def train_bi_ltsm(train_data, n_epochs=1):
train_data = list(train_data)
vocab_words, vocab_tags = make_vocabs(train_data)
model = BiLSTM_Tagger(50, 300, len(vocab_words), len(vocab_tags))
loss_function = nn.NLLLoss()
optimizer = optim.SGD(model.parameters(), lr=0.1)
for epoch in range(
n_epochs
): # again, normally you would NOT do 300 epochs, it is toy data
random.shuffle(train_data)
c = 0
for sentence, tags in batchify(train_data):
c += 1
print(f"{c}/{len(train_data)}", end="\r")
# Step 1. Remember that Pytorch accumulates gradients.
# We need to clear them out before each instance
model.zero_grad()
# Step 2. Run our forward pass.
tag_scores = model(sentence)
# Step 3. Compute the loss, gradients, and update the parameters by
# calling optimizer.step()
loss = loss_function(tag_scores, tags)
loss.backward()
optimizer.step()
return model
def train_perceptron(train_data, n_epochs=1):
word_vocab, tag_vocab = make_vocabs(train_data)
parser = PerceptronParser(word_vocab, tag_vocab)
trainer = PerceptronTrainer(parser.model)
for sample in samples(train_data, parser):
features, gold_move = sample
trainer.update(features, gold_move)
trainer.finalize()
return parser
def train_eisner(train_data, n_epochs=1):
word_vocab, tag_vocab = make_vocabs(train_data)
parser = Eisner(word_vocab)
trainer = PerceptronTrainer(parser.model)
for sample in samples(train_data, parser):
#TODO Add training loop
trainer.finalize()
return parser
def train_neural(train_data, n_epochs=1, batch_size=300):
train_data = list(train_data) # because we will shuffle in-place
vocab_words, vocab_tags = make_vocabs(train_data)
classifier = NeuralParser(vocab_words, vocab_tags)
optimizer = optim.Adam(classifier.model.parameters())
for epoch in range(1, n_epochs+1):
random.shuffle(train_data)
for bx, by in batchify(train_data, batch_size, classifier, n_epochs):
optimizer.zero_grad()
output = classifier.model.forward(bx)
loss = F.cross_entropy(output, by)
loss.backward()
optimizer.step()
return classifier
def eval_feature_parser(train_file, dev_file):
with bz2.open(train_file, 'rt', encoding="utf-8") as source:
train_data = list(read_data(source))
with bz2.open(dev_file, 'rt', encoding="utf-8") as source:
dev_data = list(read_data(source))
train_data = filter_data(train_data, [1, 3, 6])
dev_data = filter_data(dev_data, [1, 3, 6])
vocab_words, vocab_tags = make_vocabs(train_data)
perceptron_parser = train_perceptron(train_data, n_epochs=EPOCHS)
print("UAS score for feature engineered perceptron:")
print("{:.4f}".format(uas(perceptron_parser, dev_data)))
print()
def eval_tagger(train_file, dev_file):
with bz2.open(train_file, 'rt', encoding="utf-8") as source:
train_data = list(read_data(source))
with bz2.open(dev_file, 'rt', encoding="utf-8") as source:
dev_data = list(read_data(source))
train_data = filter_data(train_data, [1, 3])
dev_data = filter_data(dev_data, [1, 3])
vocab_words, vocab_tags = make_vocabs(train_data)
encoded_train_data = encode(train_data, vocab_words, vocab_tags)
encoded_dev_data = encode(dev_data, vocab_words, vocab_tags)
bi_lstm_tagger = train_bi_ltsm(encoded_train_data, n_epochs=EPOCHS)
print("Tagger accuracy for BiLSTM:")
print("{:.4f}".format(accuracy(bi_lstm_tagger, encoded_dev_data)))
print()
def train_perceptron(train_data, n_epochs=1, encoded=True):
train_data = list(train_data) # because we will shuffle in-place
vocab_words, vocab_tags = make_vocabs(train_data)
tagger = GoldTagger(vocab_words, vocab_tags, encoded)
if encoded:
trainer = PerceptronTrainer(tagger.model)
else:
trainer = PerceptronTrainer(tagger.model, vocab_tags)
#MOST FREQ
pre_sufix = {PAD: 0}
for sentence in train_data:
for i, (w, tag) in enumerate(sentence):
if w[:1] not in pre_sufix:
pre_sufix[w[:1]] = len(pre_sufix)
if w[:2] not in pre_sufix:
