def main(pred_path, gold_path):
pred_corpus = Corpus(pred_path)
train_corpus = Corpus(gold_path)
Nb = NaiveBayes(train_corpus, pred_corpus)
train_start_timestamp = time.time()
Nb.train(train_corpus)
train_elapsed = time.time() - train_start_timestamp
i = 0
for tweet in pred_corpus:
if i % 1000 == 0 and not i == 0:
print('\r%i' % i, file=sys.stderr, end='')
i += 1
tweet.prediction = Nb.predict(tweet)
eval_results = pred_corpus.evaluate()
print('\r' + pformat(eval_results) + '\n')
print(' == Macro Avg F_Score: %.2f ==\n' %
(eval_results['macro_avg_f_score']))
print(' == Eval Accuracy: %.2f ==\n' % (eval_results['accuracy']))
print(' %s\n' % ('-' * 78))
def main(pred_path: str, gold_path: str, epochs: int) -> None:
train_corpus = Corpus(train_path)
eval_corpus = Corpus(eval_path)
feature_extractor = TextualFeatureExtractor([train_corpus, eval_corpus],
caching=True)
classes = feature_extractor.classes
p = Perceptron(classes)
epoch = 0
while True:
epoch += 1
train_start_timestamp = time.time()
train_corpus.shuffle()
for i, tweet in enumerate(train_corpus):
feature_vector = feature_extractor.get_feature_vector(tweet)
prediction = p.train(feature_vector, tweet.emotion)
tweet.prediction = max(prediction.items(), key=itemgetter(1))[0]
if i % 1000 == 0 and not i == 0:
print('\r%i' % i, file=sys.stderr, end='')
train_elapsed = time.time() - train_start_timestamp
train_results = train_corpus.evaluate()
eval_start_timestamp = time.time()
eval_corpus.shuffle()
for i, tweet in enumerate(eval_corpus):
feature_vector = feature_extractor.get_feature_vector(tweet)
prediction = p.predict(feature_vector)
tweet.prediction = max(prediction.items(), key=itemgetter(1))[0]
eval_elapsed = time.time() - eval_start_timestamp
eval_results = eval_corpus.evaluate()
eval_width = 10
emotion_width = len(max(eval_results['emotions'],
key=lambda x: len(x)))
print('\rEpoch %i:' % epoch)
print(pformat(eval_results) + '\n')
print(' == Train Macro Avg F_Score: %.2f ==' %
(train_results['macro_avg_f_score']))
print(' == Eval Macro Avg F_Score: %.2f ==' %
(eval_results['macro_avg_f_score']))
print(' == Train Accuracy: %.2f ==' % (train_results['accuracy']))
print(' == Eval Accuracy: %.2f ==\n' % (eval_results['accuracy']))
print(' Time elapsed: Train: %.2f Eval %.2f' %
(train_elapsed, eval_elapsed))
print(' %s\n' % ('-' * 78))
def main(image_path: str, image_index_path: str, pred_path: str,
gold_path: str, epochs: int) -> None:
text_train_corpus = Corpus(train_path)
text_eval_corpus = Corpus(eval_path)
text_feature_extractor = TextualFeatureExtractor(
[text_train_corpus, text_eval_corpus], caching=True)
classes = text_feature_extractor.classes
text_classifier = Perceptron(classes)
predictions = {}
epoch = 0
while epoch < 10:
epoch += 1
train_start_timestamp = time.time()
text_train_corpus.shuffle()
for i, tweet in enumerate(text_train_corpus):
feature_vector = text_feature_extractor.get_feature_vector(tweet)
prediction = text_classifier.train(feature_vector, tweet.emotion)
# tweet.full_prediction = prediction
tweet.prediction = max(prediction.items(), key=itemgetter(1))[0]
predictions[tweet.tweet_id] = prediction
if i % 1000 == 0 and not i == 0:
print('\r%i' % i, file=sys.stderr, end='')
train_elapsed = time.time() - train_start_timestamp
train_results = text_train_corpus.evaluate()
eval_start_timestamp = time.time()
text_eval_corpus.shuffle()
for i, tweet in enumerate(text_eval_corpus):
feature_vector = text_feature_extractor.get_feature_vector(tweet)
prediction = text_classifier.predict(feature_vector)
# tweet.full_prediction = prediction
tweet.prediction = max(prediction.items(), key=itemgetter(1))[0]
predictions[tweet.tweet_id] = prediction
eval_elapsed = time.time() - eval_start_timestamp
eval_results = text_eval_corpus.evaluate()
eval_width = 10
emotion_width = len(max(eval_results['emotions'],
key=lambda x: len(x)))
print('\rEpoch %i:' % epoch)
print(pformat(eval_results) + '\n')
print(' == Macro Avg F_Score: %.2f ==' %
(train_results['macro_avg_f_score']))
print(' == Train Accuracy: %.2f ==' % (train_results['accuracy']))
print(' == Eval Accuracy: %.2f ==\n' % (eval_results['accuracy']))
print(' Time elapsed: Train: %.2f Eval %.2f' %
(train_elapsed, eval_elapsed))
print(' %s\n' % ('-' * 78))
image_train_corpus = Corpus(train_path,
image_path,
image_index_path,
image_res=(12, 12))
image_eval_corpus = Corpus(eval_path,
image_path,
image_index_path,
image_res=(12, 12))
image_feature_extractor = ImageFeatureExtractor(
[image_train_corpus, image_eval_corpus], [predictions], caching=True)
classes = image_feature_extractor.classes
image_classifier = Perceptron(classes)
epoch = 0
while True:
epoch += 1
train_start_timestamp = time.time()
image_train_corpus.shuffle()
for i, tweet in enumerate(image_train_corpus):
