_sentence,
root_predicted_postags):
entry.pred_pos = pos
dump = False
renew_cg()
if not dump:
yield sentence
def Train(self, conll_path, dep_epoch=0, ner_epoch=0):
eloss = 0.0
mloss = 0.0
eerrors = 0
etotal = 0
start = time.time()
dep_epoch = dep_epoch
ner_epoch = ner_epoch
with open(conll_path, 'r') as conllFP:
if ner_epoch == 0:
read_conll_nerdep = read_conll(conllFP, self.c2i)
else:
read_conll_nerdep = read_conll_ner(conllFP, self.c2i)
shuffledData = list(read_conll_nerdep)
random.shuffle(shuffledData)
errs = []
lerrs = []
posErrs = 0
postrErrs = []
nertr2Errs = []
ner2Errs = dynet.inputVector([0])
startind = 0
e = 0
for iSentence, sentence in enumerate(shuffledData):
if iSentence % 500 == 0 and iSentence != 0:
print "Processing sentence number: %d" % iSentence, ", Loss: %.4f" % (
eloss / etotal), ", Time: %.2f" % (time.time() - start)
start = time.time()
eerrors = 0
eloss = 0.0
etotal = 0
conll_sentence = [
entry for entry in sentence
if isinstance(entry, utils.ConllEntry)
]
for entry in conll_sentence:
c = float(self.wordsCount.get(entry.norm, 0))
dropFlag = (random.random() < (c / (0.25 + c)))
capvec = self.caps_lookup[entry.capInfo]
wordvec = self.wlookup[
int(self.vocab.get(entry.norm, 0)
) if dropFlag else 0] if self.wdims > 0 else None
last_state = self.char_rnn.predict_sequence(
[self.clookup[c] for c in entry.idChars])[-1]
rev_last_state = self.char_rnn.predict_sequence(
[self.clookup[c] for c in reversed(entry.idChars)])[-1]
entry.vec = dynet.dropout(
concatenate(
filter(
None,
[wordvec, last_state, rev_last_state, capvec
])), 0.33)
entry.vec2 = entry.vec
entry.pos_lstms = [entry.vec, entry.vec]
entry.headfov = None
entry.modfov = None
entry.rheadfov = None
entry.rmodfov = None
if not self.depFlag:
#NER tagging loss
lstm_forward = self.pos_builders[0].initial_state()
lstm_backward = self.pos_builders[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.pos_lstms[1] = lstm_forward.output()
rentry.pos_lstms[0] = lstm_backward.output()
for entry in conll_sentence:
entry.pos_vec = concatenate(entry.pos_lstms)
blstm_forward = self.pos_bbuilders[0].initial_state()
blstm_backward = self.pos_bbuilders[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.pos_vec)
blstm_backward = blstm_backward.add_input(
rentry.pos_vec)
entry.pos_lstms[1] = blstm_forward.output()
rentry.pos_lstms[0] = blstm_backward.output()
concat_layer = [
dynet.dropout(concatenate(entry.pos_lstms), 0.33)
for entry in conll_sentence
]
cap_info_sentence = [
self.caps_lookup[entry.capInfo]
for entry in conll_sentence
]
outputFFlayer = self.ffSeqPredictor.predict_sequence(
concat_layer)
posIDs = [
self.pos.get(entry.pos) for entry in conll_sentence
]
posErrs = (self.forward_score(outputFFlayer) -
self.pick_gold_score(outputFFlayer, posIDs))
##dependency Flag
if self.depFlag:
# Add predicted ner tags
#for entry, poses in zip(conll_sentence, outputFFlayer):
# entry.vec = concatenate([entry.vec, dynet.dropout(self.plookup[np.argmax(poses.value())], 0.33)])
for entry in conll_sentence:
entry.lstms = [entry.vec, entry.vec]
#Parsing losses
if self.blstmFlag:
lstm_forward = self.builders[0].initial_state()
lstm_backward = self.builders[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.lstms[1] = lstm_forward.output()
rentry.lstms[0] = lstm_backward.output()
if self.bibiFlag:
for entry in conll_sentence:
entry.vec = concatenate(entry.lstms)
blstm_forward = self.bbuilders[0].initial_state()
blstm_backward = self.bbuilders[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(
entry.vec)
blstm_backward = blstm_backward.add_input(
rentry.vec)
entry.lstms[1] = blstm_forward.output()
rentry.lstms[0] = blstm_backward.output()
scores, exprs = self.__evaluate(conll_sentence)
gold = [entry.parent_id for entry in conll_sentence]
heads = decoder.parse_proj(
scores, gold if self.costaugFlag else None)
if self.labelsFlag:
concat_layer = [
dynet.dropout(
self.__getRelVector(conll_sentence, head,
modifier + 1), 0.33)
for modifier, head in enumerate(gold[1:])
]
outputFFlayer = self.ffRelPredictor.predict_sequence(
concat_layer)
if dep_epoch == 1:
relIDs = [
self.rels[conll_sentence[modifier +
1].relation]
for modifier, _ in enumerate(gold[1:])
]
for pred, goldid in zip(outputFFlayer, relIDs):
lerrs.append(self.pick_neg_log(pred, goldid))
if dep_epoch == 1:
e = sum(
[1 for h, g in zip(heads[1:], gold[1:]) if h != g])
if self.sNerFlag and ner_epoch == 1:
conll_sentence[0].vec = concatenate([
conll_sentence[0].vec2,
self.rellookup[self.rels["rroot"]]
])
for entry, pred in zip(conll_sentence[1:],
outputFFlayer):
relvec = self.rellookup[np.argmax(pred.value())]
entry.vec = concatenate(
[entry.vec2,
dynet.dropout(relvec, 0.33)])
for entry in conll_sentence:
entry.ner2_lstms = [entry.vec, entry.vec]
slstm_forward = self.sner_builders[0].initial_state()
slstm_backward = self.sner_builders[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
lstm_forward = slstm_forward.add_input(entry.vec)
lstm_backward = slstm_backward.add_input(
rentry.vec)
entry.ner2_lstms[1] = lstm_forward.output()
rentry.ner2_lstms[0] = lstm_backward.output()
for entry in conll_sentence:
entry.ner2_vec = concatenate(entry.ner2_lstms)
sblstm_forward = self.sner_bbuilders[0].initial_state()
sblstm_backward = self.sner_bbuilders[1].initial_state(
)
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
blstm_forward = sblstm_forward.add_input(
entry.ner2_vec)
blstm_backward = sblstm_backward.add_input(
rentry.ner2_vec)
entry.ner2_lstms[1] = blstm_forward.output()
rentry.ner2_lstms[0] = blstm_backward.output()
concat_layer = [
dynet.dropout(concatenate(entry.ner2_lstms), 0.33)
for entry in conll_sentence
]
outputFFlayer = self.ffSeqPredictor.predict_sequence(
concat_layer)
posIDs = [
self.pos.get(entry.pos) for entry in conll_sentence
]
gold_score = self.pick_gold_score(
outputFFlayer, posIDs)
ner2Errs = (self.forward_score(outputFFlayer) -
gold_score)
if iSentence < 5:
print("ner and dep loss")
if ner2Errs != 0:
print(ner2Errs.value())
else:
print(0)
if dep_epoch != 0:
print(esum(lerrs).value())
else:
print(0)
eerrors += e
if e > 0:
loss = [(exprs[h][i] - exprs[g][i])
for i, (h, g) in enumerate(zip(heads, gold))
if h != g] # * (1.0/float(e))
eloss += (e)
mloss += (e)
errs.extend(loss)
etotal += len(conll_sentence)
if iSentence % 1 == 0:
if len(errs) > 0 or len(lerrs) > 0 or posErrs > 0 or len(
postrErrs) > 0 or ner2Errs > 0 or len(
nertr2Errs) > 0:
eerrs = 0
if len(errs + lerrs + postrErrs + nertr2Errs) > 0:
eerrs = esum(errs + lerrs + postrErrs + nertr2Errs)
eerrs += (posErrs + ner2Errs)
#print(eerrs.value())
eerrs.scalar_value()
eerrs.backward()
self.trainer.update()
errs = []
e = 0
lerrs = []
posErrs = []
postrErrs = []
ner2Errs = []
nertr2Errs = []
posErrs = 0
ner2Errs = 0
renew_cg()
print "Loss: %.4f" % (mloss / iSentence)
class jPosDepLearner:
def __init__(self, vocab, pos, rels, w2i, c2i, m2i, t2i, morph_dict, options):
self.model = ParameterCollection()
random.seed(1)
self.trainer = AdamTrainer(self.model)
#if options.learning_rate is not None:
# self.trainer = AdamTrainer(self.model, alpha=options.learning_rate)
# print("Adam initial learning rate:", options.learning_rate)
self.activations = {'tanh': tanh, 'sigmoid': logistic, 'relu': rectify,
'tanh3': (lambda x: tanh(cwise_multiply(cwise_multiply(x, x), x)))}
self.activation = self.activations[options.activation]
self.blstmFlag = options.blstmFlag
self.labelsFlag = options.labelsFlag
self.costaugFlag = options.costaugFlag
self.bibiFlag = options.bibiFlag
self.morphFlag = options.morphFlag
self.goldMorphFlag = options.goldMorphFlag
self.morphTagFlag = options.morphTagFlag
self.goldMorphTagFlag = options.goldMorphTagFlag
self.lowerCase = options.lowerCase
self.mtag_encoding_composition_type = options.mtag_encoding_composition_type
self.mtag_encoding_composition_alpha = options.mtag_encoding_composition_alpha
self.ldims = options.lstm_dims
self.wdims = options.wembedding_dims
self.mdims = options.membedding_dims
self.tdims = options.tembedding_dims
self.cdims = options.cembedding_dims
self.layers = options.lstm_layers