pre_sufix[w[:2]] = len(pre_sufix)
if w[:3] not in pre_sufix:
pre_sufix[w[:3]] = len(pre_sufix)
if w[:4] not in pre_sufix:
pre_sufix[w[:4]] = len(pre_sufix)
if w[-1:] not in pre_sufix:
pre_sufix[w[-1:]] = len(pre_sufix)
if w[-2:] not in pre_sufix:
pre_sufix[w[-2:]] = len(pre_sufix)
if w[-3:] not in pre_sufix:
pre_sufix[w[-3:]] = len(pre_sufix)
if w[-4:] not in pre_sufix:
pre_sufix[w[-4:]] = len(pre_sufix)
tagger.pre_sufix = pre_sufix
for _ in range(n_epochs):
random.shuffle(train_data)
for i, sentence in enumerate(train_data):
words, gold_tags = zip(*sentence)
pred_tags = []
for i, gold_tag in enumerate(gold_tags):
features = tagger.featurize(
words, i, pred_tags,
sentence[i + 1][1] if i + 1 < len(sentence) else PAD)
trainer.update(features, gold_tag)
pred_tags.append(gold_tag)
trainer.finalize()
return tagger
def train_perceptron(train_data, n_epochs=1, encoded=True):
train_data = list(train_data) # because we will shuffle in-place
vocab_words, vocab_tags = make_vocabs(train_data)
tagger = PerceptronTagger(vocab_words, vocab_tags, encoded)
if encoded:
trainer = PerceptronTrainer(tagger.model)
else:
trainer = PerceptronTrainer(tagger.model, vocab_tags)
for _ in range(n_epochs):
random.shuffle(train_data)
for i, sentence in enumerate(train_data):
words, gold_tags = zip(*sentence)
pred_tags = []
for i, gold_tag in enumerate(gold_tags):
features = tagger.featurize(words, i, pred_tags)
trainer.update(features, gold_tag)
pred_tags.append(gold_tag)
trainer.finalize()
return tagger
def eval_parser(train_file, dev_file):
with bz2.open(train_file, 'rt', encoding="utf-8") as source:
train_data = list(read_data(source))
with bz2.open(dev_file, 'rt', encoding="utf-8") as source:
dev_data = list(read_data(source))
train_data = filter_data(train_data, [1, 3, 6])
dev_data = filter_data(dev_data, [1, 3, 6])
vocab_words, vocab_tags = make_vocabs(train_data)
perceptron_parser = train_perceptron(train_data, n_epochs=EPOCHS)
print("UAS score for perceptron:")
print("{:.4f}".format(uas(perceptron_parser, dev_data)))
print()
# L4 read gives: ~0.6698
# Our read gives: ~0.6643
neural_parser = train_neural(train_data, n_epochs=EPOCHS)
print("UAS score for neural:")
print("{:.4f}".format(uas(neural_parser, dev_data)))
print()
def eval_tagger(train_file, dev_file):
with bz2.open(train_file, 'rt', encoding="utf-8") as source:
train_data = list(read_data(source))
with bz2.open(dev_file, 'rt', encoding="utf-8") as source:
dev_data = list(read_data(source))
train_data = filter_data(train_data, [1, 3])
dev_data = filter_data(dev_data, [1, 3])
vocab_words, vocab_tags = make_vocabs(train_data)
encoded_train_data = encode(train_data, vocab_words, vocab_tags)
encoded_dev_data = encode(dev_data, vocab_words, vocab_tags)
perceptron_tagger = train_perceptron(encoded_train_data, n_epochs=EPOCHS)
print("Tagger accuracy for perceptron:")
print("{:.4f}".format(accuracy(perceptron_tagger, encoded_dev_data)))
print()
# L3 read gives: 0.8736
# Our read gives: 0.8736
neural_tagger = train_neural(encoded_train_data, n_epochs=EPOCHS)
print("Tagger accuracy for neural:")
print("{:.4f}".format(accuracy(neural_tagger, encoded_dev_data)))
print()
from lib import Eisner
from src import eisner_parser_trainer
from util.utils import read_data, filter_data, make_vocabs
from util.parser_utils import uas, output
import bz2
with bz2.open("files/train.conllu.bz2", 'rt') as source:
train_data = list(read_data(source))
with bz2.open("files/dev.conllu.bz2", 'rt') as source:
dev_data = list(read_data(source))
train_data = filter_data(train_data, [1, 3, 6])
dev_data = filter_data(dev_data, [1, 3, 6])
vocab_words, vocab_tags = make_vocabs(train_data)
#eisner_parser_trainer.train_eisner(train_data)
eisner = Eisner.Eisner(vocab_words)
tree = eisner.build_dependency_tree(train_data[0])
print(tree)
