def run_distilibert(strategy: tf.distribute.TPUStrategy, x_train: np.array,
x_valid: np.array, _y_train: np.array, y_valid: np.array,
train_dataset: tf.data.Dataset,
valid_dataset: tf.data.Dataset,
test_dataset: tf.data.Dataset, max_len: int, epochs: int,
batch_size: int) -> tf.keras.models.Model:
"""
create and run distilbert on training and testing data
"""
logger.info('build distilbert')
with strategy.scope():
transformer_layer = TFDistilBertModel.from_pretrained(MODEL)
model = build_model(transformer_layer, max_len=max_len)
model.summary()
# train given model
n_steps = x_train.shape[0] // batch_size
history = model.fit(train_dataset,
steps_per_epoch=n_steps,
validation_data=valid_dataset,
epochs=epochs)
plot_train_val_loss(history, 'distilbert')
n_steps = x_valid.shape[0] // batch_size
_train_history_2 = model.fit(valid_dataset.repeat(),
steps_per_epoch=n_steps,
epochs=epochs * 2)
scores = model.predict(test_dataset, verbose=1)
logger.info(f"AUC: {roc_auc(scores, y_valid):.4f}")
return model
def simple_rnn(strategy: tf.distribute.TPUStrategy, x_train_padded: np.array,
x_valid_padded: np.array, y_train: np.array, y_valid: np.array,
max_len: int, embedding_size_x: int, embedding_size_y: int, epochs: int) -> tf.keras.models.Sequential:
"""
create and run simple rnn on training and testing data
"""
logger.info('build simple RNN')
with strategy.scope():
# A simpleRNN without any pretrained embeddings and one dense layer
model = tf.keras.models.Sequential()
model.add(tf.keras.layers.Embedding(embedding_size_x, embedding_size_y,
input_length=max_len))
model.add(tf.keras.layers.SimpleRNN(100))
model.add(tf.keras.layers.Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy',
optimizer='adam', metrics=['accuracy'])
model.summary()
# run model train
history = model.fit(x_train_padded, y_train, epochs=epochs, batch_size=64 *
strategy.num_replicas_in_sync)
plot_train_val_loss(history, 'simple_rnn')
scores = model.predict(x_valid_padded)
logger.info(f"AUC: {roc_auc(scores, y_valid):.4f}")
return model
def run_gru(strategy: tf.distribute.TPUStrategy, x_train_padded: np.array,
x_valid_padded: np.array, y_train: np.array, y_valid: np.array,
max_len: int, embedding_size_x: int, embedding_size_y: int,
embedding_matrix: np.array, epochs: int) -> tf.keras.models.Sequential:
"""
create and run gru on training and testing data
"""
logger.info('build gru')
# build GRU model
with strategy.scope():
# GRU with glove embeddings and dense layer
model = tf.keras.models.Sequential()
model.add(tf.keras.layers.Embedding(embedding_size_x,
embedding_size_y,
weights=[embedding_matrix],
input_length=max_len,
trainable=False))
model.add(tf.keras.layers.SpatialDropout1D(0.3))
model.add(tf.keras.layers.GRU(embedding_size_y))
model.add(tf.keras.layers.Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy',
optimizer='adam', metrics=['accuracy'])
model.summary()
history = model.fit(x_train_padded, y_train, epochs=epochs, batch_size=64 *
strategy.num_replicas_in_sync)
plot_train_val_loss(history, 'gru')
scores = model.predict(x_valid_padded)
logger.info(f"AUC: {roc_auc(scores, y_valid):.4f}")
return model
def run_rnn(strategy: tf.distribute.TPUStrategy, x_train_padded: np.array,
x_valid_padded: np.array, y_train: np.array, y_valid: np.array,
max_len: int, embedding_size_x: int, embedding_size_y: int,
embedding_matrix: np.array,
epochs: int) -> tf.keras.models.Sequential:
"""
create and run bidirectional rnn on training and testing data
"""
logger.info('build rnn')
with strategy.scope():
# A simple bidirectional LSTM with glove embeddings and one dense layer
model = tf.keras.models.Sequential()