self.wordsCount = vocab
self.vocab = {word: ind + 3 for word, ind in iter(w2i.items())}
self.pos = {word: ind for ind, word in enumerate(pos)}
self.id2pos = {ind: word for ind, word in enumerate(pos)}
self.c2i = c2i
self.m2i = m2i
self.t2i = t2i
self.i2t = {t2i[i]:i for i in self.t2i}
self.morph_dict = morph_dict
self.rels = {word: ind for ind, word in enumerate(rels)}
self.irels = rels
self.pdims = options.pembedding_dims
self.tagging_attention_size = options.tagging_att_size
self.vocab['*PAD*'] = 1
self.vocab['*INITIAL*'] = 2
self.wlookup = self.model.add_lookup_parameters((len(vocab) + 3, self.wdims))
self.clookup = self.model.add_lookup_parameters((len(c2i), self.cdims))
self.plookup = self.model.add_lookup_parameters((len(pos), self.pdims))
self.ext_embeddings = None
if options.external_embedding is not None:
ext_embeddings, ext_emb_dim = load_embeddings_file(options.external_embedding, lower=self.lowerCase, type=options.external_embedding_type)
assert (ext_emb_dim == self.wdims)
print("Initializing word embeddings by pre-trained vectors")
count = 0
for word in self.vocab:
if word in ext_embeddings:
count += 1
self.wlookup.init_row(self.vocab[word], ext_embeddings[word])
self.ext_embeddings = ext_embeddings
print("Vocab size: %d; #words having pretrained vectors: %d" % (len(self.vocab), count))
self.morph_dims = 2*2*self.mdims if self.morphFlag else 0
self.mtag_dims = 2*self.tdims if self.morphTagFlag else 0
self.pos_builders = [VanillaLSTMBuilder(1, self.wdims + self.cdims * 2 + self.morph_dims + self.mtag_dims, self.ldims, self.model),
VanillaLSTMBuilder(1, self.wdims + self.cdims * 2 + self.morph_dims + self.mtag_dims, self.ldims, self.model)]
self.pos_bbuilders = [VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model),
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model)]
if self.bibiFlag:
self.builders = [VanillaLSTMBuilder(1, self.wdims + self.cdims * 2 + self.morph_dims + self.mtag_dims + self.pdims, self.ldims, self.model),
VanillaLSTMBuilder(1, self.wdims + self.cdims * 2 + self.morph_dims + self.mtag_dims + self.pdims, self.ldims, self.model)]
self.bbuilders = [VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model),
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model)]
elif self.layers > 0:
self.builders = [VanillaLSTMBuilder(self.layers, self.wdims + self.cdims * 2 + self.morph_dims + self.mtag_dims + self.pdims, self.ldims, self.model),
VanillaLSTMBuilder(self.layers, self.wdims + self.cdims * 2 + self.morph_dims + self.mtag_dims + self.pdims, self.ldims, self.model)]
else:
self.builders = [SimpleRNNBuilder(1, self.wdims + self.cdims * 2 + self.morph_dims + self.mtag_dims, self.ldims, self.model),
SimpleRNNBuilder(1, self.wdims + self.cdims * 2 + self.morph_dims + self.mtag_dims, self.ldims, self.model)]
self.ffSeqPredictor = FFSequencePredictor(Layer(self.model, self.ldims * 2, len(self.pos), softmax))
self.hidden_units = options.hidden_units
self.hidBias = self.model.add_parameters((self.ldims * 8))
self.hidLayer = self.model.add_parameters((self.hidden_units, self.ldims * 8))
self.hid2Bias = self.model.add_parameters((self.hidden_units))
self.outLayer = self.model.add_parameters((1, self.hidden_units if self.hidden_units > 0 else self.ldims * 8))
if self.labelsFlag:
self.rhidBias = self.model.add_parameters((self.ldims * 8))
self.rhidLayer = self.model.add_parameters((self.hidden_units, self.ldims * 8))
self.rhid2Bias = self.model.add_parameters((self.hidden_units))
self.routLayer = self.model.add_parameters(
(len(self.irels), self.hidden_units if self.hidden_units > 0 else self.ldims * 8))
self.routBias = self.model.add_parameters((len(self.irels)))
self.ffRelPredictor = FFSequencePredictor(
Layer(self.model, self.hidden_units if self.hidden_units > 0 else self.ldims * 8, len(self.irels),
softmax))
self.char_rnn = RNNSequencePredictor(LSTMBuilder(1, self.cdims, self.cdims, self.model))
if self.morphFlag:
self.seg_lstm = [VanillaLSTMBuilder(1, self.cdims, self.cdims, self.model),
VanillaLSTMBuilder(1, self.cdims, self.cdims, self.model)]
self.seg_hidLayer = self.model.add_parameters((1, self.cdims*2))
self.slookup = self.model.add_lookup_parameters((len(self.c2i), self.cdims))
self.char_lstm = [VanillaLSTMBuilder(1, self.cdims, self.mdims, self.model),
VanillaLSTMBuilder(1, self.cdims, self.mdims, self.model)]
self.char_hidLayer = self.model.add_parameters((self.mdims, self.mdims*2))
self.mclookup = self.model.add_lookup_parameters((len(self.c2i), self.cdims))
self.morph_lstm = [VanillaLSTMBuilder(1, self.mdims*2, self.wdims, self.model),
VanillaLSTMBuilder(1, self.mdims*2, self.wdims, self.model)]
self.morph_hidLayer = self.model.add_parameters((self.wdims, self.wdims*2))
self.mlookup = self.model.add_lookup_parameters((len(m2i), self.mdims))
self.morph_rnn = RNNSequencePredictor(LSTMBuilder(1, self.mdims*2, self.mdims*2, self.model))
if self.morphTagFlag:
# All weights for morpheme taging will be here. (CURSOR)
# Decoder
self.dec_lstm = VanillaLSTMBuilder(1, 2 * self.cdims + self.tdims + self.cdims * 2, self.cdims, self.model)
# Attention
self.attention_w1 = self.model.add_parameters((self.tagging_attention_size, self.cdims * 2))
self.attention_w2 = self.model.add_parameters((self.tagging_attention_size, self.cdims * 2))
self.attention_v = self.model.add_parameters((1, self.tagging_attention_size))
# Attention Context
self.attention_w1_context = self.model.add_parameters((self.tagging_attention_size, self.cdims * 2))
self.attention_w2_context = self.model.add_parameters((self.tagging_attention_size, self.cdims * 2))
self.attention_v_context = self.model.add_parameters((1, self.tagging_attention_size))
# MLP - Softmax
self.decoder_w = self.model.add_parameters((len(t2i), self.cdims))
self.decoder_b = self.model.add_parameters((len(t2i)))
self.mtag_rnn = RNNSequencePredictor(VanillaLSTMBuilder(1, self.tdims, self.tdims, self.model))
self.tlookup = self.model.add_lookup_parameters((len(t2i), self.tdims))
if self.mtag_encoding_composition_type != "None":
self.mtag_encoding_f_w = self.model.add_parameters((2 * self.tdims, 4 * self.tdims))
self.mtag_encoding_f_b = self.model.add_parameters((2 * self.tdims))
self.mtag_encoding_b_w = self.model.add_parameters((2 * self.tdims, 4 * self.tdims))
self.mtag_encoding_b_b = self.model.add_parameters((2 * self.tdims))
def initialize(self):
if self.morphFlag and self.ext_embeddings:
print("Initializing word embeddings by morph2vec")
count = 0
for word in self.vocab:
if word not in self.ext_embeddings and word in self.morph_dict:
morph_seg = self.morph_dict[word]
count += 1
self.wlookup.init_row(self.vocab[word], self.__getWordVector(morph_seg).vec_value())
print("Vocab size: %d; #missing words having generated vectors: %d" % (len(self.vocab), count))
renew_cg()
def __getExpr(self, sentence, i, j):
if sentence[i].headfov is None:
sentence[i].headfov = concatenate([sentence[i].lstms[0], sentence[i].lstms[1]])
if sentence[j].modfov is None:
sentence[j].modfov = concatenate([sentence[j].lstms[0], sentence[j].lstms[1]])
_inputVector = concatenate(
[sentence[i].headfov, sentence[j].modfov, dynet.abs(sentence[i].headfov - sentence[j].modfov),
dynet.cmult(sentence[i].headfov, sentence[j].modfov)])
if self.hidden_units > 0:
output = self.outLayer.expr() * self.activation(
self.hid2Bias.expr() + self.hidLayer.expr() * self.activation(
_inputVector + self.hidBias.expr()))
else:
output = self.outLayer.expr() * self.activation(_inputVector + self.hidBias.expr())
return output
def __evaluate(self, sentence):
exprs = [[self.__getExpr(sentence, i, j) for j in range(len(sentence))] for i in range(len(sentence))]
scores = np.array([[output.scalar_value() for output in exprsRow] for exprsRow in exprs])
return scores, exprs
def pick_neg_log(self, pred, gold):
return -dynet.log(dynet.pick(pred, gold))
def binary_crossentropy(self, pred, gold):
return dynet.binary_log_loss(pred, gold)
def cosine_proximity(self, pred, gold):
def l2_normalize(x):
square_sum = dynet.sqrt(dynet.bmax(dynet.sum_elems(dynet.square(x)), np.finfo(float).eps * dynet.ones((1))[0]))
