{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"\n",
"# RNN using LSTM \n",
" \n",
"\n",
"\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"<img src=\"../imgs/RNN-rolled.png\"/ width=\"80px\" height=\"80px\">"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"<img src=\"../imgs/RNN-unrolled.png\"/ width=\"400px\" height=\"400px\">"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"<img src=\"../imgs/LSTM3-chain.png\"/ width=\"60%\">"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"_source: http://colah.github.io/posts/2015-08-Understanding-LSTMs_"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from keras.optimizers import SGD\n",
"from keras.preprocessing.text import one_hot, text_to_word_sequence\n",
"from keras.utils import np_utils\n",
"from keras.models import Sequential\n",
"from keras.layers.core import Dense, Dropout, Activation\n",
"from keras.layers.embeddings import Embedding\n",
"from keras.layers.recurrent import LSTM, GRU\n",
"from keras.preprocessing import sequence"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Reading blog post from data directory"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"import os\n",
"import pickle\n",
"import numpy as np"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"DATA_DIRECTORY = os.path.join(os.path.abspath(os.path.curdir), '..', 'data', 'word_embeddings')\n",
"print(DATA_DIRECTORY)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"male_posts = []\n",
"female_post = []"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"with open(os.path.join(DATA_DIRECTORY,\"male_blog_list.txt\"),\"rb\") as male_file:\n",
" male_posts= pickle.load(male_file)\n",
" \n",
"with open(os.path.join(DATA_DIRECTORY,\"female_blog_list.txt\"),\"rb\") as female_file:\n",
" female_posts = pickle.load(female_file)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"filtered_male_posts = list(filter(lambda p: len(p) > 0, male_posts))\n",
"filtered_female_posts = list(filter(lambda p: len(p) > 0, female_posts))"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"# text processing - one hot builds index of the words\n",
"male_one_hot = []\n",
"female_one_hot = []\n",
"n = 30000\n",
"for post in filtered_male_posts:\n",
" try:\n",
" male_one_hot.append(one_hot(post, n, split=\" \", lower=True))\n",
" except:\n",
" continue\n",
"\n",
"for post in filtered_female_posts:\n",
" try:\n",
" female_one_hot.append(one_hot(post,n,split=\" \", lower=True))\n",
" except:\n",
" continue"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"# 0 for male, 1 for female\n",
"concatenate_array_rnn = np.concatenate((np.zeros(len(male_one_hot)),\n",
" np.ones(len(female_one_hot))))"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"from sklearn.model_selection import train_test_split\n",
"\n",
"X_train_rnn, X_test_rnn, y_train_rnn, y_test_rnn = train_test_split(np.concatenate((female_one_hot,male_one_hot)),\n",
" concatenate_array_rnn, \n",
" test_size=0.2)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"maxlen = 100\n",
"X_train_rnn = sequence.pad_sequences(X_train_rnn, maxlen=maxlen)\n",
"X_test_rnn = sequence.pad_sequences(X_test_rnn, maxlen=maxlen)\n",
"print('X_train_rnn shape:', X_train_rnn.shape, y_train_rnn.shape)\n",
"print('X_test_rnn shape:', X_test_rnn.shape, y_test_rnn.shape)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"max_features = 30000\n",
"dimension = 128\n",
"output_dimension = 128\n",
"model = Sequential()\n",
"model.add(Embedding(max_features, dimension))\n",
"model.add(LSTM(output_dimension))\n",
"model.add(Dropout(0.5))\n",
"model.add(Dense(1))\n",
"model.add(Activation('sigmoid'))"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"model.compile(loss='mean_squared_error', optimizer='sgd', metrics=['accuracy'])"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"model.fit(X_train_rnn, y_train_rnn, batch_size=32,\n",
" epochs=4, validation_data=(X_test_rnn, y_test_rnn))"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"score, acc = model.evaluate(X_test_rnn, y_test_rnn, batch_size=32)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"print(score, acc)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Using TFIDF Vectorizer as an input instead of one hot encoder"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"from sklearn.feature_extraction.text import TfidfVectorizer"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"vectorizer = TfidfVectorizer(decode_error='ignore', norm='l2', min_df=5)\n",
"tfidf_male = vectorizer.fit_transform(filtered_male_posts)\n",
"tfidf_female = vectorizer.fit_transform(filtered_female_posts)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"flattened_array_tfidf_male = tfidf_male.toarray()\n",
"flattened_array_tfidf_female = tfidf_male.toarray()"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"y_rnn = np.concatenate((np.zeros(len(flattened_array_tfidf_male)),\n",
