def create_vocab_util_from_training_set(tr_input_filename: str) -> VocabUtil:
"""To keep things simple, we use the entire training set without filtering."""
tr_loader = LabeledDataLoader(tr_input_filename)
tr_tweets = tr_loader.parse_tokens_and_labels(tr_loader.load_lines())
tr_unique_tokens = set([item[0] for tweet in tr_tweets for item in tweet])
sorted_tr_tokens = sorted(tr_unique_tokens)
return VocabUtil(sorted_tr_tokens)
def create_vocab_util_from_training_set(tr_input_filename: str) -> VocabUtil:
tr_loader = LabeledDataLoader(tr_input_filename)
tr_labeled_tweets = tr_loader.parse_tokens_and_labels(
tr_loader.load_lines())
tr_unique_tokens = set([
item for labeled_tweet in tr_labeled_tweets
for item in labeled_tweet[0]
])
tr_sorted_tokens = sorted(tr_unique_tokens)
print(f"Creating VocabUtil from {len(tr_sorted_tokens)} unique tokens.")
return VocabUtil(tr_sorted_tokens)
def main_training():
print(f'Using TensorFlow version {tf.__version__}')
vu = create_vocab_util_from_training_set(TRAINING_INPUT_FILENAME)
nn_input_preparer = NNInputPreparer(vu, max_seq_len=MAX_SEQ_LEN)
tr_loader = LabeledDataLoader(TRAINING_INPUT_FILENAME)
tr_tweets = tr_loader.parse_tokens_and_labels(tr_loader.load_lines())
tr_tweets = nn_input_preparer.filter_out_long_sequences(tr_tweets)
print(
f'Training on {len(tr_tweets)} tweets, each no longer than {MAX_SEQ_LEN} tokens'
)
tr_irregular_inputs = [[
vu.nn_input_token_to_int[item[0]] for item in tweet
] for tweet in tr_tweets]
tr_irregular_targets = [[vu.nn_pos_to_int[item[1]] for item in tweet]
for tweet in tr_tweets]
tr_rectangular_inputs = nn_input_preparer.rectangularize_inputs(
tr_irregular_inputs)
tr_rectangular_targets = nn_input_preparer.rectangularize_targets(
tr_irregular_targets)
tr_targets_one_hot_encoded = nn_input_preparer.rectangular_targets_to_one_hot(
tr_rectangular_targets)
if CONTINUE_TRAINING:
print('Continuing training from', TRAINING_MODEL_FILENAME_TO_CONTINUE)
model = load_model(TRAINING_MODEL_FILENAME_TO_CONTINUE)
model.summary()
else:
print("Commencing new training run")
model_creator = LstmModelCreator(vu,
embedding_dim=EMBEDDING_DIM,
lstm_dim=LSTM_DIM,
mask_zero=MASK_ZERO)
model = model_creator.create_bi_lstm_model()
cp_filepath = BASE_DIR + 'ep_{epoch}_valacc_{val_accuracy:.5f}.h5'
checkpoint = ModelCheckpoint(cp_filepath,
monitor='val_accuracy',
verbose=1,
save_best_only=False)
rectangular_inputs, _, targets_one_hot_encoded = \
prep_validation_set(DEV_INPUT_FILENAME, nn_input_preparer, vu)
model.fit(x=tr_rectangular_inputs,
y=tr_targets_one_hot_encoded,
batch_size=32,
initial_epoch=INITIAL_EPOCH,
epochs=MAX_EPOCHS,
validation_data=(rectangular_inputs, targets_one_hot_encoded),
callbacks=[checkpoint])
def test_loads_expected_number_of_tweets(self):
# We need to use the dev set as the test set, so we will optimize based on this training set.
loader = LabeledDataLoader('../data/pos/train.conll')
train_tweets_with_labeled_tokens = loader.parse_raw_tokens_and_labels(loader.load_lines())
self.assertEqual(27893, len(train_tweets_with_labeled_tokens))