return dynet.cdiv(x, square_sum)
y_true = l2_normalize(pred)
y_pred = l2_normalize(gold)
return -dynet.sum_elems(dynet.cmult(y_true, y_pred))
def __getRelVector(self, sentence, i, j):
if sentence[i].rheadfov is None:
sentence[i].rheadfov = concatenate([sentence[i].lstms[0], sentence[i].lstms[1]])
if sentence[j].rmodfov is None:
sentence[j].rmodfov = concatenate([sentence[j].lstms[0], sentence[j].lstms[1]])
_outputVector = concatenate(
[sentence[i].rheadfov, sentence[j].rmodfov, abs(sentence[i].rheadfov - sentence[j].rmodfov),
cmult(sentence[i].rheadfov, sentence[j].rmodfov)])
if self.hidden_units > 0:
return self.rhid2Bias.expr() + self.rhidLayer.expr() * self.activation(
_outputVector + self.rhidBias.expr())
else:
return _outputVector
def __getSegmentationVector(self, word):
slstm_forward = self.seg_lstm[0].initial_state()
slstm_backward = self.seg_lstm[1].initial_state()
seg_lstm_forward = slstm_forward.transduce([self.slookup[self.c2i[char] if char in self.c2i else 0] for char in word])
seg_lstm_backward = slstm_backward.transduce([self.slookup[self.c2i[char] if char in self.c2i else 0] for char in reversed(word)])
seg_vec = []
for seg, rev_seg in zip(seg_lstm_forward,reversed(seg_lstm_backward)):
seg_vec.append(dynet.logistic(self.seg_hidLayer.expr() * concatenate([seg,rev_seg])))
seg_vec = concatenate(seg_vec)
return seg_vec
def __getMorphVector(self, morph):
clstm_forward = self.char_lstm[0].initial_state()
clstm_backward = self.char_lstm[1].initial_state()
char_lstm_forward = clstm_forward.transduce([self.mclookup[self.c2i[char] if char in self.c2i else 0] for char in morph] if len(morph) > 0 else [self.mclookup[0]])[-1]
char_lstm_backward = clstm_backward.transduce([self.mclookup[self.c2i[char] if char in self.c2i else 0] for char in reversed(morph)] if len(morph) > 0 else [self.mclookup[0]])[-1]
char_emb = self.char_hidLayer.expr() * concatenate([char_lstm_forward,char_lstm_backward])
return concatenate([self.mlookup[self.m2i[morph] if morph in self.m2i else 0], char_emb])
def __getWordVector(self, morph_seg):
mlstm_forward = self.morph_lstm[0].initial_state()
mlstm_backward = self.morph_lstm[1].initial_state()
morph_lstm_forward = mlstm_forward.transduce([self.__getMorphVector(morph) for morph in morph_seg])[-1]
morph_lstm_backward = mlstm_backward.transduce([self.__getMorphVector(morph) for morph in reversed(morph_seg)])[-1]
morph_enc = concatenate([morph_lstm_forward, morph_lstm_backward])
word_vec = self.morph_hidLayer.expr() * morph_enc
return word_vec
def attend(self, input_mat, state, w1dt):
w2 = parameter(self.attention_w2)
v = parameter(self.attention_v)
# input_mat: (encoder_state x seqlen) => input vecs concatenated as cols
# w1dt: (attdim x seqlen)
# w2dt: (attdim,1)
w2dt = w2 * concatenate(list(state.s()))
# att_weights: (seqlen,) row vector
# unnormalized: (seqlen,)
unnormalized = transpose(v * tanh(colwise_add(w1dt, w2dt)))
att_weights = softmax(unnormalized)
# context: (encoder_state)
context = input_mat * att_weights
return context
def attend_context(self, input_mat, state, w1dt_context):
w2_context = parameter(self.attention_w2_context)
v_context = parameter(self.attention_v_context)
# input_mat: (encoder_state x seqlen) => input vecs concatenated as cols
# w1dt: (attdim x seqlen)
# w2dt: (attdim,1)
w2dt_context = w2_context * concatenate(list(state.s()))
# att_weights: (seqlen,) row vector
# unnormalized: (seqlen,)
unnormalized = transpose(v_context * tanh(colwise_add(w1dt_context, w2dt_context)))
att_weights = softmax(unnormalized)
# context: (encoder_state)
context = input_mat * att_weights
return context
def decode(self, vectors, decoder_seq, word_context):
w = parameter(self.decoder_w)
b = parameter(self.decoder_b)
w1 = parameter(self.attention_w1)
w1_context = parameter(self.attention_w1_context)
input_mat = concatenate_cols(vectors)
input_context = concatenate_cols(word_context)
w1dt = None
w1dt_context = None
last_output_embeddings = self.tlookup[self.t2i["<s>"]]
s = self.dec_lstm.initial_state().add_input(concatenate([vecInput(self.cdims * 2),
last_output_embeddings,
vecInput(self.cdims * 2)]))
loss = []
for char in decoder_seq:
# w1dt can be computed and cached once for the entire decoding phase
w1dt = w1dt or w1 * input_mat
w1dt_context = w1dt_context or w1_context * input_context
vector = concatenate([self.attend(input_mat, s, w1dt),
last_output_embeddings,
self.attend_context(input_context, s, w1dt_context)])
s = s.add_input(vector)
out_vector = w * s.output() + b
probs = softmax(out_vector)
last_output_embeddings = self.tlookup[char]
loss.append(-log(pick(probs, char)))
loss = esum(loss)
return loss
def __getLossMorphTagging(self, all_encoded_states, decoder_gold, word_context):
return self.decode(all_encoded_states, decoder_gold, word_context)
def generate(self, encoded, word_context):
w = parameter(self.decoder_w)
b = parameter(self.decoder_b)
w1 = parameter(self.attention_w1)
w1_context = parameter(self.attention_w1_context)
input_mat = concatenate_cols(encoded)
input_context = concatenate_cols(word_context)
w1dt = None
w1dt_context = None
last_output_embeddings = self.tlookup[self.t2i["<s>"]]
s = self.dec_lstm.initial_state().add_input(concatenate([vecInput(self.cdims * 2),
last_output_embeddings,
vecInput(self.cdims * 2)]))
out = []
count_EOS = 0
limit_features = 10
for i in range(limit_features):
if count_EOS == 2: break
# w1dt can be computed and cached once for the entire decoding phase
w1dt = w1dt or w1 * input_mat
w1dt_context = w1dt_context or w1_context * input_context
vector = concatenate([self.attend(input_mat, s, w1dt),
last_output_embeddings,
self.attend_context(input_context, s, w1dt_context)])
s = s.add_input(vector)
out_vector = w * s.output() + b
probs = softmax(out_vector).vec_value()
next_char = probs.index(max(probs))
last_output_embeddings = self.tlookup[next_char]
if next_char == self.t2i["<s>"]:
count_EOS += 1
out.append(next_char)
return out
def Save(self, filename):
self.model.save(filename)
def Load(self, filename):
self.model.populate(filename)
def Predict(self, conll_path):
with open(conll_path, 'r') as conllFP:
for iSentence, sentence in enumerate(read_conll(conllFP, self.c2i, self.m2i, self.t2i, self.morph_dict)):
conll_sentence = [entry for entry in sentence if isinstance(entry, utils.ConllEntry)]
if self.morphTagFlag:
sentence_context = []
last_state_char = self.char_rnn.predict_sequence([self.clookup[self.c2i["<start>"]]])[-1]
rev_last_state_char = self.char_rnn.predict_sequence([self.clookup[self.c2i["<start>"]]])[-1]
sentence_context.append(concatenate([last_state_char, rev_last_state_char]))
for entry in conll_sentence:
last_state_char = self.char_rnn.predict_sequence([self.clookup[c] for c in entry.idChars])
rev_last_state_char = self.char_rnn.predict_sequence([self.clookup[c] for c in reversed(entry.idChars)])
entry.char_rnn_states = [concatenate([f,b]) for f,b in zip(last_state_char, rev_last_state_char)]
sentence_context.append(entry.char_rnn_states[-1])
for idx, entry in enumerate(conll_sentence):
wordvec = self.wlookup[int(self.vocab.get(entry.norm, 0))] if self.wdims > 0 else None
if self.morphTagFlag:
entry.vec = concatenate([wordvec, entry.char_rnn_states[-1]])
else:
last_state_char = self.char_rnn.predict_sequence([self.clookup[c] for c in entry.idChars])[-1]
rev_last_state_char = self.char_rnn.predict_sequence([self.clookup[c] for c in reversed(entry.idChars)])[-1]
entry.vec = concatenate([wordvec, last_state_char, rev_last_state_char])
for idx, entry in enumerate(conll_sentence):
if self.morphFlag:
if len(entry.norm) > 2:
if self.goldMorphFlag:
seg_vec = self.__getSegmentationVector(entry.norm)
seg_vec = dynet.vecInput(seg_vec.dim()[0][0])
seg_vec.set(entry.idMorphs)
morph_seg = utils.generate_morphs(entry.norm, seg_vec.vec_value())
entry.pred_seg = morph_seg
else:
seg_vec = self.__getSegmentationVector(entry.norm)
morph_seg = utils.generate_morphs(entry.norm, seg_vec.vec_value())
entry.pred_seg = seg_vec.vec_value()
else:
morph_seg = [entry.norm]
entry.pred_seg = entry.idMorphs
entry.seg = entry.idMorphs
