class TF(object):
def __init__(self,name,model_type,batch_size,learning_rate,num_train_epochs,use_crf): #default settings
self.batch_size = int(batch_size)
self.learning_rate = float(learning_rate)
self.epochs = int(num_train_epochs)
self.use_crf = use_crf
self.type_ = model_type
self.save_folder_path = os.path.join(os.path.join(os.getcwd(),'opennlu/data/model/tensorflow'),name)
tf.compat.v1.reset_default_graph()
self.sess = tf.compat.v1.Session()
self.graph = tf.get_default_graph()
set_session(self.sess)
#tf.compat.v1.random.set_random_seed(123)
#option 1: train new model
def train(self, train_data_dir, valid_data_dir):
self.train_data_folder_path = train_data_dir
self.val_data_folder_path = valid_data_dir
if self.type_ == 'bert':
bert_model_hub_path = 'https://tfhub.dev/google/bert_uncased_L-12_H-768_A-12/1'
is_bert = True
else:
bert_model_hub_path = 'https://tfhub.dev/google/albert_base/1'
is_bert = False
#read data
train_text_arr, train_tags_arr, train_intents = Reader.read(self.train_data_folder_path)
val_text_arr, val_tags_arr, val_intents = Reader.read(self.val_data_folder_path)
#vectorise data
self.bert_vectorizer = BERTVectorizer(self.sess, is_bert, bert_model_hub_path)
train_input_ids, train_input_mask, train_segment_ids, train_valid_positions, train_sequence_lengths = self.bert_vectorizer.transform(train_text_arr)
val_input_ids, val_input_mask, val_segment_ids, val_valid_positions, val_sequence_lengths = self.bert_vectorizer.transform(val_text_arr)
#vectorise tags
if self.use_crf: # with crf
self.tags_vectorizer = TagsVectorizer()
self.tags_vectorizer.fit(train_tags_arr)
train_tags = self.tags_vectorizer.transform(train_tags_arr, train_valid_positions)
train_tags = tf.keras.utils.to_categorical(train_tags)
val_tags = self.tags_vectorizer.transform(val_tags_arr, val_valid_positions)
val_tags = tf.keras.utils.to_categorical(val_tags)
slots_num = len(self.tags_vectorizer.label_encoder.classes_)
else: # without crf
self.tags_vectorizer = TagsVectorizer()
self.tags_vectorizer.fit(train_tags_arr)
train_tags = self.tags_vectorizer.transform(train_tags_arr, train_valid_positions)
val_tags = self.tags_vectorizer.transform(val_tags_arr, val_valid_positions)
slots_num = len(self.tags_vectorizer.label_encoder.classes_)
#encode labels
self.intents_label_encoder = LabelEncoder()
train_intents = self.intents_label_encoder.fit_transform(train_intents).astype(np.int32)
val_intents = self.intents_label_encoder.transform(val_intents).astype(np.int32)
intents_num = len(self.intents_label_encoder.classes_)
#train
if self.use_crf: # with crf
self.model = JointBertCRFModel(slots_num, intents_num, bert_model_hub_path, self.sess, num_bert_fine_tune_layers=10, is_bert=is_bert, is_crf=True, learning_rate=self.learning_rate)
self.model.fit([train_input_ids, train_input_mask, train_segment_ids, train_valid_positions, train_sequence_lengths], [train_tags, train_intents],
validation_data=([val_input_ids, val_input_mask, val_segment_ids, val_valid_positions, val_sequence_lengths], [val_tags, val_intents]),
epochs=self.epochs, batch_size=self.batch_size)
else: # without crf
self.model = JointBertModel(slots_num, intents_num, bert_model_hub_path, self.sess, num_bert_fine_tune_layers=10, is_bert=is_bert, is_crf=False, learning_rate=self.learning_rate)
self.model.fit([train_input_ids, train_input_mask, train_segment_ids, train_valid_positions], [train_tags, train_intents],
validation_data=([val_input_ids, val_input_mask, val_segment_ids, val_valid_positions], [val_tags, val_intents]),
epochs=self.epochs, batch_size=self.batch_size)
#save
if not os.path.exists(self.save_folder_path):
os.makedirs(self.save_folder_path)
self.model.save(self.save_folder_path)
with open(os.path.join(self.save_folder_path, 'tags_vectorizer.pkl'), 'wb') as handle:
pickle.dump(self.tags_vectorizer, handle, protocol=pickle.HIGHEST_PROTOCOL)
with open(os.path.join(self.save_folder_path, 'intents_label_encoder.pkl'), 'wb') as handle:
pickle.dump(self.intents_label_encoder, handle, protocol=pickle.HIGHEST_PROTOCOL)
#option 2: load existing model
def load(self):
load_folder_path = self.save_folder_path
with open(os.path.join(load_folder_path, 'params.json'), 'r') as json_file:
model_params = json.load(json_file)
slots_num = model_params['slots_num']
intents_num = model_params['intents_num']
bert_model_hub_path = model_params['bert_hub_path']
num_bert_fine_tune_layers = model_params['num_bert_fine_tune_layers']
is_bert = model_params['is_bert']
if 'is_crf' in model_params:
is_crf = model_params['is_crf']
else:
is_crf = False
self.bert_vectorizer = BERTVectorizer(self.sess, is_bert, bert_model_hub_path)
with open(os.path.join(load_folder_path, 'tags_vectorizer.pkl'), 'rb') as handle:
self.tags_vectorizer = pickle.load(handle)
slots_num = len(self.tags_vectorizer.label_encoder.classes_)
with open(os.path.join(load_folder_path, 'intents_label_encoder.pkl'), 'rb') as handle:
self.intents_label_encoder = pickle.load(handle)
intents_num = len(self.intents_label_encoder.classes_)
if is_crf:
self.model = JointBertCRFModel.load(load_folder_path, self.sess)
else:
self.model = JointBertModel.load(load_folder_path, self.sess)
#evaluate single input message
def predict(self,utterance):
with self.graph.as_default():
set_session(self.sess)
tokens = utterance.split()
input_ids, input_mask, segment_ids, valid_positions, data_sequence_lengths = self.bert_vectorizer.transform([utterance])
if self.use_crf:
predicted_tags, predicted_intents = self.model.predict_slots_intent(
[input_ids, input_mask, segment_ids, valid_positions, data_sequence_lengths],
self.tags_vectorizer, self.intents_label_encoder, remove_start_end=True,
include_intent_prob=True)
else:
predicted_tags, predicted_intents = self.model.predict_slots_intent(
[input_ids, input_mask, segment_ids, valid_positions],
self.tags_vectorizer, self.intents_label_encoder, remove_start_end=True,
include_intent_prob=True)
response = {
"intent": {
"name": predicted_intents[0][0].strip(),
"confidence": predicted_intents[0][1]
},
"slots": " ".join(predicted_tags[0])
}
return response
#evaluate test data set
def evaluate(self, test_data_dir):
from sklearn import metrics
with self.graph.as_default():
set_session(self.sess)
self.test_data_folder_path = test_data_dir
data_text_arr, data_tags_arr, data_intents = Reader.read(self.test_data_folder_path)
data_input_ids, data_input_mask, data_segment_ids, data_valid_positions, data_sequence_lengths = self.bert_vectorizer.transform(data_text_arr)
if self.use_crf:
predicted_tags, predicted_intents = self.model.predict_slots_intent(
[data_input_ids, data_input_mask, data_segment_ids, data_valid_positions, data_sequence_lengths],
self.tags_vectorizer, self.intents_label_encoder, remove_start_end=True, include_intent_prob=True)
else:
predicted_tags, predicted_intents = self.model.predict_slots_intent(
[data_input_ids, data_input_mask, data_segment_ids, data_valid_positions],
self.tags_vectorizer, self.intents_label_encoder, remove_start_end=True, include_intent_prob=True)
gold_tags = [x.split() for x in data_tags_arr]
#calculate metrics for intent and slots
slot_acc = metrics.accuracy_score(flatten(gold_tags), flatten(predicted_tags))
slot_f1 = metrics.f1_score(flatten(gold_tags), flatten(predicted_tags), average='weighted')
slot_precision = metrics.precision_score(flatten(gold_tags), flatten(predicted_tags), average='weighted')
slot_recall = metrics.recall_score(flatten(gold_tags), flatten(predicted_tags), average='weighted')
confidence = [ex[1] for ex in predicted_intents]
predicted_intents = [ex[0] for ex in predicted_intents]
intent_acc = metrics.accuracy_score(data_intents, predicted_intents)
intent_f1 = metrics.f1_score(data_intents, predicted_intents, average='weighted')
intent_precision = metrics.precision_score(data_intents, predicted_intents, average='weighted')