# The test data is NOT labeled, so we want to use this dev set as the test set.
loader = LabeledDataLoader('../data/pos/dev.conll')
dev_tweets_with_labeled_tokens = loader.parse_raw_tokens_and_labels(loader.load_lines())
self.assertEqual(4298, len(dev_tweets_with_labeled_tokens))
def prep_validation_set(input_filename: str,
nn_input_preparer: NNInputPreparer, vu: VocabUtil,
upsample: bool):
loader = LabeledDataLoader(input_filename)
labeled_tweets = loader.parse_tokens_and_labels(loader.load_lines())
labeled_tweets = nn_input_preparer.filter_out_long_tweets(labeled_tweets)
if upsample:
labeled_tweets = nn_input_preparer.crude_upsample(labeled_tweets)
irregular_inputs = [[
vu.nn_input_token_to_int[token] if token in vu.nn_input_token_to_int
else vu.nn_input_token_to_int['<OOV>'] for token in labeled_tweet[0]
] for labeled_tweet in labeled_tweets]
rectangular_inputs = nn_input_preparer.rectangularize_inputs(
irregular_inputs)
rectangular_targets = [tweet[1] for tweet in labeled_tweets]
targets_one_hot_encoded = nn_input_preparer.rectangular_targets_to_one_hot(
rectangular_targets)
return rectangular_inputs, rectangular_targets, targets_one_hot_encoded
def prep_validation_set(input_filename: str, nn_input_preparer: NNInputPreparer, vu: VocabUtil) \
-> Tuple[np.ndarray, np.ndarray, np.ndarray]:
loader = LabeledDataLoader(input_filename)
tweets = loader.parse_tokens_and_labels(loader.load_lines())
tweets = nn_input_preparer.filter_out_long_sequences(tweets)
print(
f'processing all not-too-long {len(tweets)} tweets from {input_filename}'
)
irregular_inputs = [[
vu.nn_input_token_to_int[item[0]] if item[0]
in vu.nn_input_token_to_int else vu.nn_input_token_to_int['<OOV>']
for item in tweet
] for tweet in tweets]
irregular_targets = [[vu.nn_pos_to_int[item[1]] for item in tweet]
for tweet in tweets]
rectangular_inputs = nn_input_preparer.rectangularize_inputs(
irregular_inputs)
rectangular_targets = nn_input_preparer.rectangularize_targets(
irregular_targets)
targets_one_hot_encoded = nn_input_preparer.rectangular_targets_to_one_hot(
rectangular_targets)
return rectangular_inputs, rectangular_targets, targets_one_hot_encoded
def test_has_one_of_three_labels(self):
# We need to use the dev set as the test set, so we will optimize based on this training set.
loader = LabeledDataLoader('../data/sa/train.conll')
train_tweets_with_labels = loader.parse_tokens_and_labels(
loader.load_lines())
train_labels_set = set(
[labeled_tweet[1] for labeled_tweet in train_tweets_with_labels])
self.assertEqual({'positive', 'negative', 'neutral'}, train_labels_set)
loader = LabeledDataLoader('../data/sa/dev.conll')
dev_tweets_with_labels = loader.parse_tokens_and_labels(
loader.load_lines())
dev_labels_set = set(
[labeled_tweet[1] for labeled_tweet in dev_tweets_with_labels])
self.assertEqual({'positive', 'negative', 'neutral'}, dev_labels_set)
def test_loads_expected_number_of_tokens(self):
# We need to use the dev set as the test set, so we will optimize based on this training set.
loader = LabeledDataLoader('../data/pos/train.conll')
train_tweets_with_labeled_tokens = loader.parse_tokens_and_labels(
loader.load_lines())
self.assertEqual(
217068,
sum((len(tweet) for tweet in train_tweets_with_labeled_tokens)))