last_state_morph = self.morph_rnn.predict_sequence([self.__getMorphVector(morph) for morph in morph_seg])[-1]
rev_last_state_morph = self.morph_rnn.predict_sequence([self.__getMorphVector(morph) for morph in reversed(morph_seg)])[
-1]
entry.vec = concatenate([entry.vec, last_state_morph, rev_last_state_morph])
morphtag_encodings = []
for idx, entry in enumerate(conll_sentence):
if self.morphTagFlag:
if self.goldMorphTagFlag:
morph_tags = entry.idMorphTags
entry.pred_tags = entry.idMorphTags
entry.pred_tags_tokens = [self.i2t[m_tag_id] for m_tag_id in entry.pred_tags]
else:
word_context = [c for i, c in enumerate(sentence_context) if i - 1 != idx]
entry.pred_tags = self.generate(entry.char_rnn_states, word_context)
morph_tags = entry.pred_tags
entry.tags = entry.idMorphTags
entry.pred_tags_tokens = [self.i2t[m_tag_id] for m_tag_id in entry.pred_tags]
last_state_mtag = self.mtag_rnn.predict_sequence([self.tlookup[t] for t in morph_tags])[-1]
rev_last_state_mtag = self.mtag_rnn.predict_sequence([self.tlookup[t] for t in reversed(morph_tags)])[-1]
current_encoding_mtag = concatenate([last_state_mtag, rev_last_state_mtag])
morphtag_encodings.append(current_encoding_mtag)
if self.morphTagFlag:
forward = []
for idx, encoding in enumerate(morphtag_encodings):
if idx == 0:
forward.append(encoding)
else:
updated = morphtag_encodings[idx-1]*self.mtag_encoding_composition_alpha \
+ encoding*(1-self.mtag_encoding_composition_alpha)
forward.append(updated)
if self.mtag_encoding_composition_type == "w_sum":
upper_morphtag_encodings = forward
elif self.mtag_encoding_composition_type == "bi_w_sum":
backward = []
for idx, r_encoding in enumerate(morphtag_encodings):
if idx == len(morphtag_encodings) - 1:
backward.append(r_encoding)
else:
updated = morphtag_encodings[idx+1]*self.mtag_encoding_composition_alpha \
+ r_encoding*(1-self.mtag_encoding_composition_alpha)
backward.append(updated)
upper_morphtag_encodings = [f+b for f,b in zip(forward, backward)]
elif self.mtag_encoding_composition_type == "bi_mlp":
forward = []
backward = []
for idx, encoding in enumerate(morphtag_encodings):
if idx != 0:
f = self.mtag_encoding_f_w * concatenate([encoding, morphtag_encodings[idx-1]]) \
+ self.mtag_encoding_f_b
forward.append(f)
else:
forward.append(encoding)
if idx != len(morphtag_encodings) - 1:
b = self.mtag_encoding_b_w * concatenate([encoding, morphtag_encodings[idx+1]]) \
+ self.mtag_encoding_b_b
backward.append(b)
else:
backward.append(encoding)
upper_morphtag_encodings = [f+b for f,b in zip(forward, backward)]
else:
upper_morphtag_encodings = morphtag_encodings
for entry, mtag in zip(conll_sentence, upper_morphtag_encodings):
entry.vec = concatenate([entry.vec, mtag])
for idx, entry in enumerate(conll_sentence):
entry.pos_lstms = [entry.vec, entry.vec]
entry.headfov = None
entry.modfov = None
entry.rheadfov = None
entry.rmodfov = None
#Predicted pos tags
lstm_forward = self.pos_builders[0].initial_state()
lstm_backward = self.pos_builders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.pos_lstms[1] = lstm_forward.output()
rentry.pos_lstms[0] = lstm_backward.output()
for entry in conll_sentence:
entry.pos_vec = concatenate(entry.pos_lstms)
blstm_forward = self.pos_bbuilders[0].initial_state()
blstm_backward = self.pos_bbuilders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.pos_vec)
blstm_backward = blstm_backward.add_input(rentry.pos_vec)
entry.pos_lstms[1] = blstm_forward.output()
rentry.pos_lstms[0] = blstm_backward.output()
concat_layer = [concatenate(entry.pos_lstms) for entry in conll_sentence]
outputFFlayer = self.ffSeqPredictor.predict_sequence(concat_layer)
predicted_pos_indices = [np.argmax(o.value()) for o in outputFFlayer]
predicted_postags = [self.id2pos[idx] for idx in predicted_pos_indices]
# Add predicted pos tags for parsing prediction
for entry, posid in zip(conll_sentence, predicted_pos_indices):
entry.vec = concatenate([entry.vec, self.plookup[posid]])
entry.lstms = [entry.vec, entry.vec]
if self.blstmFlag:
lstm_forward = self.builders[0].initial_state()
lstm_backward = self.builders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.lstms[1] = lstm_forward.output()
rentry.lstms[0] = lstm_backward.output()
if self.bibiFlag:
for entry in conll_sentence:
entry.vec = concatenate(entry.lstms)
blstm_forward = self.bbuilders[0].initial_state()
blstm_backward = self.bbuilders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.vec)
blstm_backward = blstm_backward.add_input(rentry.vec)
entry.lstms[1] = blstm_forward.output()
rentry.lstms[0] = blstm_backward.output()
scores, exprs = self.__evaluate(conll_sentence)
heads = decoder.parse_proj(scores)
# Multiple roots: heading to the previous "rooted" one
rootCount = 0
rootWid = -1
for index, head in enumerate(heads):
if head == 0:
rootCount += 1
if rootCount == 1:
rootWid = index
if rootCount > 1:
heads[index] = rootWid
rootWid = index
for entry, head, pos in zip(conll_sentence, heads, predicted_postags):
entry.pred_parent_id = head
entry.pred_relation = '_'
entry.pred_pos = pos
dump = False
if self.labelsFlag:
concat_layer = [self.__getRelVector(conll_sentence, head, modifier + 1) for modifier, head in
enumerate(heads[1:])]
outputFFlayer = self.ffRelPredictor.predict_sequence(concat_layer)
predicted_rel_indices = [np.argmax(o.value()) for o in outputFFlayer]
predicted_rels = [self.irels[idx] for idx in predicted_rel_indices]
for modifier, head in enumerate(heads[1:]):
conll_sentence[modifier + 1].pred_relation = predicted_rels[modifier]
renew_cg()
if not dump:
yield sentence
def morph2word(self, morph_dict):
word_emb = {}
for word in morph_dict.keys():
morph_seg = morph_dict[word]
word_vec = self.__getWordVector(morph_seg)
word_emb[word] = word_vec.vec_value()
renew_cg()
return word_emb
def morph(self):
morph_dict = {}
for morph in self.m2i.keys():
morph_dict[morph] = self.__getMorphVector(morph).vec_value()
renew_cg()
return morph_dict
def Train_Morph(self):
self.trainer.set_sparse_updates(False)
start = time.time()
for iWord, word in enumerate(list(self.morph_dict.keys())):
if iWord % 2000 == 0 and iWord != 0:
print("Processing word number: %d" % iWord, ", Time: %.2f" % (time.time() - start))
start = time.time()
morph_seg = self.morph_dict[word]
morph_vec = self.__getWordVector(morph_seg)
if self.ext_embeddings is None:
vec_gold = self.wlookup[int(self.vocab.get(word, 0))].vec_value()
elif word in self.ext_embeddings:
vec_gold = self.ext_embeddings[word]
else:
vec_gold = None
if vec_gold is not None:
y_gold = dynet.vecInput(self.wdims)
y_gold.set(vec_gold)
mErrs = self.cosine_proximity(morph_vec, y_gold)
mErrs.backward()
self.trainer.update()
renew_cg()
def embed_word(self, word):
return [self.input_lookup[char] for char in word]
def run_lstm(self, init_state, input_vecs):
s = init_state
out_vectors = []
for vector in input_vecs:
s = s.add_input(vector)
out_vector = s.output()
out_vectors.append(out_vector)
return out_vectors
def encode_word(self, word):
word_rev = list(reversed(word))
fwd_vectors = self.run_lstm(self.enc_fwd_lstm.initial_state(), word)
bwd_vectors = self.run_lstm(self.enc_bwd_lstm.initial_state(), word_rev)
bwd_vectors = list(reversed(bwd_vectors))
vectors = [concatenate(list(p)) for p in zip(fwd_vectors, bwd_vectors)]
return vectors
def Train(self, conll_path):
self.trainer.set_sparse_updates(True)
eloss = 0.0
mloss = 0.0
eerrors = 0
etotal = 0
start = time.time()
with open(conll_path, 'r') as conllFP:
shuffledData = list(read_conll(conllFP, self.c2i, self.m2i, self.t2i, self.morph_dict))
random.shuffle(shuffledData)
errs = []
lerrs = []
posErrs = []
segErrs = []
mTagErrs = []
for iSentence, sentence in enumerate(shuffledData):
if iSentence % 500 == 0 and iSentence != 0:
print("Processing sentence number: %d" % iSentence, ", Loss: %.4f" % (
eloss / etotal), ", Time: %.2f" % (time.time() - start))
start = time.time()
eerrors = 0
eloss = 0.0
etotal = 0
conll_sentence = [entry for entry in sentence if isinstance(entry, utils.ConllEntry)]
if self.morphTagFlag:
sentence_context = []
last_state_char = self.char_rnn.predict_sequence([self.clookup[self.c2i["<start>"]]])[-1]