model.add(
tf.keras.layers.Embedding(embedding_size_x,
embedding_size_y,
weights=[embedding_matrix],
input_length=max_len,
trainable=False))
model.add(
tf.keras.layers.Bidirectional(
tf.keras.layers.LSTM(embedding_size_y,
dropout=0.3,
recurrent_dropout=0.3)))
model.add(tf.keras.layers.Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['accuracy'])
model.summary()
# fit rnn model, plot the training output
history = model.fit(x_train_padded,
y_train,
epochs=epochs,
batch_size=64 * strategy.num_replicas_in_sync)
plot_train_val_loss(history, 'bidirectional_lstm')
scores = model.predict(x_valid_padded)
logger.info(f"AUC: {roc_auc(scores, y_valid):.4f}")
return model
def train_test(
clean_data: pd.DataFrame
) -> Tuple[tf.keras.models.Sequential, tf.keras.models.Sequential]:
"""
train test
creates the tensorflow models for sentiment analysis
"""
logger.info('run training and testing for lstm and cnn')
all_reviews: List[str] = clean_data[review_key]
labels: List[int] = clean_data[class_key]
train_reviews, test_reviews, train_labels, test_labels = train_test_split(
all_reviews, labels, test_size=TEST_SIZE)
train_reviews, validation_reviews, train_labels, validation_labels = train_test_split(
train_reviews, train_labels, test_size=VALIDATION_SIZE)
# create the training, validation, and testing datasets
training_dataset = tf.data.Dataset.from_tensor_slices(
(train_reviews, train_labels))
validation_dataset = tf.data.Dataset.from_tensor_slices(
(validation_reviews, validation_labels))
testing_dataset = tf.data.Dataset.from_tensor_slices(
(test_reviews, test_labels))
# buffer size is used to shuffle the dataset
buffer_size = 10000
training_dataset = training_dataset.shuffle(buffer_size).batch(
batch_size, drop_remainder=True)
validation_dataset = validation_dataset.shuffle(buffer_size).batch(
batch_size, drop_remainder=True)
testing_dataset = testing_dataset.shuffle(buffer_size).batch(
batch_size, drop_remainder=True)
# print some samples
logger.success('training data sample:')
for input_example, target_example in training_dataset.take(1):
logger.info(f"\ninput: {input_example}\ntarget: {target_example}")
vocab_size = 10000
sequence_length = 250
vectorize_layer = TextVectorization(standardize=standardize_text,
max_tokens=vocab_size,
output_mode='int',
output_sequence_length=sequence_length)
autotune = tf.data.experimental.AUTOTUNE
training_dataset = training_dataset.cache().prefetch(buffer_size=autotune)
train_text = training_dataset.map(lambda x, y: x)
vectorize_layer.adapt(train_text)
embedding_dim = 16
embedding_layer = tf.keras.layers.Embedding(vocab_size, embedding_dim)
output_layer = tf.keras.layers.Dense(1, activation='sigmoid')
# create lstm model
lstm_model = tf.keras.models.Sequential([
vectorize_layer,
embedding_layer,
tf.keras.layers.LSTM(256),
tf.keras.layers.Dense(64, activation='relu'),
output_layer,
])
learning_rate: int = 1e-3
optimizer = tf.keras.optimizers.Adam(learning_rate)
loss = tf.keras.losses.BinaryCrossentropy(from_logits=True)
# save accuracy after each epoch
metrics = [tf.keras.metrics.BinaryAccuracy()]
lstm_model.compile(optimizer=optimizer, loss=loss, metrics=metrics)
_lstm_callbacks = [UseMaxWeights()]
# tried setting callbacks in the fit function below to lstm_callbacks
# to use the max of all hidden states as the context vector for prediction
# it did not work as well as using the last hidden state, so I am
# not using the callback
# run training and save history
hist = lstm_model.fit(training_dataset,
epochs=12,
callbacks=[],
validation_data=validation_dataset)
plot_train_val_loss(hist, 'reviews_lstm')