# The test data is NOT labeled, so we need to use this dev set as the test set.
loader = LabeledDataLoader('../data/pos/dev.conll')
dev_tweets_with_labeled_tokens = loader.parse_tokens_and_labels(
loader.load_lines())
self.assertEqual(
33345, sum(
(len(tweet) for tweet in dev_tweets_with_labeled_tokens)))
def main_training():
print(f'Using TensorFlow version {tf.__version__}')
tvu = TargetVocabUtil()
btc = BertTokenConverter(model_dir=BERT_PRETRAINED_MODEL_DIR, tvu=tvu)
nn_input_preparer = NNInputPreparer(tvu=tvu, max_seq_len=MAX_SEQ_LEN)
tr_loader = LabeledDataLoader(TR_INPUT_FILENAME)
tr_labeled_tweets = tr_loader.parse_tokens_and_labels(
tr_loader.load_lines())
tr_labeled_tweets = btc.convert(tr_labeled_tweets)
tr_labeled_tweets = btc.convert_to_ids(tr_labeled_tweets)
tr_labeled_tweets = btc.prepend_cls(tr_labeled_tweets)
tr_labeled_tweets = nn_input_preparer.filter_out_long_sequences(
tr_labeled_tweets)
print(
f'Processing all not-too-long {len(tr_labeled_tweets)} tweets from {TR_INPUT_FILENAME}'
)
tr_irregular_inputs = [tweet[0] for tweet in tr_labeled_tweets]
tr_rectangular_targets = [tweet[1] for tweet in tr_labeled_tweets]
tr_rectangular_inputs = nn_input_preparer.rectangularize_inputs(
tr_irregular_inputs)
tr_targets_one_hot_encoded = nn_input_preparer.rectangular_targets_to_one_hot(
tr_rectangular_targets)
if CONTINUE_TRAINING:
print('Continuing training from', TRAINING_MODEL_FILENAME_TO_CONTINUE)
model = load_model(TRAINING_MODEL_FILENAME_TO_CONTINUE)
model.summary()
else:
print("Commencing new training run")
model_creator = BertModelCreator(model_dir=BERT_PRETRAINED_MODEL_DIR,
tvu=tvu,
max_seq_len=MAX_SEQ_LEN)
model = model_creator.create_model()
cp_filepath = BASE_DIR + 'ep_{epoch}_valacc_{val_accuracy:.5f}.h5'
checkpoint = ModelCheckpoint(cp_filepath,
monitor='val_accuracy',
verbose=1,
save_best_only=False)
dev_loader = LabeledDataLoader(DEV_INPUT_FILENAME)
dev_labeled_tweets = tr_loader.parse_tokens_and_labels(
dev_loader.load_lines())
dev_labeled_tweets = btc.convert(dev_labeled_tweets)
dev_labeled_tweets = btc.convert_to_ids(dev_labeled_tweets)
dev_labeled_tweets = btc.prepend_cls(dev_labeled_tweets)
dev_labeled_tweets = nn_input_preparer.filter_out_long_sequences(
dev_labeled_tweets)
print(
f'Processing all not-too-long {len(dev_labeled_tweets)} tweets from {DEV_INPUT_FILENAME}'
)
dev_irregular_inputs = [tweet[0] for tweet in dev_labeled_tweets]
dev_rectangular_targets = [tweet[1] for tweet in dev_labeled_tweets]
dev_rectangular_inputs = nn_input_preparer.rectangularize_inputs(
dev_irregular_inputs)
dev_targets_one_hot_encoded = nn_input_preparer.rectangular_targets_to_one_hot(
dev_rectangular_targets)
model.fit(tr_rectangular_inputs,
tr_targets_one_hot_encoded,
batch_size=32,
initial_epoch=INITIAL_EPOCH,
epochs=MAX_EPOCHS,
validation_data=(dev_rectangular_inputs,
dev_targets_one_hot_encoded),
callbacks=[checkpoint])
model.save(FINAL_TRAINED_MODEL_FILENAME)
def main_training():
print(f'Using TensorFlow version {tf.__version__}')
tr_loader = LabeledDataLoader(TRAINING_INPUT_FILENAME)
tr_tweets = tr_loader.parse_raw_tokens_and_labels(tr_loader.load_lines())