" np.ones(len(flattened_array_tfidf_female))))"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"X_train_rnn, X_test_rnn, y_train_rnn, y_test_rnn = train_test_split(np.concatenate((flattened_array_tfidf_male, \n",
" flattened_array_tfidf_female)),\n",
" y_rnn,test_size=0.2)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"maxlen = 100\n",
"X_train_rnn = sequence.pad_sequences(X_train_rnn, maxlen=maxlen)\n",
"X_test_rnn = sequence.pad_sequences(X_test_rnn, maxlen=maxlen)\n",
"print('X_train_rnn shape:', X_train_rnn.shape, y_train_rnn.shape)\n",
"print('X_test_rnn shape:', X_test_rnn.shape, y_test_rnn.shape)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"max_features = 30000\n",
"model = Sequential()\n",
"model.add(Embedding(max_features, dimension))\n",
"model.add(LSTM(output_dimension))\n",
"model.add(Dropout(0.5))\n",
"model.add(Dense(1))\n",
"model.add(Activation('sigmoid'))"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"model.compile(loss='mean_squared_error',optimizer='sgd', metrics=['accuracy'])"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"model.fit(X_train_rnn, y_train_rnn, \n",
" batch_size=32, epochs=1,\n",
" validation_data=(X_test_rnn, y_test_rnn))"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"score,acc = model.evaluate(X_test_rnn, y_test_rnn, \n",
" batch_size=32)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"print(score, acc)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Sentence Generation using LSTM"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"# reading all the male text data into one string\n",
"male_post = ' '.join(filtered_male_posts)\n",
"\n",
"#building character set for the male posts\n",
"character_set_male = set(male_post)\n",
"#building two indices - character index and index of character\n",
"char_indices = dict((c, i) for i, c in enumerate(character_set_male))\n",
"indices_char = dict((i, c) for i, c in enumerate(character_set_male))\n",
"\n",
"\n",
"# cut the text in semi-redundant sequences of maxlen characters\n",
"maxlen = 20\n",
"step = 1\n",
"sentences = []\n",
"next_chars = []\n",
"for i in range(0, len(male_post) - maxlen, step):\n",
" sentences.append(male_post[i : i + maxlen])\n",
" next_chars.append(male_post[i + maxlen])\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#Vectorisation of input\n",
"x_male = np.zeros((len(male_post), maxlen, len(character_set_male)), dtype=np.bool)\n",
"y_male = np.zeros((len(male_post), len(character_set_male)), dtype=np.bool)\n",
"\n",
"print(x_male.shape, y_male.shape)\n",
"\n",
"for i, sentence in enumerate(sentences):\n",
" for t, char in enumerate(sentence):\n",
" x_male[i, t, char_indices[char]] = 1\n",
" y_male[i, char_indices[next_chars[i]]] = 1\n",
"\n",
"print(x_male.shape, y_male.shape)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"\n",
"# build the model: a single LSTM\n",
"print('Build model...')\n",
"model = Sequential()\n",
"model.add(LSTM(128, input_shape=(maxlen, len(character_set_male))))\n",
"model.add(Dense(len(character_set_male)))\n",
"model.add(Activation('softmax'))\n",
"\n",
"optimizer = RMSprop(lr=0.01)\n",
"model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"auto_text_generating_male_model.compile(loss='mean_squared_error',optimizer='sgd')"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"import random, sys"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"# helper function to sample an index from a probability array\n",
"def sample(a, diversity=0.75):\n",
" if random.random() > diversity:\n",
" return np.argmax(a)\n",
" while 1:\n",
" i = random.randint(0, len(a)-1)\n",
" if a[i] > random.random():\n",
" return i"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"scrolled": false
},
"outputs": [],
"source": [
"# train the model, output generated text after each iteration\n",
"for iteration in range(1,10):\n",
" print()\n",
" print('-' * 50)\n",
" print('Iteration', iteration)\n",
" model.fit(x_male, y_male, batch_size=128, epochs=1)\n",
"\n",
" start_index = random.randint(0, len(male_post) - maxlen - 1)\n",
"\n",
" for diversity in [0.2, 0.4, 0.6, 0.8]:\n",
" print()\n",
" print('----- diversity:', diversity)\n",
"\n",
" generated = ''\n",
" sentence = male_post[start_index : start_index + maxlen]\n",
" generated += sentence\n",
" print('----- Generating with seed: \"' + sentence + '\"')\n",
"\n",
" for iteration in range(400):\n",
" try:\n",
" x = np.zeros((1, maxlen, len(character_set_male)))\n",
" for t, char in enumerate(sentence):\n",
" x[0, t, char_indices[char]] = 1.\n",
"\n",
" preds = model.predict(x, verbose=0)[0]\n",
" next_index = sample(preds, diversity)\n",
" next_char = indices_char[next_index]\n",
"\n",
" generated += next_char\n",
" sentence = sentence[1:] + next_char\n",
" except:\n",
" continue\n",
" \n",
" print(sentence)\n",
" print()"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.5.3"
}
},
"nbformat": 4,
"nbformat_minor": 1
}