class Conversation():
"""
This model will learn from user input to bot output.
Given the probability of the user's statement being of a certain
type, the result will be a random item in the response list.
"""
def __init__(self, intents=None, testing=False):
self.td = TopicProcessor()
self.cp = ConversationProcessor()
self.convX, self.convy = self.cp.get_dataset()
self.topX, self.topy = self.td.get_dataset()
self.topy = np_utils.to_categorical(self.topy)
self.top_classes = self.td.get_num_categories()
self.top_sentence_length = self.td.get_sentence_length()
self.graph = None
self.seq2seqmodel = seq2seqmodel()
if not testing:
# Try to load an existing model from file
try:
clear_session()
self.model, self.graph, self.sess = self._load_model()
except (FileNotFoundError, OSError):
# If the model doesn't exist, we'll have to build one
model = self._build_model()
self.model = self._train_model(model, self.convX, self.convy,
self.topX, self.topy)
self.model, self.graph, self.sess = self._load_model()
# If the user hasn't supplied a location, use the default
if not intents:
intents = 'chatbot/corpus/output.json'
with open(intents, 'r') as json_file:
self.intents = json.load(json_file)
else:
# self._test_model(self.convX, self.convy, self.topX, self.topy)
self._k_fold_test(self.convX, self.convy, self.topX, self.topy)
def _load_model(self):
"""
Load the model from the disk. This is the weights in an h5 file,
and the structure in a JSON format.
returns: keras model
"""
directory = __file__.split('\\')
directory = '/'.join(directory[:-1])
#load_model = models.load_model(os.path.join(directory, 'model.h5'))
with open(
os.path.join(directory, 'trainedconversationalmodel-old.json'),
'r') as open_file:
load_model = open_file.read()
load_model = model_from_json(load_model)
load_model.load_weights(
os.path.join(directory,
'trainedconversationalmodelweights-old.h5'))
load_model._make_predict_function()
self.graph = tf.get_default_graph()
# self.graph.finalize()
sess = K.get_session()
print("Combined model was loaded from disk")
return load_model, self.graph, sess
def _save_model(self, model):
"""
Save a trained model to the disk. The weights of the model is saved in
a H5 file. The models structure is in a JSON file format.
Args:
model: The trained model you want to save to disk.
"""
directory = __file__.split('\\')
directory = '/'.join(directory[:-1])
model.save(os.path.join(directory, 'model.h5'))
json_model = model.to_json()
with open(directory + '/trainedconversationalmodel.json',
'w') as open_file:
open_file.write(json_model)
model.save_weights(directory + '/trainedconversationalmodelweights.h5')
def _build_model(self):
"""
Build keras model and train it against the dataset.
Args:
X: input statements
y: label of the input statement
Returns:
A trained keras model
"""
conv_input = Input(shape=(self.cp.MAX_SEQUENCE_LENGTH, ),
name='conv_input')
top_input = Input(shape=(self.top_sentence_length, ), name='top_input')
inputs = concatenate([conv_input, top_input])
embedding = Embedding(input_dim=1000, output_dim=40)(inputs)
flatten = Flatten()(embedding)
conv_layer_1 = Dense(units=600, activation='sigmoid')(flatten)
conv_dropout_1 = Dropout(0.7)(conv_layer_1)
conv_layer_2 = Dense(units=400, activation='sigmoid')(conv_dropout_1)
conv_dropout_2 = Dropout(0.7)(conv_layer_2)
conv_output = Dense(units=self.cp.n_classes,
activation='softmax',
name='conversation_output')(conv_dropout_2)
topic_layer_1 = Dense(units=2000, activation='sigmoid',
init='normal')(flatten)
topic_dropout_1 = Dropout(0.5)(topic_layer_1)
topic_layer_2 = Dense(units=500, activation='sigmoid',
init='uniform')(topic_dropout_1)
topic_dropout_2 = Dropout(0.5)(topic_layer_2)
topic_output = Dense(units=self.top_classes,
activation='softmax',
name='topic_output')(topic_dropout_2)
topic_optimizer = optimizers.RMSprop(lr=0.0001)
model = Model(inputs=[conv_input, top_input],
outputs=[conv_output, topic_output])
model.compile(optimizer=topic_optimizer,
loss='categorical_crossentropy',
metrics=['accuracy'])
return model
def _k_fold_test(self, statements, labels, sentences, topics):
"""
Perform Kfold cross validation, this is a slightly different take on the _test_model function, as
the X and Y datasets will be split into k partitions and then tests k times. This means that it will
take considerably longer to get the final result, depending on the number of folds used.
args:
statements: the intent statements (X)
labels: the intent classes (Y)
sentences: the topic statements (X)
topics: the topic classes (Y)
"""
cv_scores_conv = []
cv_scores_topic = []
kfold = KFold(
labels.shape[0], n_folds=10, shuffle=True, random_state=0
) # depreciated in sklearn, but works well with multi-label classification
for train, test in kfold:
model = self._build_model()
model.fit([statements[train], sentences[train]],
[labels[train], topics[train]],
batch_size=742,
epochs=1300)
score = model.evaluate([statements[test], sentences[test]],
[labels[test], topics[test]])
print(score[3] * 100, score[4] * 100)
cv_scores_conv.append(score[3] * 100)
cv_scores_topic.append(score[4] * 100)
print("Intent Accuracy: %.2f%% (+/- %.2f%%)" %
(np.mean(cv_scores_conv), np.std(cv_scores_conv)))
print("Topic Accuracy: %.2f%% (+/- %.2f%%)" %
(np.mean(cv_scores_topic), np.std(cv_scores_topic)))
def _test_model(self, statements, labels, sentences, topics):
"""
Perform a simple testing method by splitting the corpus into training and testing sets. This method
will also output graphs for how the model performs during the training and testing methods.