tvu = TargetVocabUtil()
tokenizer = BertTokenConverter(model_dir=BERT_PRETRAINED_MODEL_DIR,
tvu=tvu)
tr_tweets = tokenizer.convert_to_tokens(tr_tweets)
tr_tweets = tokenizer.convert_to_ids(tr_tweets)
nn_input_preparer = NNInputPreparer(tvu, max_seq_len=MAX_SEQ_LEN)
tr_tweets = nn_input_preparer.filter_out_long_sequences(tr_tweets)
print(
f'Training on {len(tr_tweets)} tweets, each no longer than {MAX_SEQ_LEN} tokens'
)
tr_irregular_inputs = [[item[0] for item in tweet] for tweet in tr_tweets]
tr_irregular_targets = [[item[1] for item in tweet] for tweet in tr_tweets]
tr_rectangular_inputs = nn_input_preparer.rectangularize_inputs(
tr_irregular_inputs)
tr_rectangular_targets = nn_input_preparer.rectangularize_targets(
tr_irregular_targets)
tr_targets_one_hot_encoded = nn_input_preparer.rectangular_targets_to_one_hot(
tr_rectangular_targets)
if CONTINUE_TRAINING:
print('Continuing training from', TRAINING_MODEL_FILENAME_TO_CONTINUE)
model = load_model(TRAINING_MODEL_FILENAME_TO_CONTINUE,
custom_objects={"BertModelLayer": BertModelLayer})
model.summary()
else:
print('Commencing new training run')
model_creator = BertModelCreator(model_dir=BERT_PRETRAINED_MODEL_DIR,
tvu=tvu,
max_seq_len=MAX_SEQ_LEN,
freeze_bert_layer=True)
model = model_creator.create_model()
cp_filepath = BASE_DIR + 'ep_{epoch}_valacc_{val_accuracy:.5f}.h5'
dev_loader = LabeledDataLoader(DEV_INPUT_FILENAME)
dev_tweets = dev_loader.parse_raw_tokens_and_labels(
dev_loader.load_lines())
dev_tweets = tokenizer.convert_to_tokens(dev_tweets)
dev_tweets = tokenizer.convert_to_ids(dev_tweets)
dev_tweets = nn_input_preparer.filter_out_long_sequences(dev_tweets)
print(
f'processing all {len(dev_tweets)} not-too-long tweets from {DEV_INPUT_FILENAME}'
)
dev_irregular_inputs = [[item[0] for item in tweet]
for tweet in dev_tweets]
# print('proportion of non-pad in dev:', sum(len(tweet) for tweet in tweets) / (MAX_SEQ_LEN * len(tweets)))
dev_irregular_targets = [[item[1] for item in tweet]
for tweet in dev_tweets]
dev_rectangular_inputs = nn_input_preparer.rectangularize_inputs(
dev_irregular_inputs)
dev_rectangular_targets = nn_input_preparer.rectangularize_targets(
dev_irregular_targets)
dev_targets_one_hot_encoded = nn_input_preparer.rectangular_targets_to_one_hot(
dev_rectangular_targets)
checkpoint = ModelCheckpoint(cp_filepath,
monitor='val_accuracy',
verbose=1,
save_best_only=False)
model.fit(tr_rectangular_inputs,
tr_targets_one_hot_encoded,
batch_size=32,
initial_epoch=INITIAL_EPOCH,
epochs=MAX_EPOCHS,
validation_data=(dev_rectangular_inputs,
dev_targets_one_hot_encoded),
callbacks=[checkpoint])
model.save(FINAL_TRAINED_MODEL_FILENAME)
def main_inference():
print(f'Using TensorFlow version {tf.__version__}')
print(f'Loading model {TRAINING_MODEL_FILENAME}')
trained_model = load_model(
TRAINING_MODEL_FILENAME,
custom_objects={"BertModelLayer": BertModelLayer})
trained_model.summary()
tvu = TargetVocabUtil()
btc = BertTokenConverter(model_dir=BERT_PRETRAINED_MODEL_DIR, tvu=tvu)
nn_input_preparer = NNInputPreparer(tvu=tvu, max_seq_len=MAX_SEQ_LEN)
for input_filename in [DEV_INPUT_FILENAME]:
loader = LabeledDataLoader(input_filename)
tweets = loader.parse_tokens_and_labels(loader.load_lines())
tweets = btc.convert(tweets)
tweets = btc.convert_to_ids(tweets)
tweets = btc.prepend_cls(tweets)
tweets = nn_input_preparer.filter_out_long_sequences(tweets)
print(
f'Processing all not-too-long {len(tweets)} tweets from {input_filename}'
)
irregular_inputs = [tweet[0] for tweet in tweets]
rectangular_targets = [tweet[1] for tweet in tweets]
argmax_confusion_matrix = np.zeros(
(tvu.get_output_vocab_size(), tvu.get_output_vocab_size()),
dtype=int)
expected_sampling_confusion_matrix = np.zeros(
(tvu.get_output_vocab_size(), tvu.get_output_vocab_size()))
num_correct_argmax_predictions = 0
expected_sampling_accuracy_sum = 0.0
for irregular_input, target_index in tqdm(
zip(irregular_inputs, rectangular_targets)):
rectangular_input_singleton = nn_input_preparer.rectangularize_inputs(
[irregular_input])
predicted_probabilities = trained_model(
rectangular_input_singleton)[0]
# the predicted index if we take the class with the largest probability
argmax_index = np.argmax(predicted_probabilities)
if argmax_index == target_index:
num_correct_argmax_predictions += 1
argmax_confusion_matrix[target_index][argmax_index] += 1
# rhs is the probability of guessing target if we sample according to predicted probabilities
expected_sampling_accuracy_sum += tf.keras.backend.get_value(
predicted_probabilities[target_index])
for i in range(tvu.get_output_vocab_size()):
expected_sampling_confusion_matrix[target_index][
i] += predicted_probabilities[i]
num_tweets_in_dataset = len(rectangular_targets)
print(f'Argmax accuracy for {input_filename}:',
num_correct_argmax_predictions / num_tweets_in_dataset)
print(f'Expected sampling accuracy for {input_filename}:',
expected_sampling_accuracy_sum / num_tweets_in_dataset)
print(
f"Argmax confusion matrix of targets vs predicted for {input_filename}:\n"
f"{tvu.raw_sentiment_labels}\n", argmax_confusion_matrix)
print(
f"Expected sampling confusion matrix of targets vs predicted for {input_filename}:\n"
f"{tvu.raw_sentiment_labels}\n",
expected_sampling_confusion_matrix)
def main_inference():
print(f'Using TensorFlow version {tf.__version__}')
print(f'Loading model {TRAINING_MODEL_FILENAME}')
trained_model = load_model(
TRAINING_MODEL_FILENAME,
custom_objects={"BertModelLayer": BertModelLayer})
print('Loaded fine-tuned model:')
trained_model.summary()
tvu = TargetVocabUtil()
btc = BertTokenConverter(BERT_PRETRAINED_MODEL_DIR, tvu)
nn_input_preparer = NNInputPreparer(tvu=tvu, max_seq_len=MAX_SEQ_LEN)
for input_filename in [DEV_INPUT_FILENAME]:
loader = LabeledDataLoader(input_filename)
tweets = loader.parse_raw_tokens_and_labels(loader.load_lines())
tweets = btc.convert_to_tokens(tweets)
tweets = btc.convert_to_ids(tweets)
tweets = nn_input_preparer.filter_out_long_sequences(tweets)
print(
f'processing all {len(tweets)} not-too-long tweets from {input_filename}'
)
irregular_inputs = [[item[0] for item in tweet] for tweet in tweets]
irregular_targets = [[item[1] for item in tweet] for tweet in tweets]
num_tokens_in_dataset = num_token_level_correct_argmax_predictions = \
