def test(embeddingsFile, existingModel, predictionsFile,
documents_are_sequences, sentenceCNN, charCNN, uniLSTM, useBERT):
embeddings, vocab = event_reader.load_embeddings(embeddingsFile,
vocab_size,
word_embedding_dim)
testSentences, testBookIndex = event_reader.prepare_annotations_from_folder(
test_folder, documents_are_sequences, useBERT)
testC, testX, testP, testW, testY, testL = transform_examples(
testSentences, vocab, useBERT)
test_generator = single_generator(testC, testX, testP, testW, testY, testL)
test_metadata = event_reader.convert_to_index(testSentences)
model = event_cnn(embeddings, sentenceCNN, charCNN, uniLSTM, useBERT)
model.load_weights(existingModel)
predictionFile = predictionsFile
out = open(predictionFile, "w", encoding="utf-8")
gold = []
preds = []
c = 0
for step in range(len(testL)):
batch, y = next(test_generator)
probs = model.predict_on_batch(batch)
_, length, _ = y.shape
for i in range(length):
out.write("%s\t%s\t%s\t%.20f\n" %
('\t'.join([str(x) for x in test_metadata[c]]),
int(probs[0][i][0] > 0.5), y[0][i][0], probs[0][i][0]))
preds.append(probs[0][i][0] >= 0.5)
gold.append(y[0][i][0])
c += 1
f, p, r, correct, trials, trues = event_eval.check_f1_two_lists(
gold, preds)
print("precision: %.3f %s/%s" % (p, correct, trials))
print("recall: %.3f %s/%s" % (r, correct, trues))
print("F: %.3f" % f)
event_eval.check_f1_two_lists(gold, preds)
out.close()
def train(embeddingsFile, writePath, documents_are_sequences, sentenceCNN,
charCNN, pad, uniLSTM, useBERT):
global devC, devX, devP, devW, devY, devL
embeddings, vocab = event_reader.load_embeddings(embeddingsFile,
vocab_size,
word_embedding_dim)
trainSentences, trainBookIndex = event_reader.prepare_annotations_from_folder(
train_folder, documents_are_sequences, useBERT)
devSentences, devBookIndex = event_reader.prepare_annotations_from_folder(
dev_folder, documents_are_sequences, useBERT)
devC, devX, devP, devW, devY, devL = transform_examples(
devSentences, vocab, useBERT)
trainC, trainX, trainP, trainW, trainY, trainL = transform_examples(
trainSentences, vocab, useBERT)
max_sequence_length = 0
for length in trainL:
if length > max_sequence_length:
max_sequence_length = length
for length in devL:
if length > max_sequence_length:
max_sequence_length = length
print("Max length: %s" % max_sequence_length)
model = event_cnn(embeddings, sentenceCNN, charCNN, uniLSTM, useBERT)
tensorboard = TensorBoard(log_dir="logs/{}".format(time()))
checkpoint = ModelCheckpoint("%s.hdf5" % writePath,
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='min')
early_stopping = EarlyStopping(monitor='val_loss',
min_delta=0,
patience=15,
verbose=0,
mode='auto')
if pad:
trainC, trainX, trainP, trainW, trainY, trainL = pad_all(
trainC, trainX, trainP, trainW, trainY, trainL,
max_sequence_length, useBERT)
devC, devX, devP, devW, devY, devL = pad_all(devC, devX, devP, devW,
devY, devL,
max_sequence_length,
useBERT)
batchF1 = BatchF1()
model.fit([trainC, trainW, trainX, trainP],
trainY,
validation_data=([devC, devW, devX, devP], devY),
epochs=num_epochs,
batch_size=batch_size,
callbacks=[batchF1, tensorboard, checkpoint, early_stopping])
else:
train_generator = single_generator(trainC, trainX, trainP, trainW,
trainY, trainL)
dev_generator = single_generator(devC, devX, devP, devW, devY, devL)
generatorF1 = GeneratorF1()
model.fit_generator(
train_generator,
steps_per_epoch=len(trainL),
validation_data=dev_generator,
validation_steps=len(devL),
epochs=num_epochs,
callbacks=[generatorF1, tensorboard, checkpoint, early_stopping])
import event_eval
import event_reader
if __name__ == "__main__":
outputFile = sys.argv[1]
nlp = spacy.load('en', disable=['ner,parser'])
nlp.remove_pipe('ner')
nlp.remove_pipe('parser')
train_folder = "../data/bert/train"
dev_folder = "../data/bert/dev"
test_folder = "../data/bert/test"
testSentences, _ = event_reader.prepare_annotations_from_folder(
test_folder)
test_metadata = event_reader.convert_to_index(testSentences)
golds = []
preds = []
for sentence in testSentences:
tokens_list = [word[0] for word in sentence]
tokens = nlp.tokenizer.tokens_from_list(tokens_list)
nlp.tagger(tokens)
for idx, token in enumerate(tokens):
pred = 0
if token.tag_.startswith("V"):
pred = 1
preds.append(pred)
label = sentence[idx][1]
print(bestModel)
final_test_preds=bestModel.predict(test_X)
out=open(outputFile, "w", encoding="utf-8")
for idx, pred in enumerate(final_test_preds):
out.write("%s\t%s\n" % ('\t'.join([str(x) for x in test_metadata[idx]]), pred))
out.close()
if __name__ == "__main__":
read_countability("../data/esl.cd")
read_embeddings("../data/guten.vectors.txt")
trainSentences, _ = event_reader.prepare_annotations_from_folder(train_folder)
devSentences, _ = event_reader.prepare_annotations_from_folder(dev_folder)
testSentences, _ = event_reader.prepare_annotations_from_folder(test_folder)
test_metadata=event_reader.convert_to_index(testSentences)
functions=[get_word, get_context, get_dep, get_lemma, get_wordnet, get_pos, get_nsubj_features, get_embedding]
train_eval(trainSentences, devSentences, testSentences, "../results/featurized.preds.txt", functions)
functions=[get_context, get_dep, get_lemma, get_wordnet, get_pos, get_nsubj_features, get_embedding]
train_eval(trainSentences, devSentences, testSentences, "../results/featurized.word_ablation.preds.txt", functions)
functions=[get_word, get_context, get_dep, get_wordnet, get_pos, get_nsubj_features, get_embedding]
train_eval(trainSentences, devSentences, testSentences, "../results/featurized.lemma_ablation.preds.txt", functions)
functions=[get_word, get_context, get_dep, get_lemma, get_wordnet, get_nsubj_features, get_embedding]