rev_last_state_char = self.char_rnn.predict_sequence([self.clookup[self.c2i["<start>"]]])[-1]
sentence_context.append(concatenate([last_state_char, rev_last_state_char]))
for entry in conll_sentence:
last_state_char = self.char_rnn.predict_sequence([self.clookup[c] for c in entry.idChars])
rev_last_state_char = self.char_rnn.predict_sequence([self.clookup[c] for c in reversed(entry.idChars)])
entry.char_rnn_states = [concatenate([f,b]) for f,b in zip(last_state_char, rev_last_state_char)]
sentence_context.append(entry.char_rnn_states[-1])
for idx, entry in enumerate(conll_sentence):
c = float(self.wordsCount.get(entry.norm, 0))
dropFlag = (random.random() < (c / (0.25 + c)))
wordvec = self.wlookup[
int(self.vocab.get(entry.norm, 0)) if dropFlag else 0] if self.wdims > 0 else None
if self.morphTagFlag :
entry.vec = dynet.dropout(concatenate([wordvec, entry.char_rnn_states[-1]]), 0.33)
else:
last_state_char = self.char_rnn.predict_sequence([self.clookup[c] for c in entry.idChars])[-1]
rev_last_state_char = self.char_rnn.predict_sequence([self.clookup[c] for c in reversed(entry.idChars)])[-1]
entry.vec = dynet.dropout(concatenate([wordvec, last_state_char, rev_last_state_char]), 0.33)
for idx, entry in enumerate(conll_sentence):
if self.morphFlag:
if len(entry.norm) > 2:
if self.goldMorphFlag:
seg_vec = self.__getSegmentationVector(entry.norm)
seg_vec = dynet.vecInput(seg_vec.dim()[0][0])
seg_vec.set(entry.idMorphs)
morph_seg = utils.generate_morphs(entry.norm, seg_vec.vec_value())
else:
seg_vec = self.__getSegmentationVector(entry.norm)
morph_seg = utils.generate_morphs(entry.norm, seg_vec.vec_value())
vec_gold = dynet.vecInput(seg_vec.dim()[0][0])
vec_gold.set(entry.idMorphs)
segErrs.append(self.binary_crossentropy(seg_vec,vec_gold))
else:
morph_seg = [entry.norm]
last_state_morph = self.morph_rnn.predict_sequence([self.__getMorphVector(morph) for morph in morph_seg])[-1]
rev_last_state_morph = self.morph_rnn.predict_sequence([self.__getMorphVector(morph) for morph in reversed(morph_seg)])[
-1]
encoding_morph = concatenate([last_state_morph, rev_last_state_morph])
entry.vec = concatenate([entry.vec, dynet.dropout(encoding_morph, 0.33)])
morphtag_encodings = []
for idx, entry in enumerate(conll_sentence):
if self.morphTagFlag:
if self.goldMorphTagFlag:
morph_tags = entry.idMorphTags
else:
word_context = [c for i, c in enumerate(sentence_context) if i-1 != idx]
mTagErrs.append(
self.__getLossMorphTagging(entry.char_rnn_states, entry.idMorphTags, word_context))
predicted_sequence = self.generate(entry.char_rnn_states, word_context)
morph_tags = predicted_sequence
last_state_mtag = self.mtag_rnn.predict_sequence([self.tlookup[t] for t in morph_tags])[-1]
rev_last_state_mtag = \
self.mtag_rnn.predict_sequence([self.tlookup[t] for t in reversed(morph_tags)])[
-1]
current_encoding_mtag = concatenate([last_state_mtag, rev_last_state_mtag])
morphtag_encodings.append(current_encoding_mtag)
if self.morphTagFlag:
forward = []
for idx, encoding in enumerate(morphtag_encodings):
if idx == 0:
forward.append(encoding)
else:
updated = morphtag_encodings[idx-1]*self.mtag_encoding_composition_alpha \
+ encoding*(1-self.mtag_encoding_composition_alpha)
forward.append(updated)
if self.mtag_encoding_composition_type == "w_sum":
upper_morphtag_encodings = forward
elif self.mtag_encoding_composition_type == "bi_w_sum":
backward = []
for idx, r_encoding in enumerate(morphtag_encodings):
if idx == len(morphtag_encodings) - 1:
backward.append(r_encoding)
else:
updated = morphtag_encodings[idx+1]*self.mtag_encoding_composition_alpha \
+ r_encoding*(1-self.mtag_encoding_composition_alpha)
backward.append(updated)
upper_morphtag_encodings = [f+b for f,b in zip(forward, backward)]
elif self.mtag_encoding_composition_type == "bi_mlp":
forward = []
backward = []
for idx, encoding in enumerate(morphtag_encodings):
if idx != 0:
f = self.mtag_encoding_f_w * concatenate([encoding, morphtag_encodings[idx-1]]) \
+ self.mtag_encoding_f_b
forward.append(f)
else:
forward.append(encoding)
if idx != len(morphtag_encodings) - 1:
b = self.mtag_encoding_b_w * concatenate([encoding, morphtag_encodings[idx+1]]) \
+ self.mtag_encoding_b_b
backward.append(b)
else:
backward.append(encoding)
upper_morphtag_encodings = [f+b for f,b in zip(forward, backward)]
else:
upper_morphtag_encodings = morphtag_encodings
for entry, mtag in zip(conll_sentence, upper_morphtag_encodings):
entry.vec = concatenate([entry.vec, dynet.dropout(mtag, 0.33)])
for idx, entry in enumerate(conll_sentence):
entry.pos_lstms = [entry.vec, entry.vec]
entry.headfov = None
entry.modfov = None
entry.rheadfov = None
entry.rmodfov = None
#POS tagging loss
lstm_forward = self.pos_builders[0].initial_state()
lstm_backward = self.pos_builders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.pos_lstms[1] = lstm_forward.output()
rentry.pos_lstms[0] = lstm_backward.output()
for entry in conll_sentence:
entry.pos_vec = concatenate(entry.pos_lstms)
blstm_forward = self.pos_bbuilders[0].initial_state()
blstm_backward = self.pos_bbuilders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.pos_vec)
blstm_backward = blstm_backward.add_input(rentry.pos_vec)
entry.pos_lstms[1] = blstm_forward.output()
rentry.pos_lstms[0] = blstm_backward.output()
concat_layer = [dynet.dropout(concatenate(entry.pos_lstms), 0.33) for entry in conll_sentence]
outputFFlayer = self.ffSeqPredictor.predict_sequence(concat_layer)
posIDs = [self.pos.get(entry.pos) for entry in conll_sentence]
for pred, gold in zip(outputFFlayer, posIDs):
posErrs.append(self.pick_neg_log(pred, gold))
# Add predicted pos tags
for entry, poses in zip(conll_sentence, outputFFlayer):
entry.vec = concatenate([entry.vec, dynet.dropout(self.plookup[np.argmax(poses.value())], 0.33)])
entry.lstms = [entry.vec, entry.vec]
#Parsing losses
if self.blstmFlag:
lstm_forward = self.builders[0].initial_state()
lstm_backward = self.builders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.lstms[1] = lstm_forward.output()
rentry.lstms[0] = lstm_backward.output()
if self.bibiFlag:
for entry in conll_sentence:
entry.vec = concatenate(entry.lstms)
blstm_forward = self.bbuilders[0].initial_state()
blstm_backward = self.bbuilders[1].initial_state()
for entry, rentry in zip(conll_sentence, reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.vec)
blstm_backward = blstm_backward.add_input(rentry.vec)
entry.lstms[1] = blstm_forward.output()
rentry.lstms[0] = blstm_backward.output()
scores, exprs = self.__evaluate(conll_sentence)
gold = [entry.parent_id for entry in conll_sentence]
heads = decoder.parse_proj(scores, gold if self.costaugFlag else None)
if self.labelsFlag:
concat_layer = [dynet.dropout(self.__getRelVector(conll_sentence, head, modifier + 1), 0.33) for
modifier, head in enumerate(gold[1:])]
outputFFlayer = self.ffRelPredictor.predict_sequence(concat_layer)
relIDs = [self.rels[conll_sentence[modifier + 1].relation] for modifier, _ in enumerate(gold[1:])]
for pred, goldid in zip(outputFFlayer, relIDs):
lerrs.append(self.pick_neg_log(pred, goldid))
e = sum([1 for h, g in zip(heads[1:], gold[1:]) if h != g])
eerrors += e
if e > 0:
loss = [(exprs[h][i] - exprs[g][i]) for i, (h, g) in enumerate(zip(heads, gold)) if h != g] # * (1.0/float(e))
eloss += (e)
mloss += (e)
errs.extend(loss)
etotal += len(conll_sentence)
if iSentence % 1 == 0:
if len(errs) > 0 or len(lerrs) > 0 or len(posErrs) > 0 or len(segErrs) > 0 or len(mTagErrs) > 0:
eerrs = (esum(errs + lerrs + posErrs + segErrs + mTagErrs))
eerrs.scalar_value()
eerrs.backward()
self.trainer.update()
errs = []
lerrs = []
posErrs = []
segErrs = []
mTagErrs = []
renew_cg()
print("Loss: %.4f" % (mloss / iSentence))
class jNeRE:
def __init__(self, vocab, ner, rels, w2i, c2i, options):
self.model = ParameterCollection()
random.seed(1)
self.trainer = AdamTrainer(self.model)
if options.learning_rate is not None:
self.trainer = AdamTrainer(self.model, alpha=options.learning_rate)
print("Adam initial learning rate:", options.learning_rate)
self.activations = {
'tanh': tanh,
