train_model.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Mon Jul  2 18:30:43 2018

@author: himanshu
"""


from keras.layers import Input, Embedding, LSTM, Dense, RepeatVector, Bidirectional, Dropout, merge
from keras.optimizers import Adam, SGD
from keras.models import Model
from keras.models import Sequential
from keras.layers import Activation, Dense
from keras.callbacks import EarlyStopping
from keras.preprocessing import sequence

import os
import keras
from importlib import reload
import tensorflow as tf
print(tf.test.gpu_device_name())
import keras.backend as K

import numpy as np
np.random.seed(2018)  # for reproducibility
# import cPickle
import _pickle as cPickle
import os
import pandas as pd
import sys
import matplotlib.pyplot as plt

word_embedding_size = 100
sentence_embedding_size = 300
# dictionary_size = T.dscalar()
dictionary_size = 7000
maxlen_input = 50
maxlen_output = 50
num_subsets = 1
Epochs = 50
BatchSize = 2500 #  Check the capacity of your GPU
Patience = 0
dropout = .25
n_test = 100

vocabulary_file = '../input/vocabulary/vocabulary_movie'
questions_file = '../input/talk-data-2009/Padded_context_2009'
answers_file = '../input/talk-data-2009/Padded_answers_2009'
weights_file = 'my_model_weights20.h5'
GLOVE_DIR = '../input/glove6b100dtxt'

early_stopping = EarlyStopping(monitor='val_loss', patience=Patience)


def print_result(input):

    ans_partial = np.zeros((1,maxlen_input))
    ans_partial[0, -1] = 2  #  the index of the symbol BOS (begin of sentence)
    for k in range(maxlen_input - 1):
        ye = model.predict([input, ans_partial])
        mp = np.argmax(ye)
        ans_partial[0, 0:-1] = ans_partial[0, 1:]
        ans_partial[0, -1] = mp
    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)


# **********************************************************************
# Reading a pre-trained word embedding and addapting to our vocabulary:
# **********************************************************************

embeddings_index = {}
f = open(os.path.join(GLOVE_DIR, 'glove.6B.100d.txt'))
for line in f:
    values = line.split()
    word = values[0]
    coefs = np.asarray(values[1:], dtype='float32')
    embeddings_index[word] = coefs
f.close()

print('Found %s word vectors.' % len(embeddings_index))
embedding_matrix = np.zeros((dictionary_size, word_embedding_size))

# Loading our vocabulary:
vocabulary = cPickle.load(open(vocabulary_file, 'rb'))

# Using the Glove embedding:
i = 0
for word in vocabulary:
    embedding_vector = embeddings_index.get(word[0])
    
    if embedding_vector is not None:
        # words not found in embedding index will be all-zeros.
        embedding_matrix[i] = embedding_vector
    i += 1

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

ad = Adam(lr=0.00005) 

input_context = Input(shape=(maxlen_input,), dtype='int32', name='input_context')
input_answer = Input(shape=(maxlen_input,), dtype='int32', name='input_answer')
LSTM_encoder = LSTM(sentence_embedding_size, init= 'lecun_uniform')
LSTM_decoder = LSTM(sentence_embedding_size, init= 'lecun_uniform')
if os.path.isfile(weights_file):
    Shared_Embedding = Embedding(output_dim=word_embedding_size, input_dim=dictionary_size, input_length=maxlen_input)
else:
    Shared_Embedding = Embedding(output_dim=word_embedding_size, input_dim=dictionary_size, weights=[embedding_matrix], input_length=maxlen_input)
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 = merge([context_embedding, answer_embedding], mode='concat', concat_axis=1)
out = Dense(3500, activation="relu")(merge_layer)
out = Dense(7000, activation="softmax")(out)

model = Model(input=[input_context, input_answer], output = [out])

model.compile(loss='categorical_crossentropy', optimizer=ad)
print('model compiled')
if os.path.isfile(weights_file):
    model.load_weights(weights_file)

# ************************************************************************
# Loading the data:
# ************************************************************************

q = cPickle.load(open(questions_file, 'rb'))
a = cPickle.load(open(answers_file, 'rb'))
print('pickle load completed')
n_exem, n_words = a.shape

qt = q[0:n_test,:]
at = a[0:n_test,:]
q = q[n_test + 1:,:]
a = a[n_test + 1:,:]

print('Number of exemples = %d'%(n_exem - n_test))
step = np.around((n_exem - n_test)/num_subsets)
round_exem = step * num_subsets
step = int(step)
# *************************************************************************
# Bot training:
# *************************************************************************

x = range(0,Epochs) 
valid_loss = np.zeros(Epochs)
train_loss = np.zeros(Epochs)
for m in range(Epochs):
    
    # Loop over training batches due to memory constraints:
    for n in range(0,int(round_exem),int(step)):
        
        q2 = q[n:n+step]
        s = q2.shape
        count = 0
        for i, sent in enumerate(a[n:n+step]):
            l = np.where(sent==3)  #  the position od the symbol EOS
            limit = l[0][0]
            count += limit + 1
        print('count is',count)    
        Q = np.zeros((count,maxlen_input))
        A = np.zeros((count,maxlen_input))
        from psutil import virtual_memory
        mem = virtual_memory()
        print(mem)
        Y = np.zeros((count,dictionary_size),dtype = 'uint8')
        
        # Loop over the training examples:
        count = 0
        for i, sent in enumerate(a[n:n+step]):
            ans_partial = np.zeros((1,maxlen_input))
            
            # Loop over the positions of the current target output (the current output sequence):
            l = np.where(sent==3)  #  the position of the symbol EOS
            limit = l[0][0]

            for k in range(1,limit+1):
                # Mapping the target output (the next output word) for one-hot codding:
                y = np.zeros((1, dictionary_size))
                y[0, sent[k]] = 1

                # preparing the partial answer to input:

                ans_partial[0,-k:] = sent[0:k]

                # training the model for one epoch using teacher forcing:
                
                Q[count, :] = q2[i:i+1] 
                A[count, :] = ans_partial 
                Y[count, :] = y
                count += 1
                
        print('Training epoch: %d, training examples: %d - %d'%(m,n, n + step))
        model.fit([Q, A], Y, batch_size=BatchSize, epochs=1)
         
        test_input = qt[41:42]
        print(print_result(test_input))
        train_input = q[41:42]
        print(print_result(train_input))        
        
model.save_weights(weights_file, overwrite=True)