-
Notifications
You must be signed in to change notification settings - Fork 0
/
chat-bot.py
428 lines (328 loc) · 12.1 KB
/
chat-bot.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
'''
Panagiotis Christakakis
'''
# Import Libraries
# For parsing
import numpy as np
import pandas as pd
import random
import glob
import json
import re
import os
# Preprocessing step Model building
from keras.layers import LSTM,Dense, Dropout, Embedding, CuDNNLSTM, Bidirectional, Embedding, Input, TimeDistributed
from tensorflow.keras.layers import Dense, Embedding, LSTM, Input, Bidirectional, Concatenate, Dropout, Attention
from tensorflow.keras.preprocessing.sequence import pad_sequences
from keras.callbacks import ModelCheckpoint, EarlyStopping
from keras.models import Sequential, Model, load_model
from tensorflow.keras.preprocessing.sequence import pad_sequences
from sklearn.model_selection import train_test_split
from tensorflow.keras.utils import to_categorical
from keras.preprocessing.text import Tokenizer
from tensorflow.keras.models import Model
from sklearn.utils import shuffle
import tensorflow as tf
# Plotting
import matplotlib.pyplot as plt
# %matplotlib inline
# # For G-Drive attach
# from google.colab import drive
# drive.mount('/content/drive')
# Functions
def clean_text(txt):
# Input: text
# This function converts it's text input to lowercase letters
# and then many contraction words into their corresponding ones.
txt = txt.lower()
txt = re.sub(r"where's", "where is", txt)
txt = re.sub(r"that's", "that is", txt)
txt = re.sub(r"what's", "what is", txt)
txt = re.sub(r"won't", "will not", txt)
txt = re.sub(r"can't", "can not", txt)
txt = re.sub(r"she's", "she is", txt)
txt = re.sub(r"\'ll", " will", txt)
txt = re.sub(r"\'ve", " have", txt)
txt = re.sub(r"he's", "he is", txt)
txt = re.sub(r"\'d", " would", txt)
txt = re.sub(r"\'re", " are", txt)
txt = re.sub(r"[^\w\s]", "", txt)
txt = re.sub(r"i'm", "i am", txt)
return txt
def plot_metric(history, metric):
# Input: history of model, metrics
# This function can plot two different graphs.
# Training and Validation loss and accuracy
train_metrics = history.history[metric]
val_metrics = history.history['val_'+metric]
epochs = range(1, len(train_metrics) + 1)
plt.plot(epochs, train_metrics)
plt.plot(epochs, val_metrics)
plt.title('Training and Validation '+ metric)
plt.xlabel("Epochs")
plt.ylabel(metric)
plt.legend(["train_"+metric, 'val_'+metric])
plt.show()
# Folder specification
# %Link_To_Your_Path%
path_to_files = 'C:/Users/panos/Desktop/metalwoz-v1/metalwoz_data/'
os.chdir(path_to_files)
# Obtain all .txt path files from folder
data_folder_path = '/dialogues/*.txt'
# With glob module we obtain all the pathnames matching a specified pattern
files_list = glob.glob(path_to_files + data_folder_path)
print('A total of', len(files_list), 'files loaded.\n')
# Parsing step
list_of_dicts = []
# Loop for each file and insert in a dictionary
for filename in files_list:
with open(filename) as f:
for line in f:
list_of_dicts.append(json.loads(line))
# Create a new dict containing only useful data
new_list_of_dicts = []
# Create key-value pairs for useful data of conversation only.
for old_dict in list_of_dicts:
# k == 'turns' because from there it starts the conversation.
foodict = {k: v for k, v in old_dict.items() if (k == 'turns')}
new_list_of_dicts.append(foodict)
# Delete and replace to free memory
del(list_of_dicts)
list_of_dicts = new_list_of_dicts
questions = []
answers = []
matrix_greetings = ["Hey", "Hi", " "]
matrix_byes = ["Ok", " ", "Bye"]
# For each dictionary in the list
for dictionary in list_of_dicts:
matrix_QA = dictionary['turns']
# In order to split the QAs to 2 matrices (questions & answers),
# we will use a flag to indicate if the sentence is given from the bot or from the user
#bot_flag = True # Init
# For each Q/A in the matrix
# Remove "hey how can i help you"
matrix_QA.pop(0)
bot_flag = False
for sentence in matrix_QA:
if bot_flag == True:
# Used for bot's answers
answers.append(sentence)
# Switch
bot_flag = False
continue
else:
# Used for user's questions
questions.append(sentence)