'sigmoid': logistic,
'relu': rectify,
'tanh3': (lambda x: tanh(cwise_multiply(cwise_multiply(x, x), x)))
}
self.activation = self.activations[options.activation]
self.blstmFlag = options.blstmFlag
self.labelsFlag = options.labelsFlag
self.costaugFlag = options.costaugFlag
self.bibiFlag = options.bibiFlag
self.ldims = options.lstm_dims
self.wdims = options.wembedding_dims
self.cdims = options.cembedding_dims
self.layers = options.lstm_layers
self.wordsCount = vocab
self.vocab = {word: ind + 3 for word, ind in w2i.iteritems()}
self.ner = {word: ind for ind, word in enumerate(ner)}
self.id2ner = {ind: word for ind, word in enumerate(ner)}
self.c2i = c2i
self.rels = {word: ind for ind, word in enumerate(rels)}
self.id2rels = rels
# print self.rels
# print self.id2rels
self.nerdims = options.nembedding_dims
self.mixture_weight = options.mixture_weight
#self.posCount = postagCount
#self.pos2id = {word: ind + 1 for ind, word in enumerate(postagCount.keys())}
#self.pdims = options.pembedding_dims
self.vocab['*PAD*'] = 1
self.vocab['*INITIAL*'] = 2
self.wlookup = self.model.add_lookup_parameters(
(len(vocab) + 3, self.wdims))
self.clookup = self.model.add_lookup_parameters((len(c2i), self.cdims))
self.nerlookup = self.model.add_lookup_parameters(
(len(ner), self.nerdims))
#self.plookup = self.model.add_lookup_parameters((len(postagCount.keys()) + 1, self.pdims))
if options.external_embedding is not None:
ext_embeddings, ext_emb_dim = load_embeddings_file(
options.external_embedding, lower=True)
assert (ext_emb_dim == self.wdims)
print("Initializing word embeddings by pre-trained vectors")
count = 0
for word in self.vocab:
_word = unicode(word, "utf-8")
if _word in ext_embeddings:
count += 1
self.wlookup.init_row(self.vocab[word],
ext_embeddings[_word])
print("Vocab size: %d; #words having pretrained vectors: %d" %
(len(self.vocab), count))
self.ner_builders = [
VanillaLSTMBuilder(1, self.wdims + self.cdims * 2, self.ldims,
self.model),
VanillaLSTMBuilder(1, self.wdims + self.cdims * 2, self.ldims,
self.model)
]
self.ner_bbuilders = [
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model),
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model)
]
if self.bibiFlag:
self.builders = [
VanillaLSTMBuilder(1,
self.wdims + self.cdims * 2 + self.nerdims,
self.ldims, self.model),
VanillaLSTMBuilder(1,
self.wdims + self.cdims * 2 + self.nerdims,
self.ldims, self.model)
]
self.bbuilders = [
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model),
VanillaLSTMBuilder(1, self.ldims * 2, self.ldims, self.model)
]
elif self.layers > 0:
self.builders = [
VanillaLSTMBuilder(self.layers,
self.wdims + self.cdims * 2 + self.nerdims,
self.ldims, self.model),
VanillaLSTMBuilder(self.layers,
self.wdims + self.cdims * 2 + self.nerdims,
self.ldims, self.model)
]
else:
self.builders = [
SimpleRNNBuilder(1, self.wdims + self.cdims * 2 + self.nerdims,
self.ldims, self.model),
SimpleRNNBuilder(1, self.wdims + self.cdims * 2 + self.nerdims,
self.ldims, self.model)
]
# self.ffSeqPredictor = FFSequencePredictor(Layer(self.model, self.ldims * 2, len(self.ner), softmax))
self.hidden_units = options.hidden_units
self.char_rnn = RNNSequencePredictor(
LSTMBuilder(1, self.cdims, self.cdims, self.model))
self.crf_module = CRF(self.model, self.id2ner)
self.tanh_layer_W = self.model.add_parameters(
(self.hidden_units, 2 * self.ldims))
self.tanh_layer_b = self.model.add_parameters((self.hidden_units))
self.last_layer_W = self.model.add_parameters(
(len(self.ner), self.hidden_units))
self.last_layer_b = self.model.add_parameters((len(self.ner)))
W = orthonormal_initializer(self.hidden_units, 2 * self.ldims)
self.head_layer_W = self.model.parameters_from_numpy(W)
self.head_layer_b = self.model.add_parameters(
(self.hidden_units, ), init=dynet.ConstInitializer(0.))
self.dep_layer_W = self.model.parameters_from_numpy(W)
self.dep_layer_b = self.model.add_parameters(
(self.hidden_units, ), init=dynet.ConstInitializer(0.))
self.rel_U = self.model.add_parameters(
(len(self.rels) * self.hidden_units, self.hidden_units),
init=dynet.ConstInitializer(0.))
self.rel_W = self.model.parameters_from_numpy(
orthonormal_initializer(len(self.rels), 2 * self.hidden_units))
#self.rel_W = self.model.add_parameters((len(self.rels), self.hidden_units * 2))
self.rel_b = self.model.add_parameters((len(self.rels), ),
init=dynet.ConstInitializer(0.))
def pick_neg_log(self, pred, gold):
return -dynet.log(dynet.pick(pred, gold))
def Save(self, filename):
self.model.save(filename)
def Load(self, filename):
self.model.populate(filename)
def Predict(self, test_data):
# with open(conll_path, 'r') as conllFP:
outputPredNER = {}
id2arg2rel = {}
outputPredRel = {}
count = 0.0
nercount = 0.0
for sentenceID in test_data:
sentence = test_data[sentenceID]
for entry in sentence:
wordvec = self.wlookup[int(self.vocab.get(
entry.norm, 0))] if self.wdims > 0 else None
last_state = self.char_rnn.predict_sequence(
[self.clookup[c] for c in entry.idChars])[-1]
rev_last_state = self.char_rnn.predict_sequence(
[self.clookup[c] for c in reversed(entry.idChars)])[-1]
entry.vec = concatenate(
filter(None, [wordvec, last_state, rev_last_state]))
entry.ner_lstms = [entry.vec, entry.vec]
entry.headfov = None
entry.modfov = None
entry.rheadfov = None
entry.rmodfov = None
# Predicted ner tags
lstm_forward = self.ner_builders[0].initial_state()
lstm_backward = self.ner_builders[1].initial_state()
for entry, rentry in zip(sentence, reversed(sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.ner_lstms[1] = lstm_forward.output()
rentry.ner_lstms[0] = lstm_backward.output()
for entry in sentence:
entry.ner_vec = concatenate(entry.ner_lstms)
blstm_forward = self.ner_bbuilders[0].initial_state()
blstm_backward = self.ner_bbuilders[1].initial_state()
for entry, rentry in zip(sentence, reversed(sentence)):
blstm_forward = blstm_forward.add_input(entry.ner_vec)
blstm_backward = blstm_backward.add_input(rentry.ner_vec)
entry.ner_lstms[1] = blstm_forward.output()
rentry.ner_lstms[0] = blstm_backward.output()
concat_layer = [concatenate(entry.ner_lstms) for entry in sentence]
context_representations = [dynet.tanh(dynet.affine_transform([self.tanh_layer_b.expr(),
self.tanh_layer_W.expr(),
context])) \
for context in concat_layer]
tag_scores = [dynet.affine_transform([self.last_layer_b.expr(),
self.last_layer_W.expr(),
context]) \
for context in context_representations]
observations = [
dynet.concatenate([obs, dynet.inputVector([-1e10, -1e10])],
d=0) for obs in tag_scores
]
predicted_ner_indices, _ = self.crf_module.viterbi_decoding(
observations)
predicted_nertags = [
self.id2ner[idx] for idx in predicted_ner_indices
]
outputPredNER[sentenceID] = predicted_nertags
# Add ner embeddings
for entry, ner in zip(sentence, predicted_ner_indices):
entry.vec = concatenate([entry.vec, self.nerlookup[ner]])
entry.lstms = [entry.vec, entry.vec]
# Relation losses
if self.blstmFlag:
lstm_forward = self.builders[0].initial_state()
lstm_backward = self.builders[1].initial_state()
for entry, rentry in zip(sentence, reversed(sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.lstms[1] = lstm_forward.output()
rentry.lstms[0] = lstm_backward.output()
if self.bibiFlag:
for entry in sentence:
entry.vec = concatenate(entry.lstms)
blstm_forward = self.bbuilders[0].initial_state()
blstm_backward = self.bbuilders[1].initial_state()
for entry, rentry in zip(sentence, reversed(sentence)):
blstm_forward = blstm_forward.add_input(entry.vec)
blstm_backward = blstm_backward.add_input(rentry.vec)
entry.lstms[1] = blstm_forward.output()
rentry.lstms[0] = blstm_backward.output()
concat_layer = [concatenate(entry.lstms) for entry in sentence]
head_context_representations = [dynet.tanh(dynet.affine_transform([self.head_layer_b.expr(),
self.head_layer_W.expr(),
context])) \
for context in concat_layer]
dep_context_representations = [dynet.tanh(dynet.affine_transform([self.dep_layer_b.expr(),
self.dep_layer_W.expr(),
context])) \
for context in concat_layer]
nerids = []
for ind, tag in enumerate(predicted_nertags):
