conversation_GAN.py

# -*- coding: utf-8 -*-
# PYTHON 3 - upgrades
__author__ = 'Oswaldo Ludwig'
__version__ = '1.02'


from tensorflow.python.keras.layers import Dense, Reshape, Flatten, \
    Dropout, Input, concatenate, Conv2D, MaxPooling2D, UpSampling2D, Conv2DTranspose, Activation

from keras.layers import Input, Embedding, LSTM, Dense, RepeatVector, Dropout, merge,concatenate
from keras.optimizers import Adam 
from keras.models import Model
from keras.models import Sequential
from keras.layers import Activation, Dense
from keras.preprocessing import sequence
from six.moves import input

import keras.backend as K
import numpy as np
np.random.seed(1234)  # for reproducibility
import pickle
import theano
import os.path
import sys
import nltk
import re
import time

from keras.utils import plot_model

WORD_EMBED_HIDDEN_SIZE = 100
SENTENCE_EMBEDDING_SIZE = 300
DICTIONARY_SIZE = 7000
MAX_SEQUENCE_LENGTH_INPUT = 50

vocabulary_file = 'vocabulary_movie'
weights_file = 'my_model_weights20.h5'
weights_file_GAN = 'my_model_weights.h5'
unknown_token = 'something'
file_saved_context = 'saved_context'
file_saved_answer = 'saved_answer'
name_of_computer = 'john'

def greedy_decoder(input):

    flag = 0
    prob = 1
    ans_partial = np.zeros((1,MAX_SEQUENCE_LENGTH_INPUT))
    ans_partial[0, -1] = 2  #  the index of the symbol BOS (begin of sentence)
    for k in range(MAX_SEQUENCE_LENGTH_INPUT - 1):
        ye = model.predict([input, ans_partial])
        yel = ye[0,:]
        p = np.max(yel)
        mp = np.argmax(ye)
        ans_partial[0, 0:-1] = ans_partial[0, 1:]
        ans_partial[0, -1] = mp
        if mp == 3:  #  he index of the symbol EOS (end of sentence)
            flag = 1
        if flag == 0:    
            prob = prob * p
    text = ''
    for k in ans_partial[0]:
        k = k.astype(int)
        if k < (DICTIONARY_SIZE-2):
            w = vocabulary[k]
            text = text + w[0] + ' '
    return(text, prob)
    
    
def preprocess(raw_word, name):
    
    l1 = ['won’t','won\'t','wouldn’t','wouldn\'t','’m', '’re', '’ve', '’ll', '’s','’d', 'n’t', '\'m', '\'re', '\'ve', '\'ll', '\'s', '\'d', 'can\'t', 'n\'t', 'B: ', 'A: ', ',', ';', '.', '?', '!', ':', '. ?', ',   .', '. ,', 'EOS', 'BOS', 'eos', 'bos']
    l2 = ['will not','will not','would not','would not',' am', ' are', ' have', ' will', ' is', ' had', ' not', ' am', ' are', ' have', ' will', ' is', ' had', 'can not', ' not', '', '', ' ,', ' ;', ' .', ' ?', ' !', ' :', '? ', '.', ',', '', '', '', '']
    l3 = ['-', '_', ' *', ' /', '* ', '/ ', '\"', ' \\"', '\\ ', '--', '...', '. . .']
    l4 = ['jeffrey','fred','benjamin','paula','walter','rachel','andy','helen','harrington','kathy','ronnie','carl','annie','cole','ike','milo','cole','rick','johnny','loretta','cornelius','claire','romeo','casey','johnson','rudy','stanzi','cosgrove','wolfi','kevin','paulie','cindy','paulie','enzo','mikey','i\97','davis','jeffrey','norman','johnson','dolores','tom','brian','bruce','john','laurie','stella','dignan','elaine','jack','christ','george','frank','mary','amon','david','tom','joe','paul','sam','charlie','bob','marry','walter','james','jimmy','michael','rose','jim','peter','nick','eddie','johnny','jake','ted','mike','billy','louis','ed','jerry','alex','charles','tommy','bobby','betty','sid','dave','jeffrey','jeff','marty','richard','otis','gale','fred','bill','jones','smith','mickey']    

    raw_word = raw_word.lower()
    raw_word = raw_word.replace(', ' + name_of_computer, '')
    raw_word = raw_word.replace(name_of_computer + ' ,', '')

    for j, term in enumerate(l1):
        raw_word = raw_word.replace(term,l2[j])
        
    for term in l3:
        raw_word = raw_word.replace(term,' ')
    
    for term in l4:
        raw_word = raw_word.replace(', ' + term, ', ' + name)
        raw_word = raw_word.replace(' ' + term + ' ,' ,' ' + name + ' ,')
        raw_word = raw_word.replace('i am ' + term, 'i am ' + name_of_computer)
        raw_word = raw_word.replace('my name is' + term, 'my name is ' + name_of_computer)
    
    for j in range(30):
        raw_word = raw_word.replace('. .', '')
        raw_word = raw_word.replace('.  .', '')
        raw_word = raw_word.replace('..', '')
       
    for j in range(5):
        raw_word = raw_word.replace('  ', ' ')
        
    if raw_word[-1] !=  '!' and raw_word[-1] != '?' and raw_word[-1] != '.' and raw_word[-2:] !=  '! ' and raw_word[-2:] != '? ' and raw_word[-2:] != '. ':
        raw_word = raw_word + ' .'
    