I have disabled it it on MacOS as matplotlib doesn't work in virtualenvwrapper environments.
args:
statements: the intent statements (X)
labels: the intent classes (Y)
sentences: the topic statements (X)
topics: the topic classes (Y)
"""
X_conv_train, X_conv_test, y_conv_train, y_conv_test = train_test_split(
statements, labels, test_size=0.5, random_state=0)
X_topic_train, X_topic_test, y_topic_train, y_topic_test = train_test_split(
sentences, topics, test_size=0.5, random_state=0)
images_folder_loc = 'images/'
model = self._build_model()
training = model.fit([X_conv_train, X_topic_train],
[y_conv_train, y_topic_train],
validation_data=([X_conv_test, X_topic_test],
[y_conv_test, y_topic_test]),
batch_size=X_conv_train.shape[0],
epochs=1500)
print(training.history.keys())
if platform != 'darwin':
plt.plot(training.history['conversation_output_acc'], color='red')
plt.plot(training.history['val_conversation_output_acc'],
color='green')
plt.title('Conversation Accuracy')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.legend(['train', 'test'], loc='upper left')
plt.savefig(str(images_folder_loc + 'conv_acc.png'))
plt.clf()
plt.plot(training.history['topic_output_acc'], color='red')
plt.plot(training.history['val_topic_output_acc'], color='green')
plt.title('Topic Accuracy')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.legend(['Training', 'Testing'], loc='upper left')
plt.savefig(str(images_folder_loc + 'topic_acc.png'))
plt.clf()
plt.plot(training.history['conversation_output_loss'], color='red')
plt.plot(training.history['val_conversation_output_loss'],
color='green')
plt.title('Converation Loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend(['Training', 'Testing'], loc='upper left')
plt.savefig(str(images_folder_loc + 'conv_loss.png'))
plt.clf()
plt.plot(training.history['topic_output_loss'], color='red')
plt.plot(training.history['val_topic_output_loss'], color='green')
plt.title('Topic Loss')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.legend(['Training', 'Testing'], loc='upper left')
plt.savefig(str(images_folder_loc + 'topic_loss.png'))
plt.clf()
plot_model(model, to_file=str(images_folder_loc + 'model.png'))
print(model.metrics_names)
print(
model.evaluate([X_conv_test, X_topic_test],
[y_conv_test, y_topic_test]))
def _train_model(self, model, statements, labels, sentences, topics):
"""
Train method for the pre-compiled model. Returns a trained model
args:
model: pre-compiled keras model to train
statements: the intent statements (X)
labels: the intent classes (Y)
sentences: the topic statements (X)
topics: the topic classes (Y)
returns:
model: the trained keras model
"""
model.fit([statements, sentences], [labels, topics],
batch_size=statements.shape[0],
epochs=1)
self._save_model(model)
return model
def encode(self, statement):
"""
When the user inputs a statement into the chatbot. It will need to be converted
to a sequence vector for the model to work with. This function is an abstraction of that
process.
Args:
statement: statement that is to be converted to a vector.
Returns:
A sequence vector representing the user's input.
"""
return self.cp.encode_statment(statement)
def request(self, req, context=None, state=None):
"""
This is the function that will be called by the chatbot, when the user inputs a statement.
It will call the encode function of this class, use the model to predict the
probability that it is of all classes using a softmax function. The highest probability
is therefore deemed the correct one (with some error threshold checking).
Args:
req: statement that we need to predict the class for.
Returns:
(response, context_set). Response is the String that the chatbot should respond with.
Context_set is the tag of the intent that the bot needs context for.
"""
with self.sess.as_default():
with self.graph.as_default():
pred = self.model.predict(
[self.encode(req),
self.td.encode_statement(req)])
statement_pred = pred[0]
topic_pred = pred[1]
topic_class = np.argmax(topic_pred)
topic_confidence = topic_pred[0][topic_class]
statement_class = np.argmax(statement_pred)
statement_confidence = statement_pred[0][statement_class]
pred_class = self.cp.get_class(statement_class)
topic_pred_class = self.td.decode_topic(topic_class)
context_set = None
response = ''
for intent in self.intents['intents']:
if context:
if intent['tag'] == state:
for index, value in enumerate(intent['context_filter']):
if value == context:
if 'context_set' in intent:
try:
context_set = intent['context_set'][index]
except:
context_set = None
response = intent['responses'][index]
elif 'context_filter' in intent and intent['tag'] == pred_class:
for index, value in enumerate(intent['context_filter']):
if value == context:
if 'context_set' in intent:
try:
context_set = intent['context_set'][index]
except:
context_set = None
response = intent['responses'][index]
elif intent['tag'] == pred_class and 'context_filter' not in intent:
if 'context_set' in intent:
context_set = intent['context_set'][0]
response = random.choice(intent['responses'])
if not response: # catch all
# response = 'I am not sure I understand, could you explain it to me further?'
response = self.seq2seqmodel.decode_sequence(req)
return response, context_set, pred_class, topic_pred_class, topic_confidence
def test_decode_topic(self):
tp = TopicProcessor()
original_topic = 15
decoded_topic = tp.decode_topic(original_topic)
self.assertNotEqual(original_topic, decoded_topic)
self.assertEqual(decoded_topic, 'abuse')