feature_vector = image_feature_extractor.get_feature_vector(tweet)
prediction = image_classifier.train(feature_vector, tweet.emotion)
# tweet.full_prediction = prediction
tweet.prediction = max(prediction.items(), key=itemgetter(1))[0]
if i % 1000 == 0 and not i == 0:
print('\r%i' % i, file=sys.stderr, end='')
train_elapsed = time.time() - train_start_timestamp
train_results = image_train_corpus.evaluate()
eval_start_timestamp = time.time()
image_eval_corpus.shuffle()
for i, tweet in enumerate(image_eval_corpus):
feature_vector = image_feature_extractor.get_feature_vector(tweet)
prediction = image_classifier.predict(feature_vector)
# tweet.full_prediction = prediction
tweet.prediction = max(prediction.items(), key=itemgetter(1))[0]
eval_elapsed = time.time() - eval_start_timestamp
eval_results = image_eval_corpus.evaluate()
eval_width = 10
emotion_width = len(max(eval_results['emotions'],
key=lambda x: len(x)))
print('\rEpoch %i:' % epoch)
print(pformat(eval_results) + '\n')
print(' == Train Macro Avg F_Score: %.2f ==' %
(train_results['macro_avg_f_score']))
print(' == Eval Macro Avg F_Score: %.2f ==' %
(eval_results['macro_avg_f_score']))
print(' == Train Accuracy: %.2f ==' % (train_results['accuracy']))
print(' == Eval Accuracy: %.2f ==\n' % (eval_results['accuracy']))
print(' %s\n' % ('-' * 78))
def main(pred_path: str, gold_path: str, epochs: int) -> None:
# train_longest_tweet = max((len(x) for tweet in train_corpus for x in tweet.body))
# eval_longest_tweet = max((len(x) for tweet in eval_corpus for x in tweet.body))
# longest_tweet = max(train_longest_tweet, eval_longest_tweet)
train_corpus = Corpus(train_path)
eval_corpus = Corpus(eval_path)
word_lookup = Lookup()
emotion_lookup = Lookup()
i = 1
j = 0
for tweet in train_corpus:
for word in tweet.content:
if word not in word_lookup:
word_lookup[word] = i
i += 1
tag = tweet.emotion
if tag not in emotion_lookup:
emotion_lookup[tag] = j
j += 1
word_lookup['__pad__'] = len(word_lookup)
word_lookup['__unk__'] = len(word_lookup)
emotion_lookup['__pad__'] = len(emotion_lookup)
emotion_lookup['__unk__'] = len(emotion_lookup)
reverse_emotion_lookup = {i: x for x, i in emotion_lookup.items()}
x1 = T.imatrix('x1')
# m = T.imatrix('y')
# x2 = T.ftensor4('x2')
y = T.ivector('y')
network = build_network(text_input_var=x1,
number_of_words=len(word_lookup),
word_embedding_length=30,
num_target=len(emotion_lookup))
prediction = lasagne.layers.get_output(network)
loss = lasagne.objectives.categorical_crossentropy(prediction, y)
# loss = lasagne.objectives.categorical_crossentropy(prediction, y) * T.reshape(T.cast(m, 'float32'), [-1])
# loss = T.sum(loss)/T.sum(m)
loss = T.mean(loss)
params = lasagne.layers.get_all_params(network)
updates = lasagne.updates.adam(loss, params)
train = theano.function(inputs=[x1, y],
outputs=[prediction, loss],
updates=updates,
allow_input_downcast=True,
on_unused_input='warn')
predict = theano.function(inputs=[x1],
outputs=[prediction],
allow_input_downcast=True,
on_unused_input='warn')
prev_error = -1
epoch = 0
xts, tis, es = prepare_data(train_corpus, word_lookup, emotion_lookup)
test_xts, test_tis, test_es = prepare_data(eval_corpus, word_lookup,
emotion_lookup)
try:
prev_error = -1
while True:
epoch += 1
errors = []
for texts, _, emotions in generate_batches(xts, tis, es,
word_lookup):
ys, err = train(texts, emotions)
errors.append(err)
error = sum(errors) / len(errors)
for texts, tweet_ids, _ in generate_batches(
xts, tis, es, word_lookup):
ys = predict(texts)[0]
for y, ti in zip(ys, tweet_ids):
pred = reverse_emotion_lookup[int(np.argmax(y))]
train_corpus.get_tweet(ti).prediction = pred
for texts, tweet_ids, _ in generate_batches(
test_xts, test_tis, test_es, word_lookup):
ys = predict(texts)[0]
for y, ti in zip(ys, tweet_ids):
pred = reverse_emotion_lookup[int(np.argmax(y))]
eval_corpus.get_tweet(ti).prediction = pred
train_results = train_corpus.evaluate()
eval_results = eval_corpus.evaluate()
eval_width = 10
emotion_width = len(
max(eval_results['emotions'], key=lambda x: len(x)))
if prev_error == -1:
print('\rEpoch %i error: %.6f' % (epoch, error))
else:
err_delta = error - prev_error
print('\rEpoch %i error: %.6f error_delta: %.6f' %
(epoch, error, err_delta))
prev_error = error
print(pformat(eval_results) + '\n')
print(' == Train Macro Avg F_Score: %.2f ==' %
(train_results['macro_avg_f_score']))
print(' == Eval Macro Avg F_Score: %.2f ==' %
(eval_results['macro_avg_f_score']))
print(' == Train Accuracy: %.2f ==' % (train_results['accuracy']))
print(' == Eval Accuracy: %.2f ==\n' % (eval_results['accuracy']))
# print(' Time elapsed: Train: %.2f Eval %.2f' %(train_elapsed, eval_elapsed))
print(' %s\n' % ('-' * 78))
except KeyboardInterrupt:
pass