num_token_level_correct_argmax_predictions_incl_pads = 0
tweet_level_argmax_accuracy_sum = tweet_level_expected_sampling_accuracy_sum = \
token_level_expected_sampling_accuracy_sum = token_level_expected_sampling_accuracy_sum_incl_pads = 0.0
for (irregular_input, irregular_target_indices) in tqdm(
zip(irregular_inputs, irregular_targets)):
rectangular_inputs = nn_input_preparer.rectangularize_inputs(
[irregular_input])
rectangular_targets = nn_input_preparer.rectangularize_targets(
[irregular_target_indices])
num_tokens_in_current_tweet = num_current_tweet_correct_argmax_predictions = 0
current_tweet_expected_sampling_accuracy_sum = 0.0
predicted_probabilities_sequence = trained_model(
rectangular_inputs)
for predicted_probabilities, target_index in zip(
predicted_probabilities_sequence[0],
rectangular_targets[0]):
# the predicted index if we take the class with the largest probability
argmax_index = np.argmax(predicted_probabilities)
# probability of guessing target_index if we sample according to predicted probabilities
prob_sampling_success_on_token = tf.keras.backend.get_value(
predicted_probabilities[target_index])
if argmax_index == target_index:
num_token_level_correct_argmax_predictions_incl_pads += 1
token_level_expected_sampling_accuracy_sum_incl_pads += prob_sampling_success_on_token
if target_index != 0:
if argmax_index == target_index:
num_token_level_correct_argmax_predictions += 1
num_current_tweet_correct_argmax_predictions += 1
current_tweet_expected_sampling_accuracy_sum += prob_sampling_success_on_token
token_level_expected_sampling_accuracy_sum += prob_sampling_success_on_token
num_tokens_in_current_tweet += 1
num_tokens_in_dataset += 1
# every tweet has at least one non-padding token, so we don't worry about division by zero
current_tweet_argmax_accuracy = num_current_tweet_correct_argmax_predictions / num_tokens_in_current_tweet
current_tweet_expected_sampling_accuracy = \
current_tweet_expected_sampling_accuracy_sum / num_tokens_in_current_tweet
tweet_level_argmax_accuracy_sum += current_tweet_argmax_accuracy
tweet_level_expected_sampling_accuracy_sum += current_tweet_expected_sampling_accuracy
num_tweets_in_dataset = len(tweets)
num_tokens_in_dataset_incl_pads = MAX_SEQ_LEN * num_tweets_in_dataset
print(
f'Argmax accuracy for {input_filename} including padding:',
num_token_level_correct_argmax_predictions_incl_pads /
num_tokens_in_dataset_incl_pads)
print(
f'Expected sampling accuracy for {input_filename} including padding:',
token_level_expected_sampling_accuracy_sum_incl_pads /
num_tokens_in_dataset_incl_pads)
print(
f'Token-level argmax accuracy for {input_filename}:',
num_token_level_correct_argmax_predictions / num_tokens_in_dataset)
print(
f'Token-level expected sampling accuracy for {input_filename}:',
token_level_expected_sampling_accuracy_sum / num_tokens_in_dataset)
print(f'Tweet-level argmax accuracy for {input_filename}:',
tweet_level_argmax_accuracy_sum / num_tweets_in_dataset)
print(
f'Tweet-level expected sampling accuracy for {input_filename}:',
tweet_level_expected_sampling_accuracy_sum / num_tweets_in_dataset)