# print(ind, tag)
if str(tag).startswith("L-") or str(tag).startswith("U-"):
if not str(tag).endswith("-Other"):
nerids.append(ind)
id2arg2rel[sentenceID] = {}
for ind1 in nerids:
for ind2 in nerids:
if ind1 != ind2:
if predicted_nertags[ind1] in [
"L-Peop", "U-Peop"
] and predicted_nertags[ind2] in [
"L-Peop", "U-Peop", "L-Org", "U-Org", "L-Loc",
"U-Loc"
]:
id2arg2rel[sentenceID][(ind1, ind2)] = "NEG"
if predicted_nertags[ind1] in [
"L-Loc", "U-Loc", "L-Org", "U-Org"
] and predicted_nertags[ind2] in ["L-Loc", "U-Loc"]:
id2arg2rel[sentenceID][(ind1, ind2)] = "NEG"
# id2arg2rel[sentenceID][(ind1, ind2)] = "NEG"
for (head, dep) in id2arg2rel[sentenceID]:
# print (head, dep), pairrels[(head, dep)]
linear = self.rel_U.expr() * dep_context_representations[dep]
linear = dynet.reshape(linear,
(self.hidden_units, len(self.rels)))
bilinear = dynet.transpose(
head_context_representations[head]) * linear
biaffine = dynet.transpose(
bilinear) + self.rel_W.expr() * concatenate([
head_context_representations[head],
dep_context_representations[dep]
]) + self.rel_b.expr()
id2arg2rel[sentenceID][(head, dep)] = self.id2rels[np.argmax(
softmax(biaffine).value())]
outputPredRel[sentenceID] = {}
for (head, dep) in id2arg2rel[sentenceID]:
rel = id2arg2rel[sentenceID][(head, dep)]
if rel != "NEG":
outputPredRel[sentenceID][(head, dep)] = rel
# else:
# i_rel = id2arg2rel[sentenceID][(dep, head)]
# if str(i_rel).endswith("-1"):
# outputPredRel[sentenceID][(head, dep)] = i_rel[:-2]
renew_cg()
# print "----"
# print outputPredNER[sentenceID]
# print id2arg2rel[sentenceID]
# print outputPredRel[sentenceID]
return outputPredNER, outputPredRel
def Train(self, train_data, train_id2nerBILOU, id2arg2rel, isTrain=True):
eloss = 0.0
mloss = 0.0
eerrors = 0
etotal = 0
start = time.time()
nwtotal = 0
if isTrain:
shuffledData = train_data.keys()
random.shuffle(shuffledData)
# errs = []
lerrs = []
nerErrs = []
for iSentence, sentenceId in enumerate(shuffledData):
if iSentence % 100 == 0 and iSentence != 0:
print "Processing sentence number: %d" % iSentence, ", Loss: %.4f" % (
eloss / etotal), ", Time: %.2f" % (time.time() - start)
start = time.time()
eerrors = 0
eloss = 0.0
etotal = 0
sentence = train_data[sentenceId]
goldNers = train_id2nerBILOU[sentenceId].strip().split()
for entry in sentence:
c = float(self.wordsCount.get(entry.norm, 0))
dropFlag = (random.random() < (c / (0.25 + c)))
wordvec = self.wlookup[
int(self.vocab.get(entry.norm, 0)
) if dropFlag else 0] if self.wdims > 0 else None
last_state = self.char_rnn.predict_sequence(
[self.clookup[c] for c in entry.idChars])[-1]
rev_last_state = self.char_rnn.predict_sequence(
[self.clookup[c] for c in reversed(entry.idChars)])[-1]
entry.vec = dynet.dropout(
concatenate(
filter(None,
[wordvec, last_state, rev_last_state])),
0.33)
entry.ner_lstms = [entry.vec, entry.vec]
entry.headfov = None
entry.modfov = None
entry.rheadfov = None
entry.rmodfov = None
# ner tagging loss
lstm_forward = self.ner_builders[0].initial_state()
lstm_backward = self.ner_builders[1].initial_state()
for entry, rentry in zip(sentence, reversed(sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.ner_lstms[1] = lstm_forward.output()
rentry.ner_lstms[0] = lstm_backward.output()
for entry in sentence:
entry.ner_vec = concatenate(entry.ner_lstms)
blstm_forward = self.ner_bbuilders[0].initial_state()
blstm_backward = self.ner_bbuilders[1].initial_state()
for entry, rentry in zip(sentence, reversed(sentence)):
blstm_forward = blstm_forward.add_input(entry.ner_vec)
blstm_backward = blstm_backward.add_input(rentry.ner_vec)
entry.ner_lstms[1] = blstm_forward.output()
rentry.ner_lstms[0] = blstm_backward.output()
concat_layer = [
dynet.dropout(concatenate(entry.ner_lstms), 0.33)
for entry in sentence
]
context_representations = [dynet.tanh(dynet.affine_transform([self.tanh_layer_b.expr(),
self.tanh_layer_W.expr(),
context])) \
for context in concat_layer]
tag_scores = [dynet.affine_transform([self.last_layer_b.expr(),
self.last_layer_W.expr(),
context]) \
for context in context_representations]
nerIDs = [self.ner.get(tag) for tag in goldNers]
loss = self.crf_module.neg_log_loss(tag_scores, nerIDs)
# loss, _ = self.crf_module.viterbi_loss(tag_scores, nerIDs)
nerErrs.append(loss)
# observations = [dynet.concatenate([obs, dynet.inputVector([-1e10, -1e10])], d=0) for obs in
# tag_scores]
# predicted_ner_indices, _ = self.crf_module.viterbi_decoding(observations)
# Add ner embeddings
for entry, ner in zip(sentence, nerIDs):
entry.vec = concatenate(
[entry.vec,
dynet.dropout(self.nerlookup[ner], 0.33)])
entry.lstms = [entry.vec, entry.vec]
# Relation losses
if self.blstmFlag:
lstm_forward = self.builders[0].initial_state()
lstm_backward = self.builders[1].initial_state()
for entry, rentry in zip(sentence, reversed(sentence)):
lstm_forward = lstm_forward.add_input(entry.vec)
lstm_backward = lstm_backward.add_input(rentry.vec)
entry.lstms[1] = lstm_forward.output()
rentry.lstms[0] = lstm_backward.output()
if self.bibiFlag:
for entry in sentence:
entry.vec = concatenate(entry.lstms)
blstm_forward = self.bbuilders[0].initial_state()
blstm_backward = self.bbuilders[1].initial_state()
for entry, rentry in zip(sentence, reversed(sentence)):
blstm_forward = blstm_forward.add_input(entry.vec)
blstm_backward = blstm_backward.add_input(
rentry.vec)
entry.lstms[1] = blstm_forward.output()
rentry.lstms[0] = blstm_backward.output()
concat_layer = [
dynet.dropout(concatenate(entry.lstms), 0.33)
for entry in sentence
]
head_context_representations = [dynet.tanh(dynet.affine_transform([self.head_layer_b.expr(),
self.head_layer_W.expr(),
context])) \
for context in concat_layer]
dep_context_representations = [dynet.tanh(dynet.affine_transform([self.dep_layer_b.expr(),
self.dep_layer_W.expr(),
context])) \
for context in concat_layer]
pairrels = id2arg2rel[sentenceId]
for (head, dep) in pairrels:
# print (head, dep), pairrels[(head, dep)]
linear = self.rel_U.expr(
) * dep_context_representations[dep]
linear = dynet.reshape(linear,
(self.hidden_units, len(self.rels)))
bilinear = dynet.transpose(
head_context_representations[head]) * linear
biaffine = dynet.transpose(
bilinear) + self.rel_W.expr() * concatenate([
head_context_representations[head],
dep_context_representations[dep]
]) + self.rel_b.expr()
lerrs.append(
self.pick_neg_log(softmax(biaffine),
self.rels.get(pairrels[(head,
dep)])))
etotal += len(sentence)
nwtotal += len(sentence)
if iSentence % 1 == 0:
if len(lerrs) > 0 or len(nerErrs) > 0:
# if len(nerErrs) > 0:
eerrs = esum(nerErrs + lerrs)
eerrs.scalar_value()
eloss += eerrs.scalar_value()
mloss += eloss
eerrs.backward()
self.trainer.update()
# errs = []
lerrs = []
nerErrs = []
renew_cg()
print "Loss: %.4f" % (mloss / nwtotal)
class Learner:
def __init__(self, c2i, options):
self.model = dy.ParameterCollection()
random.seed(1)
self.trainer = dy.AdamTrainer(self.model)
self.dropput_rate = options.dropout_rate
self.ldims = options.lstm_dims
self.cdims = options.cembedding_dims
self.c2i = c2i
self.W_d = self.model.add_parameters((self.ldims, 2 * self.ldims))
self.W_db = self.model.add_parameters(self.ldims)
self.clookup = self.model.add_lookup_parameters((len(c2i), self.cdims))
self.word_encoder = RNNSequencePredictor(
dy.VanillaLSTMBuilder(1, self.cdims, self.ldims, self.model))
self.context_encoder = [
dy.VanillaLSTMBuilder(1, self.ldims, self.ldims, self.model),
dy.VanillaLSTMBuilder(1, self.ldims, self.ldims, self.model)
]
self.output_encoder = dy.VanillaLSTMBuilder(1, self.cdims, self.ldims,
self.model)
self.decoder = dy.VanillaLSTMBuilder(2, self.cdims, self.ldims,
self.model)
self.W_s = self.model.add_parameters((len(self.c2i), self.ldims))
self.W_sb = self.model.add_parameters((len(self.c2i)))
def save(self, filename):
self.model.save(filename)
def load(self, filename):
self.model.populate(filename)
def predict(self, conll_path):
with open(conll_path, 'r') as conllFP:
for iSentence, sentence in enumerate(
read_conll(
conllFP,
self.c2i,
)):
dy.renew_cg()
conll_sentence = [
entry for entry in sentence
if isinstance(entry, utils.ConllEntry)
]