logger.info('lstm model summary:')
lstm_model.summary()
loss_metric, accuracy = lstm_model.evaluate(testing_dataset)
logger.info(f'loss: {loss_metric}, accuracy: {accuracy}')
classes = range(2)
# don't print these metrics for equally-sized datasets
# precision, recall, f_score, support = get_precision_recall_fscore(
# lstm_model, testing_dataset, classes)
# for i in classes:
# logger.info(f'{reviews_class_map[i]}:')
# logger.info(f'precision: {precision[i]}')
# logger.info(f'recall: {recall[i]}')
# logger.info(f'f-score: {f_score[i]}')
# logger.info(f'support: {support[i]}')
# create cnn model
cnn_model = tf.keras.models.Sequential([
vectorize_layer,
embedding_layer,
tf.keras.layers.ZeroPadding1D(1),
tf.keras.layers.Conv1D(32, 2, activation='relu'),
tf.keras.layers.MaxPooling1D(2),
tf.keras.layers.Conv1D(64, 3, activation='relu'),
tf.keras.layers.MaxPooling1D(3),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(64, activation='relu'),
output_layer,
])
cnn_model.compile(optimizer=optimizer, loss=loss, metrics=['accuracy'])
# save cnn model history
hist = cnn_model.fit(training_dataset,
epochs=16,
callbacks=[],
validation_data=validation_dataset)
plot_train_val_loss(hist, 'reviews_cnn')
logger.info('cnn model summary')
cnn_model.summary()
# print the loss and accuracy at the end on the testing dataset
loss_metric, accuracy = cnn_model.evaluate(testing_dataset)
logger.info(f'loss: {loss_metric}, accuracy: {accuracy}')
# precision, recall, f_score, support = get_precision_recall_fscore(
# cnn_model, testing_dataset, classes)
# for i in classes:
# logger.info(f'{reviews_class_map[i]}:')
# logger.info(f'precision: {precision[i]}')
# logger.info(f'recall: {recall[i]}')
# logger.info(f'f-score: {f_score[i]}')
# logger.info(f'support: {support[i]}')
return lstm_model, cnn_model
def train_test(
clean_data: pd.DataFrame, label_list: List[BookType]
) -> Tuple[tf.keras.models.Sequential, tf.keras.models.Sequential]:
"""
train test
run training and testing for book classification
"""
logger.info('run training and testing for lstm and cnn')
all_paragraphs: List[str] = [
' '.join(paragraph) for paragraph in clean_data[paragraph_key]
]
labels: List[int] = clean_data[class_key]
train_paragraphs, test_paragraphs, train_labels, test_labels = train_test_split(
all_paragraphs, labels, test_size=TEST_SIZE)
train_paragraphs, validation_paragraphs, train_labels, validation_labels = train_test_split(
train_paragraphs, train_labels, test_size=VALIDATION_SIZE)
# create training, validation and testing datasets
training_dataset = tf.data.Dataset.from_tensor_slices(
(train_paragraphs, train_labels))
validation_dataset = tf.data.Dataset.from_tensor_slices(
(validation_paragraphs, validation_labels))
testing_dataset = tf.data.Dataset.from_tensor_slices(
(test_paragraphs, test_labels))
# buffer size is used to shuffle the dataset
buffer_size = 10000
# shuffle and batch datasets
training_dataset = training_dataset.shuffle(buffer_size).batch(
batch_size, drop_remainder=True)
validation_dataset = validation_dataset.shuffle(buffer_size).batch(
batch_size, drop_remainder=True)
testing_dataset = testing_dataset.shuffle(buffer_size).batch(
batch_size, drop_remainder=True)
# print some samples
logger.success('training data sample:')
for input_example, target_example in training_dataset.take(1):
logger.info(f"\ninput: {input_example}\ntarget: {target_example}")
vocab_size = 10000
sequence_length = 250
# decided not to use pre-trained embedding layer (https://www.tensorflow.org/hub/tutorials/cord_19_embeddings_keras#training_a_citaton_intent_classifier)