    if raw_word == ' !' or raw_word == ' ?' or raw_word == ' .' or raw_word == ' ! ' or raw_word == ' ? ' or raw_word == ' . ':
        raw_word = 'what ?'
    
    if raw_word == '  .' or raw_word == ' .' or raw_word == '  . ':
        raw_word = 'i do not want to talk about it .'
      
    return raw_word

def tokenize(sentences):

    # Tokenizing the sentences into words:
    tokenized_sentences = nltk.word_tokenize(sentences)
    index_to_word = [x[0] for x in vocabulary]
    word_to_index = dict([(w,i) for i,w in enumerate(index_to_word)])
    tokenized_sentences = [w if w in word_to_index else unknown_token for w in tokenized_sentences]
    X = np.asarray([word_to_index[w] for w in tokenized_sentences])
    s = X.size
    Q = np.zeros((1,MAX_SEQUENCE_LENGTH_INPUT))
    if s < (MAX_SEQUENCE_LENGTH_INPUT + 1):
        Q[0,- s:] = X
    else:
        Q[0,:] = X[- MAX_SEQUENCE_LENGTH_INPUT:]
    
    return Q

# Open files to save the conversation for further training:
qf = open(file_saved_context, 'w')
af = open(file_saved_answer, 'w')

print('Starting the model...')

# *******************************************************************
# Keras model of the chatbot: 
# *******************************************************************

ad = Adam(lr=0.00005) 

input_context = Input(shape=(MAX_SEQUENCE_LENGTH_INPUT,), dtype='int32', name='thecontexttext')
input_answer = Input(shape=(MAX_SEQUENCE_LENGTH_INPUT,), dtype='int32', name='theanswertextuptothecurrenttoken')
LSTM_encoder = LSTM(SENTENCE_EMBEDDING_SIZE, kernel_initializer= 'lecun_uniform', name='Encodecontext')
LSTM_decoder = LSTM(SENTENCE_EMBEDDING_SIZE, kernel_initializer= 'lecun_uniform', name='Encodeansweruptothecurrenttoken')

Shared_Embedding = Embedding(output_dim=WORD_EMBED_HIDDEN_SIZE, input_dim=DICTIONARY_SIZE, input_length=MAX_SEQUENCE_LENGTH_INPUT, name='Shared')
word_embedding_context = Shared_Embedding(input_context)
context_embedding = LSTM_encoder(word_embedding_context)

word_embedding_answer = Shared_Embedding(input_answer)
answer_embedding = LSTM_decoder(word_embedding_answer)



merge_layer = concatenate([context_embedding, answer_embedding], axis=1)
out  = Dense(int(DICTIONARY_SIZE/2), activation="relu", name='reluactivation')(merge_layer)
out = Dense(DICTIONARY_SIZE, activation="softmax")(out)

# out = (merge_layer)
# out = Dense(DICTIONARY_SIZE, activation="softmax", name='likelihoodofthecurrenttokenusingsoftmaxactivation')(out)

model = Model(inputs=[input_context, input_answer], outputs = [out])

model.compile(loss='categorical_crossentropy', optimizer=ad) 

# Loading the data:
vocabulary = pickle.load(open(vocabulary_file, 'rb'))

print("\n \n \n \n    CHAT:     \n \n")

# Processing the user query:
prob = 0
que = ''
last_query  = ' '
last_last_query = ''
text = ' '
last_text = ''
print('computer: hi ! please type your name.\n')
name = input("user: ")
print('computer: hi , ' + name +' ! My name is ' + name_of_computer + '.\n') 


while que != 'exit .':
    
    que = input("user: ")
    que = preprocess(que, name_of_computer)
    # Collecting data for training:
    q = last_query + ' ' + text
    a = que
    qf.write(q + '\n')
    af.write(a + '\n')
    # Composing the context:
    if prob > 0.2:
        query = text + ' ' + que
    else:    
        query = que
   
    last_text = text
    
    Q = tokenize(query)
    
    # Using the trained model to predict the answer:
    model.load_weights(weights_file)
    predout, prob = greedy_decoder(Q[0:1])
    start_index = predout.find('EOS')
    text = preprocess(predout[0:start_index], name)
    print(('computer_BP: ' + text + '    (with probability of %f)'%prob))
    
    model.load_weights(weights_file_GAN)
    predout, prob = greedy_decoder(Q[0:1])
    start_index = predout.find('EOS')
    text = preprocess(predout[0:start_index], name)
    print(('computer_GAN: ' + text + '    (with probability of %f)'%prob))
    
    last_last_query = last_query    
    last_query = que

qf.close()
af.close()