# I- Word encoding
for entry in conll_sentence:
c_embeddings = []
for c in entry.idChars:
# TODO : try different formulas like alpha/(alpha + #(w))
dropFlag = False # random.random() < self.dropput_rate
c_embedding = self.clookup[c if not dropFlag else 0]
c_embeddings.append(c_embedding)
e_i = self.word_encoder.predict_sequence(c_embeddings)[-1]
entry.word_enc = dy.dropout(e_i, self.dropput_rate)
entry.context_lstms = [entry.word_enc, entry.word_enc]
# II- Context encoding
blstm_forward = self.context_encoder[0].initial_state()
blstm_backward = self.context_encoder[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.word_enc)
blstm_backward = blstm_backward.add_input(rentry.word_enc)
entry.context_lstms[1] = blstm_forward.output()
rentry.context_lstms[0] = blstm_backward.output()
for entry in conll_sentence:
entry.context_enc = dy.concatenate(entry.context_lstms)
"""
# III- Output encoding
c_embeddings = []
for entry in conll_sentence:
for f in entry.idFeats:
# TODO : try different formulas like alpha/(alpha + #(w))
dropFlag = False # random.random() < self.dropput_rate
c_embedding = self.clookup[f if not dropFlag else 0]
c_embeddings.append(c_embedding)
t_i = self.output_encoder.predict_sequence(c_embeddings)
for entry, vec in zip(conll_sentence, t_i):
entry.output_enc = dy.dropout(vec, self.dropput_rate)
"""
# IV- Decoder
# Init for Context encoding
for entry in conll_sentence:
entry.context_enc = dy.rectify(self.W_d.expr() *
entry.context_enc +
self.W_db.expr())
"""
# Init for Word and Output encoding
for entry in conll_sentence:
entry.comb = entry.word_enc + entry.output_enc
"""
# predicted_sequences = []
output_state = self.output_encoder.initial_state()
for entry in conll_sentence:
if output_state.output():
entry.comb = entry.word_enc + output_state.output()
else:
entry.comb = entry.word_enc
decoder_state = self.decoder.initial_state().set_s([
entry.context_enc,
dy.tanh(entry.context_enc), entry.comb,
dy.tanh(entry.comb)
])
predicted_sequence = []
predicted_char = self.c2i["<s>"]
counter = 0
while True:
counter += 1
decoder_state.add_input(self.clookup[predicted_char])
probs = self._get_probs(decoder_state.output())
predicted_char = probs.npvalue().argmax()
if predicted_char != self.c2i["</s>"] and counter < 50:
predicted_sequence.append(predicted_char)
else:
break
for seq_i in predicted_sequence:
tag_embedding = self.clookup[seq_i]
decoder_state.add_input(tag_embedding)
entry.predicted_sequence = predicted_sequence
# predicted_sequences.append(predicted_sequence)
yield conll_sentence
def _get_probs(self, rnn_output):
output_w = dy.parameter(self.W_s)
output_b = dy.parameter(self.W_sb)
probs = dy.softmax(output_w * rnn_output + output_b)
return probs
def train(self, conll_path):
total = 0.0
with open(conll_path, 'r') as conllFP:
shuffledData = list(read_conll(conllFP, self.c2i))
random.shuffle(shuffledData)
for iSentence, sentence in enumerate(shuffledData):
conll_sentence = [
entry for entry in sentence
if isinstance(entry, utils.ConllEntry)
]
# I- Word encoding
for entry in conll_sentence:
c_embeddings = []
for c in entry.idChars:
# TODO : try different formulas like alpha/(alpha + #(w))
dropFlag = False # random.random() < self.dropput_rate
c_embedding = self.clookup[c if not dropFlag else 0]
c_embeddings.append(c_embedding)
e_i = self.word_encoder.predict_sequence(c_embeddings)[-1]
entry.word_enc = dy.dropout(e_i, self.dropput_rate)
entry.context_lstms = [entry.word_enc, entry.word_enc]
# II- Context encoding
blstm_forward = self.context_encoder[0].initial_state()
blstm_backward = self.context_encoder[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.word_enc)
blstm_backward = blstm_backward.add_input(rentry.word_enc)
entry.context_lstms[1] = blstm_forward.output()
rentry.context_lstms[0] = blstm_backward.output()
for entry in conll_sentence:
entry.context_enc = dy.concatenate(entry.context_lstms)
# III- Output encoding
c_embeddings = []
for entry in conll_sentence:
for f in entry.idFeats:
# TODO : try different formulas like alpha/(alpha + #(w))
dropFlag = False # random.random() < self.dropput_rate
c_embedding = self.clookup[f if not dropFlag else 0]
c_embeddings.append(c_embedding)
output_encoder = RNNSequencePredictor(self.output_encoder)
t_i = output_encoder.predict_sequence(c_embeddings)
for entry, vec in zip(conll_sentence, t_i):
entry.output_enc = dy.dropout(vec, self.dropput_rate)
# IV- Decoder
# Init for Context encoding
for entry in conll_sentence:
entry.context_enc = dy.rectify(self.W_d.expr() *
entry.context_enc +
self.W_db.expr())
# Init for Word and Output encoding
for entry in conll_sentence:
entry.comb = entry.word_enc + entry.output_enc
probs = []
losses = []
for entry in conll_sentence:
decoder_state = self.decoder.initial_state().set_s([
entry.context_enc,
dy.tanh(entry.context_enc), entry.comb,
dy.tanh(entry.comb)
])
for g_c in entry.decoder_input:
decoder_state = decoder_state.add_input(
self.clookup[g_c])
p = self._get_probs(decoder_state.output())
probs.append(p)
losses += [
-dy.log(dy.pick(p, o))
for p, o in zip(probs, entry.decoder_input)
]
total_losses = dy.esum(losses)
cur_loss = total_losses.scalar_value()
total += cur_loss
total_losses.backward()
self.trainer.update()
if iSentence != 0 and iSentence % 500 == 0:
print("Sentence:" + str(iSentence) + " Loss:" +
str(total / (iSentence + 1)))
dy.renew_cg()
def train(self, conll_path):
total = 0.0
with open(conll_path, 'r') as conllFP:
shuffledData = list(read_conll(conllFP, self.c2i))
random.shuffle(shuffledData)
for iSentence, sentence in enumerate(shuffledData):
conll_sentence = [
entry for entry in sentence
if isinstance(entry, utils.ConllEntry)
]
# I- Word encoding
for entry in conll_sentence:
c_embeddings = []
for c in entry.idChars:
# TODO : try different formulas like alpha/(alpha + #(w))
dropFlag = False # random.random() < self.dropput_rate
c_embedding = self.clookup[c if not dropFlag else 0]
c_embeddings.append(c_embedding)
e_i = self.word_encoder.predict_sequence(c_embeddings)[-1]
entry.word_enc = dy.dropout(e_i, self.dropput_rate)
entry.context_lstms = [entry.word_enc, entry.word_enc]
# II- Context encoding
blstm_forward = self.context_encoder[0].initial_state()
blstm_backward = self.context_encoder[1].initial_state()
for entry, rentry in zip(conll_sentence,
reversed(conll_sentence)):
blstm_forward = blstm_forward.add_input(entry.word_enc)
blstm_backward = blstm_backward.add_input(rentry.word_enc)
entry.context_lstms[1] = blstm_forward.output()
rentry.context_lstms[0] = blstm_backward.output()
for entry in conll_sentence:
entry.context_enc = dy.concatenate(entry.context_lstms)
# III- Output encoding
c_embeddings = []
for entry in conll_sentence:
for f in entry.idFeats:
# TODO : try different formulas like alpha/(alpha + #(w))
dropFlag = False # random.random() < self.dropput_rate
c_embedding = self.clookup[f if not dropFlag else 0]
c_embeddings.append(c_embedding)
output_encoder = RNNSequencePredictor(self.output_encoder)
t_i = output_encoder.predict_sequence(c_embeddings)
for entry, vec in zip(conll_sentence, t_i):
entry.output_enc = dy.dropout(vec, self.dropput_rate)
# IV- Decoder
# Init for Context encoding
for entry in conll_sentence:
entry.context_enc = dy.rectify(self.W_d.expr() *
entry.context_enc +
self.W_db.expr())
# Init for Word and Output encoding
for entry in conll_sentence:
entry.comb = entry.word_enc + entry.output_enc
probs = []
losses = []
for entry in conll_sentence:
decoder_state = self.decoder.initial_state().set_s([
entry.context_enc,
dy.tanh(entry.context_enc), entry.comb,
dy.tanh(entry.comb)
])
for g_c in entry.decoder_input:
decoder_state = decoder_state.add_input(
self.clookup[g_c])
p = self._get_probs(decoder_state.output())
probs.append(p)
losses += [
-dy.log(dy.pick(p, o))
for p, o in zip(probs, entry.decoder_input)
]
total_losses = dy.esum(losses)
cur_loss = total_losses.scalar_value()
total += cur_loss
total_losses.backward()
self.trainer.update()
if iSentence != 0 and iSentence % 500 == 0:
print("Sentence:" + str(iSentence) + " Loss:" +
str(total / (iSentence + 1)))
dy.renew_cg()