# because I already wrote this text vectorizor system, and did not want to redo
# the same work. I tried using the universal sentence encoder from tf-hub:
# https://tfhub.dev/google/universal-sentence-encoder/4
# the proof of concept worked well, but implementing it would require changing
# all of the code below.
# create vectorization layer
vectorize_layer = TextVectorization(standardize=standardize_text,
max_tokens=vocab_size,
output_mode='int',
output_sequence_length=sequence_length)
autotune = tf.data.experimental.AUTOTUNE
training_dataset = training_dataset.cache().prefetch(buffer_size=autotune)
train_text = training_dataset.map(lambda x, y: x)
# adapt vectorization layer to text input
vectorize_layer.adapt(train_text)
embedding_dim = 16
embedding_layer = tf.keras.layers.Embedding(vocab_size, embedding_dim)
num_classes: int = len(label_list)
output_layer = tf.keras.layers.Dense(num_classes)
# create lstm model
lstm_model = tf.keras.models.Sequential([
vectorize_layer,
embedding_layer,
tf.keras.layers.LSTM(128),
tf.keras.layers.Dense(64, activation='relu'),
output_layer,
])
learning_rate: int = 1e-3
# create optimizer and loss function
optimizer = tf.keras.optimizers.Adam(learning_rate)
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
# save accuracy metric on each epoch
metrics = [tf.keras.metrics.SparseCategoricalAccuracy()]
lstm_model.compile(optimizer=optimizer, loss=loss, metrics=metrics)
# optional callback, described below
_lstm_callbacks = [UseMaxWeights()]
# tried setting callbacks in the fit function below to lstm_callbacks
# to use the max of all hidden states as the context vector for prediction
# it did not work as well as using the last hidden state, so I am
# not using the callback
# train model and save history
hist = lstm_model.fit(training_dataset,
epochs=12,
callbacks=[],
validation_data=validation_dataset)
# plot training loss
plot_train_val_loss(hist, 'books_lstm')
# print model summary
logger.info('lstm model summary:')
lstm_model.summary()
loss_metric, accuracy = lstm_model.evaluate(testing_dataset)
logger.info(f'loss: {loss_metric}, accuracy: {accuracy}')
classes = range(num_classes)
precision, recall, f_score, support = get_precision_recall_fscore(
lstm_model, testing_dataset, classes)
for i in classes:
current_book = label_list[i]
logger.info(f'{class_map[current_book]}:')
logger.info(f'precision: {precision[i]}')
logger.info(f'recall: {recall[i]}')
logger.info(f'f-score: {f_score[i]}')
logger.info(f'support: {support[i]}')
cnn_model = tf.keras.models.Sequential([
vectorize_layer,
embedding_layer,
tf.keras.layers.ZeroPadding1D(1),
tf.keras.layers.Conv1D(32, 3, activation='relu'),
tf.keras.layers.MaxPooling1D(3),
tf.keras.layers.Conv1D(24, 2, activation='relu'),
tf.keras.layers.MaxPooling1D(2),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(64, activation='relu'),
output_layer,
])
cnn_model.compile(optimizer=optimizer, loss=loss, metrics=['accuracy'])
hist = cnn_model.fit(training_dataset,
epochs=14,
callbacks=[],
validation_data=validation_dataset)
plot_train_val_loss(hist, 'books_cnn')
logger.info('cnn model summary')
cnn_model.summary()
loss_metric, accuracy = cnn_model.evaluate(testing_dataset)
logger.info(f'loss: {loss_metric}, accuracy: {accuracy}')
precision, recall, f_score, support = get_precision_recall_fscore(
cnn_model, testing_dataset, classes)
for i in classes:
current_book = label_list[i]
logger.info(f'{class_map[current_book]}:')
logger.info(f'precision: {precision[i]}')
logger.info(f'recall: {recall[i]}')
logger.info(f'f-score: {f_score[i]}')
logger.info(f'support: {support[i]}')
return lstm_model, cnn_model
logger.info('')
loss = torch.nn.MSELoss(reduction='sum')
# alternative loss: MAE
torch.nn.L1Loss(reduction='sum') # MAE
best_val_loss = math.inf
best_epoch = 0
logger.info("Starting training...")
for epoch in range(1, args.n_epochs):
train_loss = train(model, train_loader, optimizer, loss, device, scheduler,
logger if args.verbose else None)
logger.info("Epoch {}: Training Loss {}".format(epoch, train_loss))
val_loss = test(model, val_loader, loss, device, log_dir, epoch)
logger.info("Epoch {}: Validation Loss {}".format(epoch, val_loss))
if scheduler and not isinstance(scheduler, NoamLR):
scheduler.step(val_loss)
if val_loss <= best_val_loss:
best_val_loss = val_loss
best_epoch = epoch
# torch.save(model.state_dict(), os.path.join(log_dir, 'best_model'))
logger.info("Best Validation Loss {} on Epoch {}".format(
best_val_loss, best_epoch))
log_file = os.path.join(log_dir, log_file_name + '.log')
plot_train_val_loss(log_file)