# Switch
bot_flag = True
# The last loop (ideally) ends with a bot's answer, thus making bot_flag equal to False.
# Although, with data visualization and exploring, we can see that this does not happen all the time.
# Corner case: If the last answers was from the user,
# then we need to add one artificial 'ghost' response from the bot to make the dataset even.
if bot_flag == True:
answers.append(random.choice(matrix_byes))
# If list of questions and answers aren't the same return an error.
assert len(questions) == len(answers), "ERROR: The length of the questions and answer matrices are different."
print('A total of', len(questions), 'Questions-Answers were loaded.\n')
# We will shuffle them to ensure that our bot isn't overfitting on
# limited goal-oriented dialogs like setting an alarm or a exlplaining a catalogue
# Last, but not least, this way will enrich the vocabulary of our bot.
questions, answers = shuffle(np.array(questions), np.array(answers))
sorted_ques = []
sorted_ans = []
# We'll try to keep a lower number of dialogs, based on the lenght of
# the question. We do this in order to avoid RAM errors.
# So questions with lenght of 13 words and lower will be filtered and
# create new lists for QAs.
# Define the of length questions you want to keep.
q_len = 13
for i in range(len(questions)):
if len(questions[i]) < q_len:
sorted_ques.append(questions[i])
sorted_ans.append(answers[i])
clean_ques = []
clean_ans = []
# Apply the clean_text function we defined earlier.
for line in sorted_ques:
clean_ques.append(clean_text(line))
for line in sorted_ans:
clean_ans.append(clean_text(line))
for i in range(len(clean_ans)):
clean_ans[i] = ' '.join(clean_ans[i].split()[:11])
# Delete lists to free memory
del(sorted_ans, sorted_ques, answers, questions)
# # Keep a subset of the QA lists. Only used when training
# # on Google Colab because of RAM error messages.
# NUM_DIALOGS = 30000
# clean_ques = clean_ques[:NUM_DIALOGS]
# clean_ans = clean_ans[:NUM_DIALOGS]
# Count occurences
word2count = {}
for line in clean_ques:
for word in line.split():
if word not in word2count:
word2count[word] = 1
else:
word2count[word] += 1
for line in clean_ans:
for word in line.split():
if word not in word2count:
word2count[word] = 1
else:
word2count[word] += 1
## # Delete to free memory
del(word, line)
# Remove the least frequent key-value pairs from the vocabulary.
# Set a threshold of minimum value.
thresh = 5
# Create the new vocabulary.
vocab = {}
word_num = 0
for word, count in word2count.items():
if count >= thresh:
vocab[word] = word_num
word_num += 1
# Delete to free memory
del(word2count, word, count, thresh)
del(word_num)
# Import <Start Of Sentence> and <End Of Sentence> tokens
# in front and after each answer.
for i in range(len(clean_ans)):
clean_ans[i] = '<SOS> ' + clean_ans[i] + ' <EOS>'
tokens = ['<PAD>', '<EOS>', '<OUT>', '<SOS>']
x = len(vocab)
for token in tokens:
vocab[token] = x
x += 1
vocab['us'] = vocab['<PAD>']
vocab['<PAD>'] = 0
# Delete to free memory.
del(token, tokens)
del(x)
# Invert dictionary of answers
inv_vocab = {w:v for v, w in vocab.items()}
# Delete to free memory.
del(i)
# Starting to build the encoder-decoder
# by defining and creating its inputs.
encoder_inp = []
for line in clean_ques:
lst = []
for word in line.split():
if word not in vocab:
lst.append(vocab['<OUT>'])
else:
lst.append(vocab[word])
encoder_inp.append(lst)
decoder_inp = []
for line in clean_ans:
lst = []
for word in line.split():
if word not in vocab:
lst.append(vocab['<OUT>'])
else:
lst.append(vocab[word])
decoder_inp.append(lst)
# Delete to free memory.
del(clean_ans, clean_ques, line, lst, word)
