def get_summary(self, text, max_summary_length=50, num_beams=4, no_repeat_ngram_size=2, num_return_sequences=1,
early_stopping=True):
modelInfo = self.models.get("summarization")
model = modelInfo['model']
tokenizer = modelInfo['tokenizer']
text = [text]
text = add_prefix(text, "summarize: ")
result = model.generate(
tf.convert_to_tensor(tokenize_for_input_ids(text, tokenizer, max_length=modelInfo['max_text_length'])),
max_length=max_summary_length, num_beams=num_beams,
no_repeat_ngram_size=no_repeat_ngram_size, num_return_sequences=num_return_sequences,
early_stopping=early_stopping)
return [tokenizer.decode(summary) for summary in result]
def summarization_query(self,
instruction,
preprocess=True,
label_column=None,
drop=None,
epochs=5,
batch_size=32,
learning_rate=3e-5,
max_text_length=512,
gpu=False,
test_size=0.2,
random_state=49,
generate_plots=True,
save_model=False,
save_path=os.getcwd()):
'''
function to apply algorithm for text summarization
:param many params: used to hyperparametrize the function.
:return a dictionary object with all of the information for the algorithm.
'''
if test_size < 0:
raise Exception("Test size must be a float between 0 and 1")
if test_size >= 1:
raise Exception(
"Test size must be a float between 0 and 1 (a test size greater than or equal to 1 results in no training "
"data)")
if max_text_length < 2:
raise Exception("Text and summary must be at least of length 2")
if epochs < 1:
raise Exception(
"Epoch number is less than 1 (model will not be trained)")
if batch_size < 1:
raise Exception("Batch size must be equal to or greater than 1")
if max_text_length < 1:
raise Exception("Max text length must be equal to or greater than 1")
if save_model:
if not os.path.exists(save_path):
raise Exception("Save path does not exist")
if test_size == 0:
testing = False
else:
testing = True
if gpu:
if tf.test.gpu_device_name():
print('Default GPU Device: {}'.format(tf.test.gpu_device_name()))
else:
raise Exception("Please install GPU version of Tensorflow")
device = '/device:GPU:0'
else:
device = '/device:CPU:0'
tf.random.set_seed(random_state)
np.random.seed(random_state)
data = DataReader(self.dataset)
data = data.data_generator()
if drop is not None:
data.drop(drop, axis=1, inplace=True)
if preprocess:
data.fillna(0, inplace=True)
logger("Preprocessing data...")
if label_column is None:
label = "summary"
else:
label = label_column
tokenizer = T5Tokenizer.from_pretrained("t5-small")
# Find target columns
X, Y, target = get_target_values(data, instruction, label)
logger("->", "Target Column Found: {}".format(target))
logger("Establishing dataset walkers")
# Clean up text
if preprocess:
logger("Preprocessing data")
X = add_prefix(lemmatize_text(text_clean_up(X.array)), "summarize: ")
Y = add_prefix(lemmatize_text(text_clean_up(Y.array)), "summarize: ")
# tokenize text/summaries
X = tokenize_for_input_ids(X, tokenizer, max_text_length)
Y = tokenize_for_input_ids(Y, tokenizer, max_text_length)
logger('Fine-Tuning the model on your dataset...')
# Suppress unnecessary output
with NoStdStreams():
model = TFT5ForConditionalGeneration.from_pretrained(
"t5-small", output_loading_info=False)
if testing:
X_train, X_test, y_train, y_test = train_test_split(
X, Y, test_size=test_size, random_state=random_state)
test_dataset = tf.data.Dataset.from_tensor_slices(
(X_test, y_test)).shuffle(10000).batch(batch_size)
else:
X_train = X
y_train = Y
train_dataset = tf.data.Dataset.from_tensor_slices(
(X_train, y_train)).shuffle(10000).batch(batch_size)
optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate)
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
total_training_loss = []
total_validation_loss = []
# Training Loop
with tf.device(device):
for epoch in range(epochs):
total_loss = 0
total_loss_val = 0
for data, truth in train_dataset:
with tf.GradientTape() as tape:
out = model(inputs=data, decoder_input_ids=data)
loss_value = loss(truth, out[0])
total_loss += loss_value
grads = tape.gradient(loss_value, model.trainable_weights)
optimizer.apply_gradients(
zip(grads, model.trainable_weights))
total_training_loss.append(total_loss)
# Validation Loop
if testing:
for data, truth in test_dataset:
logits = model(inputs=data,
decoder_input_ids=data,
training=False)
val_loss = loss(truth, logits[0])
total_loss_val += val_loss
total_validation_loss.append(total_loss_val)
logger(
"->", "Final training loss: {}".format(
str(total_training_loss[len(total_training_loss) - 1].numpy())))
if testing:
total_loss_val_str = str(
total_validation_loss[len(total_validation_loss) - 1].numpy())
else:
total_loss_val = [0]
total_loss_val_str = str("0, No validation done")
logger("->", "Final validation loss: {}".format(total_loss_val_str))
if testing:
losses = {
"Training loss":
total_training_loss[len(total_training_loss) - 1].numpy(),
"Validation loss":
total_validation_loss[len(total_validation_loss) - 1].numpy()
}
else:
losses = {
"Training loss":
total_training_loss[len(total_training_loss) - 1].numpy()
}
plots = None
if generate_plots:
logger("Generating plots")
plots = {
"loss":
libra.plotting.nonkeras_generate_plots.plot_loss(
total_training_loss, total_validation_loss)
}
if save_model:
logger("Saving model")
model.save_weights(save_path + "summarization_checkpoint.ckpt")
logger("Storing information in client object under key 'summarization'")
self.models["summarization"] = {
"model": model,
"max_text_length": max_text_length,
"plots": plots,
"tokenizer": tokenizer,
'losses': losses
}
clearLog()
return self.models["summarization"]