intent_recall = metrics.recall_score(data_intents, predicted_intents, average='weighted')
metrics = {
'slot_acc':slot_acc,
'slot_f1':slot_f1,
'slot_precision':slot_precision,
'slot_recall':slot_recall,
'intent_acc':intent_acc,
'intent_f1':intent_f1,
'intent_precision':intent_precision,
'intent_recall':intent_recall
}
predicted_tags = [ " ".join(ex)+"\n" for ex in predicted_tags]
# for report, confusion matrix & histogram
self.intent_true = data_intents
self.intent_pred = predicted_intents
self.slot_true = data_tags_arr
self.slot_pred = predicted_tags
self.confidence_score = [round(float(score),2) for score in confidence]
return [metrics, predicted_intents, data_intents, predicted_tags, data_tags_arr, confidence] #confidence score
# get individual intent and entity type report; only to be executed after evaluate_metrics()
def evaluation_get_individual_report(self):
from sklearn import metrics
intent_report = metrics.classification_report(
self.intent_true, self.intent_pred, output_dict=True
)
slot_report = metrics.classification_report(
self.slot_true, self.slot_pred, output_dict=True
)
return [intent_report, slot_report]
# creates confusion matrix after evaluation, with intent results
def compute_confusion_matrix(self,save_path):
from sklearn.metrics import confusion_matrix
from sklearn.utils.multiclass import unique_labels
import matplotlib.pyplot as plt
from rasa.nlu.test import plot_confusion_matrix
plt.gcf().clear()
# compute confusion matrix
cnf_matrix = confusion_matrix(self.intent_true, self.intent_pred)
# get list of unique labels from target and predicted
labels = unique_labels(self.intent_true, self.intent_pred)
plot_confusion_matrix(
cnf_matrix,
classes=labels,
title="Intent Confusion matrix",
out=save_path,
)
# create histogram of confidence distribution
def compute_histogram(self,save_path):
import matplotlib.pyplot as plt
from rasa.nlu.test import plot_histogram
plt.gcf().clear()
# hits histogram
pos_hist = [
score for true, pred, score in zip(self.intent_true, self.intent_pred, self.confidence_score) if true == pred
]
# miss histogram
neg_hist = [
score for true, pred, score in zip(self.intent_true, self.intent_pred, self.confidence_score) if true != pred
]
plot_histogram([pos_hist, neg_hist],save_path)
intents_label_encoder = LabelEncoder()
train_intents = intents_label_encoder.fit_transform(train_intents).astype(
np.int32)
val_intents = intents_label_encoder.transform(val_intents).astype(np.int32)
intents_num = len(intents_label_encoder.classes_)
model = JointBertModel(slots_num,
intents_num,
sess,
num_bert_fine_tune_layers=10)
h = model.fit([
train_input_ids, train_input_mask, train_segment_ids, train_valid_positions
], [train_tags, train_intents],
validation_data=([
val_input_ids, val_input_mask, val_segment_ids,
val_valid_positions
], [val_tags, val_intents]),
epochs=epochs,
batch_size=batch_size)
### saving
print('Saving ..')
if not os.path.exists(save_folder_path):
os.makedirs(save_folder_path)
print('Folder `%s` created' % save_folder_path)
model.save(save_folder_path)
with open(os.path.join(save_folder_path, 'tags_vectorizer.pkl'),
'wb') as handle:
pickle.dump(tags_vectorizer, handle, protocol=pickle.HIGHEST_PROTOCOL)
with open(os.path.join(save_folder_path, 'intents_label_encoder.pkl'),
Y_val = [X_val[:, 4 * split_width:5 * split_width], Y_val]
X_val = [
X_val[:, 0:split_width], X_val[:, split_width:2 * split_width],
X_val[:, 2 * split_width:3 * split_width],
X_val[:, 3 * split_width:4 * split_width]
]
X_train = (X_train[0], X_train[1], X_train[2],
model.prepare_valid_positions(X_train[3]))
X_val = (X_val[0], X_val[1], X_val[2],
model.prepare_valid_positions(X_val[3]))
hist = model.fit(X_train,
Y_train,
validation_data=(X_val, Y_val),
epochs=1,
batch_size=batch_size)
if history:
history = {
key: history[key] + hist.history[key]
for key in hist.history
}
else:
history = hist.history
plot_history(history)
plt.show()
plt.close()
print('Saving ..')