# Ensure that all sequences in a list have
# the same lenght with the help of padding.
encoder_inp = pad_sequences(encoder_inp, q_len, padding = 'post', truncating = 'post')
decoder_inp = pad_sequences(decoder_inp, q_len, padding = 'post', truncating = 'post')
decoder_final_output = []
for i in decoder_inp:
decoder_final_output.append(i[1:])
decoder_final_output = pad_sequences(decoder_final_output, q_len, padding = 'post', truncating='post')
# Delete to free memory.
del(i)
decoder_final_output = to_categorical(decoder_final_output, len(vocab))
# Model creation
enc_inp = Input(shape=(q_len, ))
dec_inp = Input(shape=(q_len, ))
VOCAB_SIZE = len(vocab)
embed = Embedding(VOCAB_SIZE + 1, output_dim = 20, input_length=q_len, trainable=True)
enc_embed = embed(enc_inp)
enc_lstm = LSTM(400, return_state = True, dropout = 0.30, return_sequences = True)
enc_op, h, c = enc_lstm(enc_embed)
enc_states = [h, c]
dec_embed = embed(dec_inp)
dec_lstm = LSTM(400, return_state = True, dropout = 0.30, return_sequences = True)
dec_op, _, _ = dec_lstm(dec_embed, initial_state = enc_states)
dense = Dense(VOCAB_SIZE, activation = 'softmax')
dense_op = dense(dec_op)
# Input and output of our model.
model = Model([enc_inp, dec_inp], dense_op)
# Print a summary of the created model.
model.summary()
# Compile the model.
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate = 0.001),
loss='categorical_crossentropy', metrics=['acc'])
# Train the model
history = model.fit([encoder_inp, decoder_inp], decoder_final_output,
epochs = 50, batch_size = 128, validation_split = 0.15)
# Save the models and it's weights.
model.save('chatbot.h5')
model.save_weights('chatbot_weights.h5')
# Plot the metrics function we defined earlier.
plot_metric(history, 'loss')
plot_metric(history, 'acc')
enc_model = Model([enc_inp], enc_states)
decoder_state_input_h = Input(shape = (400,))
decoder_state_input_c = Input(shape = (400,))
decoder_states_inputs = [decoder_state_input_h, decoder_state_input_c]
decoder_outputs, state_h, state_c = dec_lstm(dec_embed,
initial_state=decoder_states_inputs)
decoder_states = [state_h, state_c]
dec_model = tf.keras.models.Model([dec_inp, decoder_states_inputs],
[decoder_outputs] + decoder_states)
print("**Starting the chat**")
prepro1 = ""
while prepro1 != 'q':
prepro1 = input("User : ")
## prepro1 = "Hey"
prepro1 = clean_text(prepro1)
## prepro1 = "hey"
prepro = [prepro1]
## prepro1 = ["hey"]
txt = []
for x in prepro:
# x = "hey"
lst = []
for y in x.split():
## y = "hey"
try:
lst.append(vocab[y])
## vocab['hey'] = (i.e: 454)
except:
lst.append(vocab['<OUT>'])
txt.append(lst)
## txt = [[454]]
txt = pad_sequences(txt, q_len, padding='post')
## txt = [[454,0,0,0,.........q_len]]
stat = enc_model.predict( txt )
empty_target_seq = np.zeros( ( 1 , 1) )
## empty_target_seq = [0]
empty_target_seq[0, 0] = vocab['<SOS>']
## empty_target_seq = [255]
stop_condition = False
decoded_translation = ''
while not stop_condition :
dec_outputs , h, c= dec_model.predict([ empty_target_seq] + stat )
decoder_concat_input = dense(dec_outputs)
## decoder_concat_input = [0.1, 0.2, .4, .0, ...............]
sampled_word_index = np.argmax( decoder_concat_input[0, -1, :] )
## sampled_word_index = [2]
sampled_word = inv_vocab[sampled_word_index] + ' '
## inv_vocab[2] = 'hi'
## sampled_word = 'hi '
if sampled_word != '<EOS> ':
decoded_translation += sampled_word
if sampled_word == '<EOS> ' or len(decoded_translation.split()) > q_len:
stop_condition = True
empty_target_seq = np.zeros( ( 1 , 1 ) )
empty_target_seq[ 0 , 0 ] = sampled_word_index
## <SOS> - > hi
## hi --> <EOS>
stat = [h, c]
print("Chatbot: ", decoded_translation )
print("==============================================")